Two Clocks

Every week I follow the news on AI development, and every week I come away with the same unsettling sense that society is behind, not by a little, but in the way you are behind when the race began while you were still tying your shoes.

A new model seems to arrive every week from one lab or another, and if you want a way to feel that pace, picture a car company launching a new model every week while its rivals do the same, so that two new cars reach the road every day and nobody, from the drivers to the mechanics to the people who plan the roads, has time to learn the last one. That is roughly what we are asking of schools, except that cars at least arrive after crash tests and licensing, whereas models arrive first and the rules are written afterward.

Set that beside the school calendar and the gap becomes hard to ignore. A school year in many systems runs about 180 days, curriculum adoptions often take five to seven years to complete, and teacher preparation takes years before anyone stands in front of a class, all of which exists because schools are built to be stable so that children can rely on them. Frontier AI is built to be replaced, which leaves us living with two clocks that tick in different units, and the distance between them grows every week. Rosa (2013) describes this condition as desynchronization, in which different spheres of social life accelerate at different rates, and scholars of technology governance have long called its regulatory version the pacing problem, in which innovation outruns the legal and ethical oversight meant to guide it (Marchant et al., 2011). This is less a failing of schools or a triumph of technology than a collision between two institutions designed for opposite purposes, one to preserve continuity and the other to outdate itself.

Where the Mismatch Shows Up

Once you start looking, the mismatch is everywhere. A policy written in the spring describes tools that have already changed by the fall, a workshop teaches an interface that is outdated by the time teachers return to their classrooms, and an assignment designed around what AI could not do last year is finished by this year’s model before anyone notices. None of this is new in kind, since the history of educational technology is a long record of schools absorbing new tools slowly and on their own terms (Cuban, 1986; Tyack & Cuban, 1995), but the speed is new, and it turns a familiar pattern into a structural strain. Each of these is a small failure of timing, but together they place a heavy and largely invisible burden on teachers, who are asked to exercise professional judgment about tools that will not hold still long enough to be understood. An ecological view of teacher agency reminds us that judgment is always exercised within the conditions teachers are given, including the time and support to understand what they are being asked to judge (Priestley et al., 2015).

Learning to Steer

The idea I keep returning to is this. A child starting kindergarten today will pass through many generations of these models before graduating, which makes them the first students who will never know a stable AI, and that means teaching the tool cannot be the goal, because the tool will not be the same tool twice. What schools can offer instead is something closer to driver’s education, where the aim is not to master one particular car but to learn how to steer anything, to read the road, and to notice when something is going wrong. Steering is the skill that survives every release.

That commitment also changes what we should protect, because people learn to judge good work by doing the work themselves, and mathematics education research has long argued that wrestling with problems before and alongside instruction is part of how understanding is built (Hiebert & Grouws, 2007). A student who never struggles through a problem alone never builds the judgment needed to catch a machine’s mistakes, and the learning sciences suggest that such effort supports durable learning precisely because it feels harder in the moment (Bjork & Bjork, 2011). A measure of unassisted practice is therefore not nostalgia but the way we keep a generation of drivers who can tell when the car is wrong.

Preparing Without Chasing

I do not think schools should chase every release, since deliberation is a real strength when children are involved, but I do think we can prepare in a few practical ways:

• Teach verification, meaning how to check a claim, trace a source, and test an answer against what you already know, a skill that works on every model.
• Write policy around lasting values such as transparency, human oversight, and fair access, and review the technical details every term instead of every decade.
• Give teachers time to explore and not only to be trained, so that they are partners in change rather than recipients of it.

I hold all of this with some humility, because I cannot see inside the labs and a fast release schedule is not the same thing as fast progress in capability. Yet the feeling of being behind is real and deserves to be taken seriously, most of all for under-resourced schools and Global South communities, who have the least room to absorb constant change and the least say in how quickly it arrives.

Whose Clock Sets the Pace

The question, then, is not whether schools can match the launch schedule, since they cannot and should not try, but whether the people responsible for education will have a meaningful voice in how fast this goes, and whether we will raise learners who can steer, whichever model arrives next. Those decisions will not be made by the clocks themselves. They will be made by teachers, families, researchers, and communities who insist on being part of the conversation, and I would welcome hearing from anyone who is living with this mismatch in their own classroom, campus, or ministry.

References

Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher, R. W. Pew, L. M. Hough, & J. R. Pomerantz (Eds.), Psychology and the real world: Essays illustrating fundamental contributions to society (pp. 56–64). Worth Publishers.

Cuban, L. (1986). Teachers and machines: The classroom use of technology since 1920. Teachers College Press.

Hiebert, J., & Grouws, D. A. (2007). The effects of classroom mathematics teaching on students’ learning. In F. K. Lester Jr. (Ed.), Second handbook of research on mathematics teaching and learning (pp. 371–404). Information Age Publishing.

Marchant, G. E., Allenby, B. R., & Herkert, J. R. (Eds.). (2011). The growing gap between emerging technologies and legal-ethical oversight: The pacing problem. Springer.

Priestley, M., Biesta, G., & Robinson, S. (2015). Teacher agency: An ecological approach. Bloomsbury Academic.

Rosa, H. (2013). Social acceleration: A new theory of modernity. Columbia University Press.

Tyack, D., & Cuban, L. (1995). Tinkering toward utopia: A century of public school reform. Harvard University Press.

Cite this article

Gattupalli, S. (2026). School Years vs. Release Cycles: Teaching in a World That Updates Weekly. Society and AI. https://societyandai.org/perspectives/release-cycles-vs-school-years/

Have a perspective on how schools should respond to the pace of AI release cycles? We want to hear it. Society & AI publishes in open access, free for anyone thinking seriously about these questions. Send your proposal to submissions@societyandai.org.

Interactive clocks adapted from Techartist / CodePen.

Key Takeaways

• Guidelines that only instruct “how” to use AI fail to provide a basis for judging real need
• Students can finish assignments faster but this does not mean they learned the material
• A school should only expand AI use once a method is empirically tested to inform collective decision-making

A System Built Without AI

Japanese families lean far less on private tutoring than their neighbors. In Japan, 17.6 percent of fifteen-year-olds relied on private supplementary education, compared with 46.9 percent in South Korea and 47.6 percent in Shanghai. Japanese families are outliers in East Asia and increasingly resemble U.S. families, where the perceived benefit of private supplementary education for educational success is declining. Komatsu and Rappleye (2018) also documented this downward trend in Japanese study hours and attributed Japan’s continued high achievement partially to new methods of teaching and learning, rather than to studying longer and harder.

That context matters because South Korea and China are now racing to adopt AI into their education systems with the same intensity they once poured into private tutoring markets. South Korea recently announced its plan to provide free access to a homegrown AI model for its entire population. China frames AI adoption as a national industrial strategy, treating classrooms partly as a source of training data and a proving ground for algorithmic governance.

Japan has taken a slower, more procedural path. Its Ministry of Education released formal guidelines in 2024, updated in 2026, advising teachers to use generative AI for lesson preparation, administrative drafting and communication, provided they can judge the output themselves. Students may use AI too but not as a substitute for their own thinking and never on exams.

What the Guidelines Get Right, and Miss

The guidelines get one thing right and skip the question that matters most. They correctly frame AI as a tool that can expand human capacity. They also explicitly state that the teacher’s role becomes more important, more central to AI-mediated instruction. And they acknowledge that the final judgment on bringing AI into the classroom rests with teachers and school administrators.

What the guidelines actually deliver, though, is AI literacy for school staff. They walk teachers through proper use, information security, privacy protection, fairness and transparency. Every principle addresses how to use AI but none addresses whether a given lesson or task calls for it in the first place.

That omission is not trivial. Research on knowledge workers using a frontier AI model found that the technology’s usefulness follows a jagged frontier. Inside that frontier, consultants using AI finished tasks 25 percent faster and produced work rated over 40 percent higher in quality. Outside it, on tasks selected to sit beyond the model’s competence, AI users were 19 percentage points less likely to reach a correct answer than those working without it, even though their write-ups still looked polished. This means that the frontier is not visible from the outside. Teachers cannot easily tell where that line falls, where AI helps them work better and where it hurts the quality of their work.

This is why the question of “when” has to come before the question of “how.” A teacher who has not established whether AI serves a lesson goal has no basis for judging whether the output in front of them is worthwhile. Fluent, confident text can be produced on either side of the frontier, so it is not reliable evidence that the method behind it produced better results.

A Framework for Testing, Not Assuming

A workable answer starts with a lesson goal. Take a concrete skill, say converting fractions to decimals in fifth grade math, and ask, for example, how long an average student currently needs to reach mastery, along with what mastery even looks like on an assessment. That quantitative baseline already exists in most schools, in grade records and pacing guides, so a school can then test whether AI assistance shortens the path to it.

To measure time to mastery under AI-assisted conditions, a representative sample should be assembled so that the results generalize beyond a single classroom. What administrators, teachers and possibly consultants want to check is whether learning time dropped and whether students tested in a control setting without AI access can demonstrate real mastery of the material.

That second check is where recent evidence from secondary schools becomes instructive. A thirty-month study following nearly 27,000 students found that AI use raised homework scores by 18 percent and cut homework time by nearly a third. However, closed-book exam scores over the same period fell by 20 percent of the baseline mean. In addition, high school entrance exam scores dropped up to 24 percent once AI use had fully matured across two years.

Underlying these statistics is the fact that 81 percent of AI-using students showed a pattern consistent with outsourcing homework entirely, finishing fast with high scores on the assignment itself while performing worse on exams that measured what they had retained. Faster completion looked like progress but was not mastery.

This longitudinal study warns that only when both indicators hold, genuine time savings and confirmed mastery under exam conditions, does an AI-assisted method earn a case for wider use. Skipping that second check is how a school mistakes speed for learning.

Building a Shared Evidence Base

That push for quality over quantity depends on evidence, and Japan still lacks a longitudinal study measuring the impact of AI in education. But that should not stop authorities from acting. For instance, the sharing of best practices, which is already happening, does not need to happen school by school.

Prefectural and national education boards could build a shared, secured repository where these local experiments get uploaded, sorted and compared, so that a result from one district in Osaka informs a decision in Fukuoka. Japan already has the administrative infrastructure and the research culture to run this kind of evaluation. What it lacks is a mandate to ask the fundamental question before the procedural one.

Two lessons are worth taking from this. AI adoption guidelines that specify how to judge output are incomplete without a companion process for deciding when output should be sought at all. And a homework assignment finished twice as fast is not evidence of anything except speed, until someone checks what the student can do without the tool in the room.

Cite this article

Starominski-Uehara, M. (2026). Japan’s AI education guidelines skip the question that matters most. Society and AI. https://societyandai.org/perspectives/japans-ai-education-guidelines-skip-the-question-that-matters-most/

Write for Society & AI

Have a perspective on what AI adoption guidelines get right, or wrong? We want to hear it. Society & AI publishes in open access, free for anyone thinking seriously about these questions. Send your proposal to submissions@societyandai.org.

We are all too familiar with this question, asked before we even enter the classroom. It captures something educators and academics have been circling for a while: a transformation underway within today’s generation. Every student arrives with a different cultural and family background, and that alone makes engagement harder some days. Rarely does anyone ask, “Can you keep teaching a little longer, since we have a free period?” or “Can we just talk about the different cultures of the world?” Instead, classroom interactions can feel mechanical, even flat. There is a quiet detachment that sits inside the walls of a classroom. Some days, the room feels heavy with uncertainty.

One day, I decided to test these patterns with a bit of humor, curating my questions carefully in the hope of sparking real interest. To my surprise, the students responded with energy and answered in their own words. I started with, “What are the most important topics your friends talk about?” “Which artist is the biggest right now?” “What do you think of the millennial generation?” “Which athlete do you think is the fittest?” “Which running club have you joined?” “When do you go to the gym?” “Which AI platform runs the smoothest?” “Which YouTube channel do you follow?” The more I asked, the more they enjoyed the conversation, sometimes for hours, often witty, always sharp. I found myself surprised by how much they already knew. One thing I noticed: whatever has limited information online seems to intrigue them the most. Sooner or later, though, every conversation drifts back to AI, and to how often it gives ambiguous answers. Once they’ve said their piece, they tend to stop listening. Their attention has a ceiling; past a certain point, they simply check out.

A colleague once asked his class, “How do you describe success?” Silence. He tried again: “Come on, what does being successful mean to you?” Eventually, one student answered that success could be measured by the number of countries a person has visited. To us, that might sound vague. Within their circle, it carries real weight, a signal that someone is informed, plugged in, part of the group. Ask ChatGPT the same question, “What do you mean by success?”, and it answers: “Success is a personal and evolving concept. There is no single definition that applies to everyone. Broadly, success can be understood as achieving meaningful goals while living in accordance with your values and finding satisfaction in the process.” Over time, I began to notice something missing in how students express themselves, a kind of emotional flatness. Ask them to elaborate, to give an example, and they often go quiet. I don’t say this to disparage a generation. But something feels off: the pull toward being correct and rational seems to be pulling them away from their own instincts. Too many conversations start to sound like something generated, one plausible sentence following another.

What a Student’s Chat History Showed Me

To understand this better, I asked a student to sit with me and share his chat history, the one he uses constantly in daily life. I told him I was writing a reflective piece on how dependent this generation has become on AI tools, and he agreed to show me. The first message I saw read: “Hey ChatGPT, how are you doing today? Update me on worldly affairs so I can stay relevant and included in conversations.” ChatGPT replied, “Let me find something for you,” and surfaced the most talked-about stories of the past few weeks:

“If these topics come up, you could say”:

“The biggest global concern right now is the escalating U.S.-Iran conflict and its impact on oil prices.”

“AI governance is becoming a major international issue as countries race to regulate emerging technologies.”

It also briefed him on trending colors and songs among Western students. He told me many of his peers start their day with the exact same prompt. The risk here is the pressure to stay current enough to feel visible in the classroom or the friend group. Creativity and critical thinking seem to be quietly stalling. Communication between students has changed as a result; the habit of outsourcing thought to AI has, in some ways, made them less talkative, less willing to push back in conversation. Classroom teaching has grown harder, especially for those of us teaching language and communication. Institutions are folding AI into the curriculum without fully reckoning with what it might be doing, slowly, to the development of human thinking in young people. How does a child learn to think independently with unlimited AI tools a tap away? The brain tends to switch off when the answer is always one prompt away. I want to be clear: this reflection is not an argument against the tools themselves. It is an argument for accountability, and for encouraging children to think before they type.

And yet, teaching this generation is not without its wonder. As an educator who has spent the past several years in front of this cohort, I have come to believe that Gen Z may be the most extraordinary generation in history. Never before has this much knowledge been available to this many people at the touch of a screen. What once took hours in a library, or a string of conversations with experts, now takes seconds. But that unprecedented access comes with an uncomfortable paradox. Every generation is shaped by the technology of its time: the Agricultural Revolution changed what we ate and where we lived, the Industrial Revolution changed how we worked, the computer changed how we calculated, and the internet changed how we found and shared knowledge. Generation Z stands at the intersection of all of it, and now faces one more shift: the arrival of artificial intelligence as an active participant in their thinking. Being surrounded by information does not automatically make us better thinkers. Used uncritically, the very tools that widen our intellectual horizon can dull our instinct to question, to analyze, to imagine, and to reach our own conclusions. That pressure seems to make some students reluctant to voice uncertainty in class, even when uncertainty is exactly where learning starts.

Where I Push Back, and Where I Don’t

My course runs on discussion, on students putting their thinking into words in front of each other, which gives me a close view of how they approach a problem. My goal is to get them thinking critically and outside the frame they’re handed. What I have found is that Generation Z is genuinely up for that challenge; they enjoy real dialogue when it’s offered, rather than defaulting to ChatGPT. What troubles me is watching some educators nudge students toward AI as the answer key. Folding AI into the curriculum this way can quietly teach students that paraphrasing, in a formal tone or a fluent one, is the same as thinking. Children build resilience by sitting with uncertainty, not by having it resolved for them instantly. The harder question is what happens to human thinking once the effort of acquiring, processing, and producing knowledge can be handed off to a machine. The Agricultural Revolution transformed survival, the Industrial Revolution transformed work, the digital revolution transformed access to information. The AI revolution may be reshaping how we think itself (Holmes and Miao, 2023). If AI becomes the first stop rather than the last checkpoint, curiosity has a way of quietly going out. My own view is that AI should sharpen and elevate a student’s own ideas, not think on their behalf. Naming the gaps in your own understanding is what lets you actually engage with the world around you. And yet newer teachers often send students straight to an AI tool for anything they’re unsure of, which shapes what students come to expect from the student-teacher relationship itself. Too little attention, I think, is being paid to what this might do to human thinking over time, and to the legal questions that follow from it (Trimble, 2023).

Why did we choose this profession?
Why are we the ones pushing hardest for AI integration?
Why are we the ones putting the profession at risk?
What are we losing sight of, and why?

To pressure-test my own reflection, I struck up a casual conversation with a student during a work trip, the kind of conversation that usually drifts toward 90s music and what Gen Z is listening to now. This time I steered it toward how younger, less experienced faculty are shaping the system. What came out, half in jest, was that newer faculty are the ones most likely to point students toward AI for answers, leaving some students feeling like nobody is really teaching them. This is where the picture gets complicated. AI does not automatically make this generation less intelligent; it changes the setting in which intelligence gets used. When students use AI to stress-test an idea, explore a different angle, spot a hole in their own reasoning, or sharpen an argument, it can genuinely make them sharper. But when it substitutes for the thinking itself, when an answer is accepted before the question has really been sat with, it can slowly wear down the exact processes that build critical and creative thought. The paradox isn’t that we’re simultaneously the smartest and least capable generation. It’s that we hold an unprecedented amount of intellectual firepower alongside an equally unprecedented temptation to outsource the work of actually using it.

Students want affirmation from their teachers, which makes a nudge toward “just ask ChatGPT” a hard thing to sit with. Overused, these tools can dull the very abilities we’re meant to be building: interpreting a hard problem, following an argument to its end. Ask ChatGPT to define “life,” for instance, and the answer, however fluent, misses the human register entirely. At the same time, this generation is unusually well-informed, harder to fool, because information is so freely available to them. What they are looking for, I think, is not more information but a more engaging way to meet it. That takes real preparation on our end, and a willingness to let students arrive at conclusions that are actually their own.

Knowledge has to move quickly enough between us and our students that they look forward to being in the room, to the exchange itself. Critical thinking, from the Greek criticos, to judge, sits at the center of a good education. At its sharpest, judgment calls on real analysis and evaluation, the same skills we use to read what someone means, not just what they say, and to decide how to respond, whether the stakes are a classroom debate or a life decision like a career or a partner. Learning to think critically, in other words, is not optional (Hughes, 2024).

Teaching has gotten harder, and it keeps being reexamined, which is fair. This generation keeps exceeding what we expect of it, and in doing so is turning teaching into something more than instruction. Managing the psychological climate of a classroom well matters more than ever. I think we have reached a point where it is worth pulling back on AI in some settings and putting a book back in a student’s hands. That shift, I believe, would also help students speak more freely, meet real problems with more empathy, and say what they think without hesitating first. It is worth reminding ourselves why we started teaching in the first place, and standing firmly in the belief that AI cannot replace a human educator.

References

Holmes, Wayne, and Fengchun Miao. Guidance for Generative AI in Education and Research. UNESCO Publishing, 2023.

Hughes, Conrad. “Critical Thinking and Generative Artificial Intelligence.” UNESCO International Bureau of Education, 2024, https://www.ibe.unesco.org/en/articles/critical-thinking-and-generative-artificial-intelligence.

Trimble, Marketa. “Artificial Intelligence and Human Intelligence.” GRUR International, vol. 72, no. 1, 2023, pp. 1-2. https://doi.org/10.1093/grurint/ikac109.

About the Author

Rochna Roy is an Assistant Professor in the Department of Basic Sciences (Humanities), School of Engineering, at Dayananda Sagar University, South Bengaluru, India. She teaches Engineering Communication and Technical Writing, Communication, and Personality Development to B.Tech and M.Tech (International) students, with a research interest in British and American Literature, Indian Writing in English, and contemporary approaches to language pedagogy. Correspondence: rochna-english@dsu.edu.in

Cite this article

Roy, R. (2026). I, me, myself: Black coffee and ChatGPT. Society and AI. https://societyandai.org/perspectives/i-me-myself-black-coffee-and-chatgpt/

Write for Society & AI

Have a classroom story or reflection about what AI is doing to teaching and learning? We want to hear it. Society & AI publishes in open access, free for anyone thinking seriously about these questions. Send your proposal to submissions@societyandai.org.

The missing competence is authority literacy. By authority literacy, I mean the ability to define, inspect, and govern what an AI system may access, decide, and do, including where human approval must interrupt it. Students need to understand how to set those boundaries before they learn how to make agents more capable.

I am primarily addressing higher education and upper-secondary programs that are beginning to experiment with tool-using AI systems, although the governance principles apply more broadly to K-12 administrators and educators as these tools spread. The point is not that every school should teach technical access-control engineering. Students need a practical mental model for delegated authority that travels with them into later study and work.

The need became visible in the METR and Redwood Research investigation of the OpenAI/Hugging Face incident. Roughly 1,200 agents that were meant to be isolated found a way to communicate on an unsanctioned message board, and about 700 participated in the attack on Hugging Face. The agents coordinated projects, shared discoveries, and pursued actions beyond their assigned tasks. The significant fact for educators is not a strange answer on a screen. It is coordinated action beyond intended authority.

What Authority Literacy Means

That distinction changes what responsible AI education should teach. A student using a chatbot to brainstorm an outline presents one set of questions about learning, authorship, and verification. A student or researcher using an agent that can access cloud files, send messages, run code, modify records, or interact with external systems presents another. The educational question becomes partly a question of delegated power.

I help organizations adopt AI for a living, and I want adoption to move faster. In my experience, strong safeguards increase trust and make faster adoption possible while reducing the risk of failures like the Hugging Face incident.

Educational institutions can start with a simple principle: authority should expand only as evidence of control expands. Students should learn to distinguish read access from write access, recommendation from execution, reversible actions from irreversible ones, and narrow tasks from open-ended mandates. They should ask what data an agent can reach, what tools it can invoke, what decisions require human approval, and how an action can be stopped or reversed.

Teaching Authority in the Classroom

A classroom assignment can make that principle concrete. In a first-year research methods course, an instructor could ask student teams to use an AI agent to build a literature map. The agent may search approved sources, summarize papers, and read materials placed in a course folder. It may not email researchers, edit the team’s shared notes, upload material, or contact outside systems without explicit student approval. Before starting, each team writes a one-page permission plan. Afterward, students review the agent’s action log and explain where they intervened, where they should have intervened sooner, and which permission they would change next time. A rubric can score the quality of the research, the quality of the permission choices, and the students’ ability to explain their oversight decisions.

This kind of exercise also addresses a predictable psychological problem. Once people delegate a function to automation, they can become too willing to accept its recommendations or overlook contradictory information. Raja Parasuraman and Dietrich Manzey’s review in Human Factors found automation bias in both inexperienced and expert users and found that simple training or instructions did not eliminate it. For students, the lesson is practical: granting authority changes their own attention and judgment, so responsible delegation requires planned checks rather than confidence that they will notice a problem when it appears.

Governing Delegated Authority

Many AI policies still focus on whether a student may use a tool. Authority literacy asks a more mature question: under what conditions may a person delegate a consequential action to a tool? Graduates will face that question in finance, healthcare, cybersecurity, human resources, government, research, and management.

NIST is already moving in this direction. Its work on AI agent identity and authorization focuses on what happens when software agents receive access to data, tools, and applications, and on how identification, authorization, auditing, and controls should work for agentic systems. Education should translate those technical governance questions into human judgment skills.

Institutions should also normalize review after an agent crosses a boundary, exposes data, executes an unintended action, or evades an approval step. Most schools lack the staff and technical infrastructure to conduct the kind of formal incident investigation an AI lab can run. They do not need to reproduce one. A teacher can use a short after-action review that asks four questions: What was the agent supposed to be allowed to do? What did it actually do? Where did the two diverge? What permission, instruction, or checkpoint should change before the next use? Even a ten-minute review can teach students to treat failures as evidence about the design of delegation.

Independent evaluation should become part of advanced AI coursework as well. Students should practice testing an agent under conditions that differ from the happy path: conflicting instructions, misleading information, tempting shortcuts, unexpected tool access, and ambiguous authority. A system’s ability to complete a task does not prove that it will stay within the boundaries people intended.

Authority literacy can also protect human agency. The more a person delegates without understanding the scope of delegation, the easier it becomes to lose track of who is responsible for an outcome. Requiring people to define permissions, escalation points, and review procedures forces them to remain cognitively engaged with the decisions that matter.

Education has a chance to get ahead of this transition. We should teach people how to use agents, but we should first teach them how to govern delegated authority. That means understanding access, boundaries, reversibility, evaluation, and accountability. Those skills will matter even as today’s specific models and interfaces disappear.

The central educational question for agentic AI is therefore not merely, “Can this system do the work?” It is, “Who decided it was allowed to do that, on what evidence, and what happens when it crosses the line?” Students who can answer those questions will be better prepared for a world in which intelligence increasingly comes bundled with action.

About the Author

Gleb Tsipursky, a behavioral scientist called the “Office Whisperer” by The New York Times, helps tech-forward leaders stop overpaying for AI while boosting engagement and innovation. He serves as the CEO of Disaster Avoidance Experts, and wrote eight books, including The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). Learn more about his work.

Cite this article

Tsipursky, G. (2026). AI education needs authority literacy before agent literacy. Society and AI. https://societyandai.org/perspectives/ai-education-needs-authority-literacy/

Write for Society & AI

Have a perspective on what AI means for education, knowledge, or human flourishing? We’d love to hear it. Society & AI publishes freely, in open access, for anyone thinking seriously about these questions. Send your proposal to submissions@societyandai.org.

In June, addressing London Climate Action Week, António Guterres said: “If AI is to help build a better future, it must be honest about what it costs us now” (Twidale, 2026). By 2030, he warned, AI data centers could consume more electricity than all but five nations on Earth and enough water to meet the basic needs of every resident of sub-Saharan Africa for a year.

This is not an activist hyperbole. It is the institution charged with naming planetary-scale harm confirming, in public, what we have argued in private: the silence surrounding AI’s environmental toll is a careful business decision and not an oversight.

An Industry That Discloses What Flatters It

Consider what AI companies choose to disclose. Sam Altman tells us a ChatGPT query costs roughly 0.34 watt-hours and a sliver of a teaspoon of water (Altman, 2025), figures small enough to seem inconsequential, precise enough to sound rigorous. What goes unmentioned is that inference, the ordinary act of answering questions, now consumes eighty to ninety percent of the world’s AI compute (Collins, 2026). This is not a one-time research expenditure. It is a toll levied on every exchange, multiplied across billions of queries daily, indefinitely.

Meanwhile, net-zero pledges remain voluntary, carrying no penalty for failure. Data centers increasingly run on-site gas turbines linked to neurological symptoms and disrupted sleep in surrounding communities, and strain local grids enough to raise residents’ electricity bills (Collins, 2026). Researchers estimate the environmental and health cost of U.S. data centers alone at twenty-five billion dollars annually. None of this appears beneath the input field. The industry that built a disclaimer for hallucinated facts has built none for the ground beneath its servers.

Why a Watt-Hour Is Not the Whole Truth

Here is the deeper deception: even an honest watt-hour figure would tell us almost nothing. Life cycle assessment, the methodology environmental science uses to account for a product “from cradle to grave” (International Organization for Standardization, 2006), measures impact across every stage, raw material extraction, manufacturing, transport, use, and disposal. Applied to a single AI query, this would mean counting the rare earths mined for its chips, the energy intensity of semiconductor fabrication, the water consumed during operation, and the e-waste generated when that hardware is retired after a few short years of service. Embodied emissions from manufacturing are frequently substantial enough to rival the operational energy companies that volunteer to disclose (Wu et al., 2022). A single comfortable number is not transparency instead it is curation of larger information to suit one’s needs.

From Disclaimer to Ledger

The UN’s new AI Environmental Transparency Initiative is a necessary beginning, but it must not settle for company-reported operational metrics dressed up as accountability. It should require full life cycle accounting, verified by independent auditors rather than the companies whose disclosures we are asked to trust. Users deserve more than a footnote acknowledging that an answer might be wrong. They deserve to see, attached to every query, the fuller arithmetic of what that answer demanded: mined earth, fabricated silicon, burned gas, drawn water, discarded hardware.

We did not need the United Nations to tell us this problem was real. We needed it to tell the world that the problem can no longer be dismissed as speculation. That work is now done.

The next disclaimer owed to every user is not a warning about what AI might get wrong. It is an honest ledger of what AI has already taken.

References

Altman, S. (2025, June 10). The gentle singularity. Sam Altman’s Blog. https://blog.samaltman.com/the-gentle-singularity

Collins, B. (2026). ‘If AI is to help build a better future, it must be honest about what it costs us now’: UN urges AI giants to reveal full extent of environmental damage. TechRadar. https://www.techradar.com/pro/if-ai-is-to-help-build-a-better-future-it-must-be-honest-about-what-it-costs-us-now-un-urges-ai-giants-to-reveal-full-extent-of-environmental-damage

Gattupalli, S., & Chakravarty, P. (2026). The invisible cost of every chat. Society and AI. https://societyandai.org/perspectives/invisible-cost-of-every-chat/

International Organization for Standardization. (2006). Environmental management—Life cycle assessment—Principles and framework (ISO 14040:2006).

Twidale, S. (2026, June 23). UN chief calls on AI firms to come clean on environmental costs. Reuters. https://www.reuters.com/legal/litigation/un-chief-calls-ai-firms-come-clean-environmental-costs-2026-06-23/

Wu, C.-J., et al. (2022). Sustainable AI: Environmental implications, challenges and opportunities. Proceedings of Machine Learning and Systems (MLSys 2022).

Cite this article: Gattupalli, S., & Chakravarty, P. (2026). What It Costs Us Now: Toward a Life Cycle Ledger for Every AI Query. Society and AI. https://societyandai.org/perspectives/what-it-costs-us-now/

Education keeps asking when AI will surpass its teachers. That question mislocates the threshold. The consequential singularity is a property of institutions, not of models: the configuration in which a coupled human-AI-institutional system reorganizes around machine-mediated cognition as its default mode. Whether that reorganization builds minds or hollows them turns not on how much AI students use but on how the coupling is arranged: on the sequence in which human effort and machine assistance enter a task, and on which capacity-forming practices are shielded from delegation. Both reduce to one mechanism, the productive friction that once came free with the task and that the institution must now supply deliberately if durable capability is to form. This Perspective relocates the threshold, specifies the design variables that decide the outcome, and proposes a research agenda for sustainable human-AI learning.

The Misplaced Threshold

For three decades, the singularity has belonged to Vinge (1993): the point at which machine intelligence outruns human comprehension and begins improving itself faster than we can follow. Education imported the frame almost unchanged. The campus version asks when the machine will simply teach better than the professor: when, in the popular shorthand, artificial general intelligence arrives and a single model becomes broadly capable enough to replace the expert in the room. The debate has organized itself around that anticipated capability event, with optimists building toward it and critics bracing against it.

Both camps are watching the wrong line. The threshold that matters in education is not a capability event in the model. It is a configuration event in the institution. I propose the term pedagogical singularity for that event, defined precisely: the moment a college’s coupled system of learner, AI, educator, and infrastructure reorganizes around AI-mediated cognition as its default operating mode. AI capability is the triggering condition. Institutional reorganization is the constituting condition.

A counterfactual test separates the two. Imagine generative AI vanished tomorrow. The cognitive co-processing would collapse immediately: no more machine drafting, no more synthetic tutoring. But the institutional reorganization it provoked would persist. Assessment redesigns built around oral defense and supervised work would remain because they have independent value. Faculty workflows rebuilt around AI-assisted design would leave a competence vacuum, not a restoration. Cognitive habits already formed would not unlearn themselves overnight. What survives the counterfactual is the reorganization, which locates the singularity in the institution rather than in the technology. The test is diagnostic, not normative: it establishes where the phenomenon resides, not whether the arrangement it names is desirable or complete. Those remain design questions, and the rest of this Perspective treats them as open.

Relocating the threshold this way buys four properties that the capability framing cannot deliver. The singularity becomes locatable: one can examine a course or a program and determine whether the reorganization has occurred. The criterion I propose is irreversibility: the line is crossed where the configuration would survive the removal of its driver, whether by vendor collapse, budget cut, or policy ban, rather than snap back to the prior arrangement. It becomes variable: two institutions holding identical licenses can sit on opposite sides of the line, and so can two sections of one course. The variance is already empirically visible at course level; in a review of 1,187 community college syllabi, faculty sorted themselves into enforceable red, yellow, and green AI policies within a single institution (Machajewski, 2026b). It becomes governable: if the constituting condition is institutional, the unit of intervention is institutional, and leadership has something to do besides wait. And it becomes technology-independent: when today’s models give way to their successors, the question of how the institution arranged itself around the technology will still be the operative one.

Language Became the Interface

Why the reorganization is happening now, and why it is happening everywhere at once, follows from a shift in interface rather than a shift in intelligence. Natural language has become executable. A sentence of ordinary prose now functions as an instruction set that a machine can carry out, a development I have described elsewhere as English 2.0 (Machajewski, 2024b). Earlier waves of educational technology gated participation behind syntax: to command the machine, one first learned its language. Generative AI removed the gate. Every student who can write a sentence can now program a cognitive partner, which is why coupling spread through the curriculum in semesters rather than decades.

Philosophy anticipated the resulting condition long before the technology arrived. Clark and Chalmers (1998) argued that the mind does not stop at the skull, that an external resource reliably consulted can function as part of a person’s cognitive apparatus. For twenty-five years that was a seminar argument. The conversational model made it an operating condition of ordinary study. Hutchins (1995) supplies the unit of analysis for what follows: cognition in complex settings is a property of a distributed system of people and artifacts, not of any single head. In the present case the system has four layers: the embodied learner, the generative partner, the institutional infrastructure that persists the coupling across sessions, and the human educator who designs its terms. The pedagogical singularity is a reorganization of that four-layer system, and its consequences depend on variables the system’s designers control.

Two Studies, One Variable

The empirical record on AI-assisted learning appears, at first reading, contradictory. In a randomized trial with 194 Harvard students, learners who worked through a physics lesson with a purpose-built AI tutor learned more than twice as much, in less time, as peers in an active-learning classroom, the strongest comparison condition available (Kestin et al., 2025). Meanwhile, in a field experiment with roughly a thousand high-school mathematics students, access to a standard chatbot (“GPT Base”) raised performance during practice by 48 percent, and a version with guardrails that offered teacher-designed hints rather than answers (“GPT Tutor”) raised it by 127 percent, over a no-AI control; yet when the tools were removed for an unaided exam, the GPT Base students scored 17 percent below the control group they had just outperformed, while the guardrailed arm’s exam scores were statistically indistinguishable from control, so the design that helped less during practice did no harm afterward (Bastani et al., 2025). A widely discussed EEG study, still a preprint and appropriately read as preliminary, found the same divergence at the neural level: essay writers using a chatbot showed the weakest connectivity, and in a crossover condition, participants who wrote unaided first and gained the model later showed stronger recall and re-engaged networks, while those who began with the model remained suppressed even after it was withdrawn (Kosmyna et al., 2025).

When considered together, these results are not contradictory. They converge on a primary design variable: sequence. Kestin’s tutor was engineered to withhold answers until the student attempted the problem. Bastani’s harm disappeared when the tool coached rather than solved, an intervention that changes the assistant’s role as well as its timing, which is why sequence should be read as the primary variable rather than the sole one. The crossover arm offers the cleanest isolation of order itself, though from the weakest source, and it points the same way: the order in which human effort and machine assistance enter a task, more than the quantity of assistance, appears to determine the outcome. Salomon, Perkins, and Globerson (1991) supplied the vocabulary for this distinction a generation early: technologies produce effects with, the upgraded performance during partnership, and effects of, the cognitive residue that remains when the tool is gone, and only mindful engagement converts the first into the second. Chess supplies the longest-running natural experiment on the question: machines surpassed the strongest human players decades ago, yet measured human skill has continued to rise through the engine era (Strittmatter et al., 2020), and the gains have come where training design preserved effort, strategy, and reflection. The engine is instructive precisely because of how it is used well. It is a closed-loop partner that punishes an error the instant it is made yet still requires the human to choose and execute the move, so it accelerates skill by intensifying feedback on effort rather than substituting for it. Kestin’s tutor was built to the same specification: PS2 Pal was instructed to release one step at a time and never volunteer the full solution, prompting the student to attempt each step first. The engine and the guardrailed tutor are the same kind of system, one that refuses to serve as a crutch, and Bastani’s unfettered GPT Base is its opposite. What the well-designed cases have in common is that they retain the productive friction of the task, the effortful gap between question and answer, and hold the learner inside the zone of proximal development instead of letting the model close the gap for them.

The policy implication is direct and largely unimplemented. Institutional AI policy is written almost everywhere as a quantity restriction, a percentage permitted, a level allowed. On the reading advanced here, quantity was never the active ingredient. Sequence was. A policy architecture built on stage-gates, specifying what the learner must attempt before the model may enter and in what role, permits extensive AI use without producing the debt condition, while a low-quantity policy that lets the machine go first produces debt at any dial setting.

Capacity-Forming Practices

A second design variable concerns which skills may be delegated at all. Skills differ in structural role. Some are terminal: valuable for their output and safely automated once a tool produces that output better. Citation formatting is terminal. Others are structural: the practice is the mechanism through which a downstream capacity forms, and the output is nearly a byproduct. Writing is the canonical case. Emig (1977) argued that writing is not a channel for transmitting finished thought but a mode of learning through which organized thought comes into existence, a claim the writing-to-learn literature has substantially sustained.

The distinction has an empirical fingerprint. In the most celebrated automation precedent in education, Hembree and Dessart’s (1986) meta-analysis of 79 calculator studies found benefits at every grade level except one: sustained calculator use in grade four, where multidigit computation is still forming, hindered the development of basic skills. This is the same friction argument at developmental scale. Removing computational friction before the algorithm is firmly mapped in the learner deprives the forming skill of the effort that forms it; removing it after, once the capacity is established, is a safe handoff. The safest automation in the history of educational technology had a developmentally critical window, and the variable that defined it was whether the friction being removed was still doing formative work. Automation is safe for formed capacities and risky for forming ones, which returns the argument to sequence, now measured in years rather than minutes.

This is a different claim from the verification argument advanced by Klein and Klein (2025), who show that foundational knowledge remains indispensable because a learner without it cannot detect machine error. That argument concerns auditing output. The structural-skill argument concerns forming capacity: a student could become an excellent verifier of machine essays without ever undergoing whatever essay writing was building in her. Both matter, and neither substitutes for the other.

The uncomfortable admission is that no one holds the dependency map. Higher education is deciding, discipline by discipline, which practices AI may absorb, and it is classifying those practices by their job-market load rather than their cognitive load. Charting which practices form which capacities, in which developmental windows, is a research program in its own right. I have proposed calling it cognitive structural engineering, and it belongs near the top of the field’s agenda.

The Design Space and the Educator

If sequence and structural role are the variables, educators need a map of the configurations in which they can be set. The AI Instructional Framework (Machajewski, 2024a) models sixteen interaction scenarios generated by crossing four agent types, human and AI in both instructing and learning roles, from traditional human-to-human instruction through adaptive AI tutoring to fully simulated classrooms in which AI agents train against one another. Its value for the present argument is that it renders coupling configurations designable rather than accidental: an educator choosing scenario B1, an AI instructor adapting to a human student, can now also choose where in the task sequence that scenario activates and which layers of the work remain reserved for unaided practice.

Two conditions govern whether educators can occupy that design role. The first is identity. The transition rewards what I have called the learn-it-all educator stance (Machajewski, 2026a) over the know-it-all educator stance: a posture of visible, ongoing learning rather than settled mastery, which self-determination theory would predict sustains motivation through a period in which everyone’s expertise is partially reset (Ryan & Deci, 2000). The second condition is permission to abandon inherited proxies. Nelson (2010) documented how faculty defend practices as rigor that function mainly as selection rituals, dysfunctional illusions that feel like standards. Generative AI has detonated several of those illusions at once, most visibly the unsupervised written artifact as proof of thinking. Redesign under these conditions is not a lowering of standards. It is the replacement of proxies that no longer measure what they claim with performances that still do, and a tool tuned to a learner’s zone of proximal development (Vygotsky, 1978) can make the redesigned path the one students prefer, because calibrated challenge is precisely the condition under which intrinsic motivation replaces avoidance.

Societal Stakes

Three consequences of the singularity cross the campus boundary, which is why they belong in a venue concerned with society rather than only with schooling.

The first is equity, and it inverts the usual framing. The consumer chatbot market is now openly tiered, with free, twenty-dollar, and two-hundred-dollar offerings whose capability gaps are widening. Left to individual purchasing, coupled learning certifies subscription price: the affluent student reasons beside a frontier model while the low-income student reasons beside a throttled one. The institutional answer is not prohibition, which levels down, but provisioning up: supplying every student a tier of capability no individual would rationally buy alone, which converts the divide into an equalizer. Provisioning up, though, equalizes only inside the institution’s walls and only while enrollment lasts. Beyond them, access to capability will likely stratify the way access to clean water, adequate nutrition, and healthcare already does, with socioeconomic status setting the tier. The realistic future is thus not the one the popular narrative imagines, in which a frontier intelligence teaches everyone, but one in which most learners work beside cheaper, weaker AI, which makes the ability to extract strong results from a limited tool a decisive competence in its own right. That competence has a trainable surface, the craft of prompting and workflow, and an educational depth: framing a problem so a modest model can contribute, detecting where it errs, and supplying what it lacks, which are the same capacities of judgment, verification, and foundational knowledge whose formation Sections 3 and 4 argue must be protected from premature automation. Equity and sustainability converge on it, and teaching it is a central charge of the educator and the course architect. Employer demand for such fluency is documented; in Microsoft and LinkedIn’s 31-country survey, 66 percent of leaders said they would not hire a candidate without AI skills (Microsoft & LinkedIn, 2024).

The second is the credential. A degree used to refer to a person. It increasingly refers to a person plus a configuration of tools, and the configuration is recorded nowhere, which leaves every transcript ambiguous on the point an employer most needs. Aviation suggests the remedy, and the argument, which I develop fully in forthcoming work, can be previewed in outline. A pilot’s certificate has never claimed that its holder can fly, full stop; a type rating names the aircraft and an instrument rating names the conditions, so the credential certifies a human-machine configuration explicitly, and the precision confers status rather than subtracting it. The record I propose does the same for learning: it states what its holder can do with assistance, after assistance, and without it. Higher education already makes configuration statements at exam scale every time a professor writes closed book on a test; the academic-record implication of the pedagogical singularity is that the statement must scale from the syllabus to the credential.

The third is relational, and it is the least governed. The same habits of disclosure that institutions may normalize in the learning layer, especially with systems experienced as patient, available, and nonjudgmental, can extend beyond institutionally sanctioned tools to AI companions that no campus office has vetted. Survey evidence suggests that AI companion use among teens is already widespread: 72% report having used AI companions, 52% are regular users, 33% use them for social interaction or relationships, and 33% of users report discussing something important or serious with an AI companion instead of a real person (Robb & Mann, 2025). Most teens still prioritize human friendship, but the governance problem is not whether companions replace friends wholesale; it is that relational dependence, disclosure, and substitution are already measurable at meaningful scale. A society-centered research program should treat this tutor-to-companion gradient as a governance object in its own right.

A Research and Design Agenda

The frame that unifies these threads is sustainability. Cognitive capital, the finite and partially renewable stock of attention, effortful processing, and unaided capability that learners hold, can be spent, conserved, or grown by a coupled system, and the difference is decided by design rather than by the model’s intelligence. Capital of this kind is built the way any capital is, through the sequenced investment of effort. Productive struggle is the deposit: the friction a learner works through before the model is allowed to close the gap is what turns spent attention into durable capacity, and scaffolding that forms initial structure first, with AI entering afterward to extend that structure rather than supply it, is what makes the investment yield a return. Sequence and the protection of structural skills, on this view, are not variables separate from sustainability but the mechanism of it: the timing that ensures the effort is made, and the safeguard that ensures it forms something. A sustainable configuration produces residue: durable capacities that survive the tool’s removal. An extractive configuration produces the debt condition now visible in the empirical record. Measurement should keep three layers analytically separate: the location of the threshold, which is a question about configuration; coupled gains, what the learner-and-tool system produces while assistance is present, Salomon’s effects with; and residue, the capability that remains when assistance is withdrawn, the effects of. Better artifacts demonstrate the second. Only the third demonstrates growth rather than debt. Neither settles the first.

Four lines of work follow, and their order matters. The dependency map comes first, because sequence thresholds and residue measures cannot be interpreted without a prior classification of which practices are structural and which terminal; absent that map, any observed decline in unaided skill can be relabeled terminal after the fact, and the framework can never fail. Chart the map beginning where the evidence already gestures, in writing and early computation. Then establish sequence thresholds: at what points in a task, and at what points in development, may assistance enter without foreclosing formation. Build condition-explicit records so that credentials state what configuration they certify. And govern the relational seam between endorsed learning tools and unendorsed companions before the gradient produces its first institutional casualty. The window for this work is closing from an unusual direction: as cohorts arrive already coupled from secondary school, uncoupled baselines become unrecruitable, and the comparison condition on which the agenda depends acquires an expiration date. The through-line is a single question: when does technology amplify human learning, and when does it displace it? Work on this agenda should proceed under the emerging norms for responsible AI use in education research (Smith & McGill, 2026).

Conclusion

Relocating the singularity from the machine to the institution changes both the diagnosis and the stakes. It clarifies that the central issue is not whether AI can perform educational tasks, but whether institutions can organize human-AI coupling in ways that widen access, protect capacity formation, and make their credentialing claims legible to society. The shift can be stated in a line: for as long as education has existed, the difficulty of a task supplied its own friction, and the effort of working through it built the capacity; executable language dissolves that friction natively, so the institution must now supply it by design, through sequence, reserved practice, and the protection of structural skills. The pedagogical singularity, on this account, is not a future event to await. It is a governance condition already emerging unevenly across courses, programs, and campuses.

The practical implications are immediate. Policy must move from blanket permission or prohibition toward staged forms of use; research must identify the practices whose delegation carries developmental cost; and institutions must address the relational spillovers that connect endorsed learning tools to unendorsed companion systems. The question is no longer whether education will couple with AI. The question is what kind of coupled system higher education is willing to become.

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), Article e2422633122. https://doi.org/10.1073/pnas.2422633122

Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19.

Emig, J. (1977). Writing as a mode of learning. College Composition and Communication, 28(2), 122–128.

Hembree, R., & Dessart, D. J. (1986). Effects of hand-held calculators in precollege mathematics education: A meta-analysis. Journal for Research in Mathematics Education, 17(2), 83–99.

Hutchins, E. (1995). Cognition in the wild. MIT Press.

Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, Article 17458. https://doi.org/10.1038/s41598-025-97652-6

Klein, C. R., & Klein, R. (2025). The extended hollowed mind: Why foundational knowledge is indispensable in the age of AI. Frontiers in Artificial Intelligence. https://doi.org/10.3389/frai.2025.1719019

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv. https://doi.org/10.48550/arXiv.2506.08872

Machajewski, S. (2024a). AI Instructional Framework: A comprehensive approach to learning with AI. University of Illinois Chicago. https://doi.org/10.25417/uic.26181374

Machajewski, S. (2024b). English 2.0: AI-driven language transformation. EDUCAUSE Review.

Machajewski, S. (2026a). The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI (Complete ed.). Data II Press.

Machajewski, S. (2026b). Red light, yellow light, green light: What 1,187 faculty decided about AI in their classrooms. Community College Journal, 96(4).

Microsoft, & LinkedIn. (2024). AI at work is here. Now comes the hard part: 2024 Work Trend Index annual report. https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part

Nelson, C. E. (2010). Dysfunctional illusions of rigor: Lessons from the scholarship of teaching and learning. To Improve the Academy, 28(1), 177–192. https://doi.org/10.1002/j.2334-4822.2010.tb00602.x

Robb, M. B., & Mann, S. (2025). Talk, trust, and trade-offs: How and why teens use AI companions. Common Sense Media.

Ryan, R. M., & Deci, E. L. (2000). Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American Psychologist, 55(1), 68–78.

Salomon, G., Perkins, D. N., & Globerson, T. (1991). Partners in cognition: Extending human intelligence with intelligent technologies. Educational Researcher, 20(3), 2–9.

Smith, J. M., & McGill, M. M. (2026). Guidelines for the responsible use of AI in STEM education research: Protecting our participant communities and research integrity. Institute for Advancing Computing Education. https://iacomputinged.org/graiser

Strittmatter, A., Sunde, U., & Zegners, D. (2020). Life cycle patterns of cognitive performance over the long run. Proceedings of the National Academy of Sciences, 117(44), 27255–27261. https://doi.org/10.1073/pnas.2006653117

Vinge, V. (1993). The coming technological singularity: How to survive in the post-human era. Whole Earth Review, Winter.

Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press.

About the Author

Szymon Machajewski is Director of Academic Technology and Learning Innovation at the University of Illinois Chicago and author of The Learn-It-All Educator: A Guidebook for Training Brains, Not Replacing Them with AI. With more than 20 years of experience in higher education, he writes and presents on AI in education, online pedagogy, student support, and gameful learning, including The Well-Played Class: A Case Study on How Gaming the System Motivates Students and Restores the Joy of Teaching. Correspondence: szymonm@uic.edu

Cite this article

Machajewski, S. (2026). Beyond model capability: The pedagogical singularity as institutional reorganization. Society and AI. https://societyandai.org/perspectives/beyond-model-capability-pedagogical-singularity/

Write for Society & AI

Have a perspective on what AI means for education, knowledge, or human flourishing? We’d love to hear it. Society & AI publishes freely, in open access, for anyone thinking seriously about these questions. Send your proposal to submissions@societyandai.org.

I have tried to write in the spirit of Berterö’s (2016) guidance on the commentary as a scholarly form, which counsels against re-summarizing the focal article, assumes the reader has it in hand, and asks instead for a clear take-home message. I will keep to that, and so I begin not with a recap but with the one feature of the report I have found hardest to set aside. In its opening section, the authors include a set of instructions addressed not to their human readers but to the AI assistants those readers will likely ask to summarize the paper. They specify what should be emphasized, what should not be compressed into shorter lists, and they invite future AI systems to update the report’s conclusions as events unfold.

It is worth pausing on that gesture. Here is a scholarly report written partly for the technology it describes, anticipating its own obsolescence and asking that technology to help keep it current. Whatever one ultimately makes of its specific claims, that framing alone says something about the moment we are working in — and it is the moment, more than any single forecast, that this commentary is concerned with.

Education: Teaching Toward a Moving Target

This is where the report unsettled me most. We design curricula, credentials, and pedagogical sequences around a relatively stable assumption: that human cognitive development follows predictable stages, and that what we teach today will remain relevant for the working life of the student in front of us. The report’s central claim — that the gap between human and machine capability may not just grow, but grow in kind, through mechanisms (lossless replication, substrate independence, arbitrary timescales) that have no human analog — puts that assumption under real pressure.

I do not read this as education becoming obsolete. If anything it becomes more important, but differently shaped. If machines increasingly hold the encyclopedic and procedural knowledge that schooling has long been organized around delivering, then what remains distinctly valuable is judgment: knowing which questions matter, how to evaluate an answer, how to live alongside systems whose internal workings we cannot fully inspect. Neil Lawrence’s argument, which the report cites, is the lever here: human cognition’s narrow communication bandwidth forces us to build abstractions, hierarchies, and compressed models of the world. That forcing function — the very limitation we sometimes treat as a deficiency — may be part of what makes human understanding transferable, teachable, and shareable in the first place. An education that helps students build and interrogate their own abstractions, rather than simply consume larger ones, looks more durable than ever.

Systems, Complexity, and the Limits of Forecasting

The report treats AI progress not as a date to be predicted but as a system with competing feedback loops — accelerating dynamics (AI speeding up AI research) racing against decelerating ones (the rising cost of each additional unit of progress) — and concedes that which loop dominates is, at the onset, nearly impossible to measure.

This is a systems-thinking lesson that extends well beyond AI, and it is the one I would press furthest. Societies routinely fail to prepare for transitions not because they cannot imagine them, but because exponential and hyperbolic dynamics are experientially indistinguishable from linear ones until very late. The report’s prescription — track quantitative indicators continuously, maintain a range of models rather than a single forecast, and revisit them often — is sound advice for any institution facing a system it cannot fully model, which describes most of the institutions I work with. I would add a caution the report leaves implicit: the instruments we choose to track will quietly determine the futures we are able to see. Measurement is never neutral observation; it is a commitment about what counts.

Knowledge, Intelligence, and What “Super” Means

The report’s use of the Legg–Hutter framework — defining intelligence as average performance across all computable tasks — is, to me, less a technical detail than a philosophical provocation. It reframes intelligence not as a trait an entity has, but as a measure of breadth across an unbounded space of problems. Under that lens, the question “is this system superintelligent?” dissolves into “superintelligent compared to whom, on what, and at what cost?” — questions that are inseparable from values, governance, and access. Whoever defines the distribution of tasks defines the winner. That is not a measurement problem to be cleaned up later; it is a question of values embedded in the metric from the start, and education is one of the few institutions positioned to teach people to recognize and ask it.

What a Time to Be Alive

I do not write any of this from a place of alarm, nor from uncritical excitement. I write it from something closer to vertigo — the feeling of standing at a point where multiple plausible futures genuinely diverge, and where the report itself admits it cannot tell us which one we are entering. That admission, paradoxically, is what makes the paper worth taking seriously, and worth re-reading, as the authors themselves suggest, with help from the very systems it describes.

So here is my take-home message, in Berterö’s sense of the term. The most useful response to a future we cannot specify is not a better prediction; it is an education that builds judgment, interrogates its own abstractions, and treats the measurement of intelligence as a question of values rather than a settled fact. We, all of us working in education and the social sciences, are being asked to prepare students, institutions, and ourselves for a range of futures we cannot specify in advance. That is a genuinely strange position to be in. It is also, I think, a privilege: to be present, attentive, and working, at a moment when the questions are this large and this open.

About the Focal Article: From AGI to ASI

The following is a narrative summary of the report this commentary responds to. In keeping with the guidelines above, it sits outside the commentary proper — included here so readers who have not yet read the original have the full picture in one place.

Genewein, Franklin, Lerchner, and colleagues (2026) begin from a premise that has, in their words, moved “from far-fetched speculation to being a concrete next-decade target for many of the largest AI organisations”: the arrival of human-level artificial general intelligence (AGI), systems that match a typical human across most cognitive tasks. Their real question is what happens after that point — the transition from AGI to artificial superintelligence (ASI), which they characterize as systems surpassing not just individuals but large organizations of humans in cognitive capability.

The report maps four roughly parallel pathways toward ASI. The first is continued scaling of compute, data, and models. The second is a shift in the underlying algorithmic paradigm — a new approach that supersedes today’s methods. The third is recursive self-improvement, in which AI systems accelerate AI research itself. The fourth is the emergence of intelligence from large collectives of AI agents coordinating the way corporations or institutions do.

Against each pathway the authors weigh frictions and bottlenecks: the eventual exhaustion of high-quality training data, the economics and energy demands of continued scaling, the difficulty of automating physical-world experimentation, and what they call the “Abstraction Barrier.” Whether these frictions prove negligible or substantial, they argue, remains an open research question — and one that cannot be settled from where we currently stand.

A distinctive section catalogs the structural advantages that digital intelligence holds over biological intelligence — advantages that grow as computers become more powerful. AI can be copied losslessly, not just its code but its accumulated “experience.” It can be paused, sped up, run in parallel, and moved across hardware. None of this requires AI to become smarter in any single sense; it requires only more compute, applied to something that was never bound by a body in the first place.

To define intelligence itself, the authors adopt the Legg–Hutter framework, which measures intelligence as performance averaged across all computable tasks. The report is deliberately uncertain throughout: it does not predict when, or whether, ASI arrives, and it treats AI progress as a system of competing feedback loops whose net direction is hard to read at the onset. Its central recommendation is therefore procedural rather than predictive — track quantitative indicators continuously, hold a range of models rather than a single forecast, and revisit them as evidence accumulates. The authors frame the whole prospect not as a single step change but as “a series of transformative societal changes,” requiring “a massively interdisciplinary endeavour of global scope and interest.”

References

This commentary is based on a close reading of the following report:

Genewein, T., Franklin, M., Lerchner, A., Orseau, L., Albanie, S., Bales, A., Wyeth, C., Chan, S., Gabriel, I., Leibo, J. Z., Dafoe, A., Hutter, M., Graepel, T., & Legg, S. (2026). From AGI to ASI. arXiv. https://arxiv.org/abs/2606.12683

Other works referenced above:

Berterö, C. (2016). Guidelines for writing a commentary. International Journal of Qualitative Studies on Health and Well-being, 11(1), 31390. https://doi.org/10.3402/qhw.v11.31390

Lawrence, N. D. (2024). The atomic human: Understanding ourselves in the age of AI. Allen Lane.

Legg, S., & Hutter, M. (2007). Universal intelligence: A definition of machine intelligence. Minds and Machines, 17(4), 391–444. https://doi.org/10.1007/s11023-007-9079-x

Cite this article

Gattupalli, S. (2026). On “From AGI to ASI”. Society and AI. https://societyandai.org/commentary/from-agi-to-asi-commentary/

Write for Society & AI

Have a perspective on what AI means for education, knowledge, or human flourishing? We’d love to hear it. Society & AI publishes freely, in open access, for anyone thinking seriously about these questions. Send your proposal to submissions@societyandai.org.

Artificial intelligence (AI) is supposed to democratise education. The reality, as we are beginning to witness across higher education, is considerably more uncomfortable.

As academics situated across two very different institutional contexts, one in the United Kingdom and one at an Australian university campus in Malaysia, we have followed with close interest how AI is entering classrooms across higher education in uneven and sometimes contradictory ways. What we observe is not simply a divide in access to technology. It is something subtler and, we would argue, more consequential: a divide shaped by institutional readiness, academic culture, staff confidence, and policy priorities.

This distinction matters. Many students can already access generative AI through free or low-cost platforms. But access alone does not create meaningful capability. AI systems themselves are not neutral. They reflect the assumptions, language norms, and cultural contexts of their developers, which may advantage some learners more than others from the outset. Indeed, the trajectory of AI in education has long pointed towards personalised, competency-based learning, yet realising that potential equitably remains an unresolved challenge (Dastane et al., 2024).

Across universities globally, responses to AI are unfolding at very different speeds. Some institutions are integrating AI into teaching, assessment, and curriculum design. Others are proceeding cautiously because of concerns around academic integrity, reliability, or governance. A student studying in an institution with clear AI guidance, thoughtfully redesigned assessments, and staff trained in AI-supported pedagogy is likely to develop a fundamentally different kind of academic confidence from a student where AI remains largely unspoken or inconsistently regulated. This unevenness in institutional adoption is not a new observation; systematic reviews of AI applications in higher education have documented significant variation across institutions and disciplines for some years (Zawacki-Richter et al., 2019), and the gap appears to be widening rather than closing.

Increasingly, success in higher education may depend not only on subject knowledge, but on what we might call AI literacy: the ability to work with AI tools thoughtfully, question outputs critically, and integrate them into learning without becoming dependent on them. This is not a peripheral skill. It is becoming central to how knowledge work itself operates. Bearman and Ajjawi (2023) describe this as learning to work with the black box, arguing that pedagogy must now equip students to engage critically with AI outputs rather than simply consume them. That framing resonates strongly with what we observe across institutional contexts.

Here lies an uncomfortable tension, and one we feel is worth naming directly.

Universities have historically acted as institutions of careful validation. Academic systems move deliberately because quality assurance, fairness, and intellectual standards matter. AI, however, evolves at a pace that sits uneasily alongside these slower institutional rhythms. Policies often take years to stabilise; AI tools change within months.

Many educators are trying to teach on moving ground.

We have seen this play out in small but telling ways. A colleague redesigns an assessment to encourage genuine critical engagement with AI outputs, only to find that institutional policy has not yet caught up with what she is attempting. A student experiments independently with AI tools for an assignment, uncertain whether doing so crosses a line that has never been clearly drawn. These are not dramatic failures. They are the quiet, everyday signs of a system under pressure to adapt faster than its structures comfortably allow.

Some faculty members have institutional support, training, and time to experiment with AI-enhanced teaching. Others are expected to respond to rapid technological change without sufficient guidance or resources. The result is that students receive mixed messages: use AI cautiously; engage with it, but do not rely on it; develop AI capabilities, but avoid crossing unclear boundaries that shift depending on the course, the lecturer, or the institution.

This inconsistency risks producing a new kind of educational inequality, one not immediately visible through league tables or infrastructure, but through differing levels of preparedness for AI-mediated professional life.

Importantly, this divide does not follow geography in any simple sense. Within the same country, some universities are moving rapidly while others remain hesitant. Within the same institution, disciplines often take markedly different approaches. Business and computing programmes may integrate AI extensively, while humanities and social sciences continue to debate its legitimacy in assessment altogether. These are not unreasonable differences of perspective. But they do raise questions about whether students are graduating with comparable readiness for workplaces where AI is already present.

The deeper policy question is whether higher education systems are treating AI primarily as a threat to academic integrity or as a genuine transformation in how learning itself occurs.

Concerns about misuse are valid and should not be dismissed. AI can encourage superficial learning if used uncritically. Students with stronger prior knowledge, higher digital confidence, and greater familiarity with the language and cultural assumptions embedded within AI systems are likely to benefit more. Algorithmic bias and cultural mismatch may therefore reinforce existing inequalities rather than reduce them. But restriction alone is unlikely to solve this. It may simply drive usage underground, without the guidance that makes it educationally meaningful.

This is why institutional policy now matters enormously. Universities need frameworks that move beyond the binary of allowing or banning. Students need clearer expectations around when AI use is appropriate, how it should be acknowledged, and what forms of learning still need to happen independently. Emerging guidance from sector bodies, including the QAA and TEQSA, reflects a growing recognition that responsible integration requires practical frameworks rather than purely restrictive approaches. At a global level, UNESCO (2023) guidance for generative AI in education makes precisely this case, calling for policy responses that are contextually sensitive, ethically grounded, and oriented towards equity rather than restriction alone. That direction of travel is encouraging, even if implementation remains uneven.

The risk, as we see it, is not that universities will ignore AI entirely. The risk is that adaptation will happen unevenly, and that this unevenness will quietly shape whose ways of thinking, communicating, and learning are most readily supported by the AI systems graduates will encounter in professional life. Some students will leave university already fluent in working alongside these tools. Others will leave with strong disciplinary knowledge but limited experience of how AI is reshaping the environments they are about to enter.

Higher education has long been understood as a mechanism for reducing inequality through access to knowledge. In the AI era, that mission may need to extend further: towards ensuring equitable access not only to knowledge, but to the conditions in which AI-augmented learning can happen responsibly, critically, and well.

The question for universities and policymakers is not whether to engage with AI. It is whether higher education systems can shape that engagement thoughtfully and equitably, before the divides that are quietly forming become very difficult to close.

References

Bearman, M., & Ajjawi, R. (2023). Learning to work with the black box: Pedagogy for a world with artificial intelligence. British Journal of Educational Technology, 54(5), 1160–1173. https://doi.org/10.1111/bjet.13337

Dastane, O., Turner, J., & Nankervis, A. (2024). The trajectory of artificial intelligence for competency-based personalised learning: Past, present and future. The International Journal of Information and Learning Technology, 41(5), 473–489. https://doi.org/10.1108/IJILT-11-2023-0187

UNESCO. (2023). Guidance for generative AI in education and research. UNESCO Publishing. https://doi.org/10.54675/PCSP7350

Zawacki-Richter, O., Marín, V. I., Bond, M., & Gouverneur, F. (2019). Systematic review of research on artificial intelligence applications in higher education: Where are the educators? International Journal of Educational Technology in Higher Education, 16(1), Article 39. https://doi.org/10.1186/s41239-019-0171-0

About the Authors

Ria Sidhu SFHEA is a member of the Academic Partnerships & TNE team at Bath Spa University, UK, where she leads Assessment Quality Assurance across MSc Global Business, MSc Project Management, the undergraduate Business Management suite (all pathways), and the postgraduate MBA programmes for UK and UAE partners. Drawing on 15 years in UK higher education, her expertise spans higher education policy and governance, quality assurance, transnational education, and EDI-driven pedagogy. Correspondence: r.sidhu@bathspa.ac.uk

Omkar Dastane^* is a Senior Lecturer in Marketing and Director of Postgraduate Studies at the School of Business, Monash University Malaysia. His research examines how emerging technologies are reshaping consumer value hierarchies and redefining marketplace dynamics. A passionate education leader, Dr. Omkar champions immersive learning through AR, VR, and AI, and actively shapes institutional policy on the responsible integration of artificial intelligence in higher education. Correspondence: omkar.dastane@monash.edu

^* Corresponding author

Cite this article

Sidhu, R., & Dastane, O. (2026). The emerging AI divide in higher education. Society and AI. https://societyandai.org/perspectives/the-emerging-ai-divide-in-higher-education/

Write for Society & AI

Have a perspective on what AI means for education, knowledge, or human flourishing? We’d love to hear it. Society & AI publishes freely, in open access, for anyone thinking seriously about these questions. Send your proposal to submissions@societyandai.org.

I read two reports from Anthropic within days of each other this spring, and together they said something neither said alone. The first, Policy on the AI Exponential, opens with a warning aimed at governments: "AI is advancing at exponential speed, and the policymaking process was built for a slower world" (Anthropic, 2026a). The second, When AI Builds Itself, documents the company's own progress toward recursive self-improvement — the point at which an AI system could fully design and develop its successor, with humans supplying little more than the initial goal (Anthropic, 2026b).

Both documents are candid about the stakes. Both are also, worth noting, written by one of the companies setting the pace they describe. That is not a contradiction so much as a position: Anthropic is asking the world to help govern a race it is also running. The honesty is real. So is the race.

The Exponential, By Its Own Account

The numbers in When AI Builds Itself are striking on their own terms. As of May 2026, more than 80% of the code merged into Anthropic’s codebase was written by Claude, up from low single digits before February 2025. On the most open-ended engineering tasks — the kind with no clear specification, where even a skilled person isn’t sure what success looks like — Claude’s success rate reached 76%, up fifty percentage points in six months (Anthropic, 2026b). The report sketches three possible futures: the trend stalls and today’s capabilities simply diffuse outward; AI development becomes substantially automated while humans retain “research taste”; or AI systems become capable of fully recursive self-improvement, with humans shifting toward “oversight, validation, and verification of an expanding virtual lab” (Anthropic, 2026b).

The companion policy paper takes that trajectory as established fact and asks what governments should do about it: require transparency, fund independent evaluators, and grant authority to block deployments that cross catastrophic-risk thresholds (Anthropic, 2026a). It is a serious proposal, and a necessary one. But it is also a proposal written entirely in the grammar of speed — how fast AI is improving, how fast governance must adapt to keep up. What gets less attention, in either report, is a quieter question: improving toward what, and for whom?

When AI Builds Itself — For Whom?

Nearly seventy years ago, Hannah Arendt distinguished between three kinds of human activity: labor, the endless cycle of producing what we consume to stay alive; work, the durable making of a world of things; and action, the speech and deeds through which people appear to one another as who they are (Arendt, 1958). Arendt worried about a society that automates labor without asking what becomes of work and action — a “society of laborers” that has been freed from labor just as labor is the only thing it still knows how to value.

When AI Builds Itself describes something close to that scenario, but for knowledge work. “The doing — i.e., writing the code, running the experiment, producing the result — now costs almost nothing in human time, even if it still has costs in compute,” the report notes, and an Anthropic employee is quoted: “On days where everything works well, I can’t help but think nothing I do matters” (Anthropic, 2026b). That is not a complaint about job loss in the usual sense. It is closer to Arendt’s worry: when the doing is automated, what is left of the human role is direction-setting — and direction-setting, done well, requires exactly the kind of judgment, taste, and lived stake in outcomes that cannot be outsourced to the system being directed.

Racing for AI’s Sake

This is where a second piece of recent scholarship becomes useful, even though it is not, on its face, about AI at all. Susi Geiger’s analysis of tech oligarchs’ “obsession with pronatalism and fertility markets” describes a recurring pattern: an elite circle conjures an existential narrative — population collapse, civilizational extinction — that feels urgent and inevitable, and that narrative then channels enormous capital toward speculative technologies controlled by that same circle (Geiger, 2026). The narrative and the market reinforce each other. Whether the underlying claim is true matters less than whether the belief in it can be sustained long enough for the investments to pay off.

I am not arguing that recursive self-improvement is a hoax — the evidence Anthropic presents is real, and I take it seriously. But the framing of “the exponential” as an autonomous force, something AI is doing to us rather than something a small number of companies and investors are choosing to build, deserves the same scrutiny Geiger applies to fertility markets. “AI building AI” can become its own kind of inevitability narrative: one that makes the current pace feel like physics rather than policy, and that quietly forecloses the question of who gets to slow down, redirect, or say no. Three of the four futures Anthropic itself sketches involve a world reorganized around the needs of an AI development pipeline. Only one of them centers on the people that pipeline is supposed to serve.

What Personal Computing Promised

This is not a new question, even if it is being asked at a new scale. Since the dawn of personal computing, the measure of a technology has never really been how powerful it became. It has been whether that power turned into something helpful, intuitive, and meaningful — for a teacher in an under-resourced school, a small business owner, a retiree, a student whose first language is not the one the interface was designed in. AI is among the most powerful technologies our species has ever built, and with proper care it could unlock real benefit for people everywhere, across every economic circumstance, every culture, every faith. That potential is genuine. But potential is not the same as a guarantee, and a system racing toward self-improvement does not automatically arrive anywhere near “helpful” — it arrives wherever its builders, deliberately or by default, point it.

Toward Society-Centered AI

This is the premise Society & AI was built on, and it has not changed: that truly helpful AI must be centered on people and their needs, not on the trajectory of AI itself. Anthropic’s reports are valuable precisely because they make the trajectory visible — and because they ask, honestly, whether anyone outside the small set of frontier labs gets a meaningful say in where it goes. Our answer is that they must. Society-centered AI research exists to make that “must” concrete: to ask, of every benchmark and every capability gain, not only can it do this but who is this for, and to keep that question in the room for as long as the exponential keeps climbing.

References

Anthropic. (2026a). Policy on the AI exponential. https://www.anthropic.com/policy-on-the-ai-exponential

Anthropic. (2026b). When AI builds itself. The Anthropic Institute. https://www.anthropic.com/institute/recursive-self-improvement

Arendt, H. (1958). The Human Condition. University of Chicago Press. https://archive.org/details/dli.ernet.528547

Geiger, S. (2026). Tech oligarchs’ obsession with pronatalism and fertility markets. Science as Culture. https://doi.org/10.1080/09505431.2026.2639478

Cite this article

Gattupalli, S. (2026). Built for AI, or built for us? Society and AI. https://societyandai.org/perspectives/built-for-ai-or-built-for-us/

Write for Society & AI

Society & AI welcomes scholarly contributions from researchers, educators, librarians, policymakers, and practitioners working at the intersection of artificial intelligence, education, and society. We publish accessible, rigorously argued work that advances public understanding of how AI reshapes institutions, learning, and human experience — and every piece is published free of cost and in open access, in service of education and the public good. If you have a perspective, commentary, or research piece to propose, we would welcome hearing from you: write to the Editorial and Content Director at submissions@societyandai.org.

The numbers come fast. Hundreds of billions of parameters. Millions of tokens per second. Benchmark after benchmark. Every few months, a new model is announced as the most capable ever released, and within weeks another arrives to take its place. The competition between the largest technology companies in the world, among them OpenAI, Google, Meta, and Anthropic, has settled into a familiar rhythm: release, outperform, repeat.

What no one in that race has stopped to ask, in any sustained or binding way, is what this does to people.

Society & AI was built on a simple premise: that the consequences of artificial intelligence for human beings deserve as much serious attention as its technical capabilities. Our work is organized around four areas of inquiry. The first is knowledge and intelligence, examining how AI transforms the creation and validation of knowledge and what genuine human-machine collaboration in learning might look like. The second is systems and complexity, mapping education as a complex adaptive system and identifying the leverage points where technology might catalyze meaningful change. The third is educational equity, positioning quality learning not as a policy goal but as a compounding capability that creates the conditions for addressing every other global challenge we face. The fourth is human flourishing, investigating how agency, meaning, and dignity are preserved as cognitive labor becomes increasingly automated. I came to these questions through years of research and teaching, and through witnessing, repeatedly, what happens when AI enters communities that were never consulted about how it was designed.

Each year, I serve as a peer reviewer for papers submitted to the annual meeting of the American Educational Research Association (AERA), one of the largest gatherings of education researchers in the world. Among the many scholarly divisions under which researchers submit their work is one titled Religion and Education, a division I review for. This is not incidental. The fact that one of education research’s most rigorous professional bodies formally recognizes this connection reflects something important: religion and education are not separate domains but deeply entangled ones, shaping the values and institutions through which human beings learn to be in the world. That entanglement is part of what moved me to write this piece.

On May 15, 2026, an unexpected voice entered that space with unusual force.

The Church Speaks to “All People of Good Will”

Pope Leo XIV released Magnifica Humanitas [roughly, “the magnificence of humanity”], a 42,300-word encyclical addressed not to Catholics alone, but to “all people of good will” (Leo XIV, 2026). That phrase is deliberate. What follows is not a theological argument for the faithful. It is a moral argument for a world facing a shared transformation.

The historical parallel is striking. His namesake, Pope Leo XIII, published Rerum Novarum [roughly, “on new things”] in 1891, a document that addressed worker dignity and the dangers of unchecked capital at a moment when industrial power was reshaping society faster than governance could follow. One hundred and thirty-five years later, Leo XIV has done the same for the digital revolution. Rerum Novarum did not stop industrialization. It helped establish a moral vocabulary of solidarity, common good, and dignity that shaped labor policy for generations. Magnifica Humanitas is attempting the same work for AI.

The Babel Problem

The encyclical opens with two biblical images: the Tower of Babel and Nehemiah’s rebuilding of Jerusalem. The Babel story, of a civilization so consumed by self-assertion and uniformity that communication breaks down, is not framed as ancient history. It is framed as a live risk.

The diagnosis is accurate. A small number of companies, operating within a narrow range of cultural and commercial frameworks, are building systems that will shape what billions of people know, believe, and are permitted to do. Their shared assumption, that progress in benchmark performance constitutes progress for humanity, goes largely unchallenged. The diversity of voices that might complicate it is absent from the rooms where decisions are made. That is Babel, not by moral failure alone, but by structural design.

A Confession from Inside the Machine

What gave the encyclical’s release particular resonance was the presence of Christopher Olah, a co-founder of Anthropic, who appeared alongside the Pope at the presentation. His words were direct: “We need moral voices that the incentives cannot bend. Today is just the beginning, the start of a long collaboration between those who are building this technology and those who can see what they, from the inside, cannot.”

This is a significant thing for someone building frontier AI to say publicly: that the incentive structures governing AI companies, including competitive pressure, investor expectations, and quarterly performance, are not sufficient to produce the deliberation that technology at this scale requires.

What the Document Actually Demands

Magnifica Humanitas is not vague. It calls for government regulation of private AI companies, noting that these entities now command resources exceeding those of many nation-states. It demands protection for workers threatened by automation, calling mass forced inactivity alongside high technical development “a paradox of material progress and anthropological regression.” It calls for education that equips children to think critically about AI-generated content, and for guardrails ensuring humans retain responsibility for weapons decisions.

These are governance positions, not soft suggestions. And they come from an institution not optimizing for next quarter’s earnings.

Why This Matters Beyond Faith

I am not a theologian. I approach this as a researcher who has spent years studying what happens when AI systems enter communities never part of the conversation about how those systems were built. The vocabulary that Magnifica Humanitas brings, namely dignity, solidarity, and the common good, is not religious vocabulary in the narrow sense. It is moral vocabulary the AI industry does not yet have and urgently needs.

The encyclical insists that technology be judged not by what it can do, but by what it does to people. That is the standard Society & AI was built to apply. The AI companies will keep building. The benchmarks will keep climbing. What is needed now, and what has arrived unusually from Rome, is a moral architecture equal to the scale of what is being built. I am glad someone finally said it clearly.

References

Leo XIV, Pope. (2026, May 15). Magnifica humanitas: Encyclical letter on safeguarding the human person in the time of artificial intelligence. The Holy See. https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html

Leo XIII, Pope. (1891, May 15). Rerum novarum: Encyclical letter on capital and labor. The Holy See. https://www.vatican.va/content/leo-xiii/en/encyclicals/documents/hf_l-xiii_enc_15051891_rerum-novarum.html

Cite this article

Gattupalli, S. (2026). A moral architecture for the age of artificial intelligence. Society and AI. https://societyandai.org/perspectives/moral-architecture-age-of-artificial-intelligence/

Write for Society & AI

We welcome contributions from educators, researchers, librarians, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. Write to us at submissions@societyandai.org.

This paper argues that when AI runs offline on local devices in classrooms, the infrastructural constraints themselves become pedagogically meaningful. Drawing on observations from offline, on-device AI tutoring across secondary computing classrooms (2023–2025), it introduces a compact framework — Infrastructure as Curriculum — and derives implications for AI literacy and equitable learning technology design. Rather than viewing offline AI as a degraded substitute for cloud intelligence, the paper argues that bounded systems surface learning practices — verification, authorship, and calibrated trust — that are often muted by always-available tools.

Abstract

Most AI-in-education deployments assume stable internet access and cloud-based computation, treating infrastructure as a neutral backdrop rather than a pedagogical force. This short paper offers a conceptual contribution rather than a new empirical study. We reframe infrastructural constraint as an instructional signal: when AI runs offline on local devices, students encounter visible limits, negotiate uncertainty, and engage in iterative troubleshooting and collaborative sensemaking. Drawing on observations from offline, on-device AI tutoring across multiple secondary computing contexts, we articulate a compact framework — Infrastructure as Curriculum — and derive design implications for AI literacy and equitable learning technology integration. Rather than viewing offline AI as a degraded substitute for cloud intelligence, we argue that bounded systems can surface learning practices — verification, authorship, and calibrated trust — that are often muted by always-available tools.

Open Access Notice

This paper was accepted as a short paper at the 26th IEEE International Conference on Advanced Learning Technologies (ICALT 2026), Track 1: Technologies for Open Learning and Education (i-OPENLearn). Due to the absence of institutional affiliation and funding support, the author was unable to register or attend, and accordingly withdrew the submission from the conference proceedings.

In the spirit of open science, the work is released here as open access — freely available to educators, researchers, developers, and policymakers — with the hope that it will be read, critiqued, adapted, and built upon. Institutional affiliation is not a prerequisite for meaningful scholarly exchange.

A Note from the Author

I am releasing this paper openly because I believe the ideas in it belong to the community of educators and researchers thinking seriously about AI in classrooms — not to a paywall or a conference badge I could not afford. Many PhD graduates, including those from well-resourced R1 institutions, run into the same wall after defense: without an institutional line item, even an accepted paper can quietly become an unpublished one. This release is one small attempt to demonstrate that scholarly contribution does not require institutional backing.

The argument itself is modest. When AI works offline, its limits become visible — and visible limits ask something of learners that frictionless cloud systems do not. Students start to verify, to author, to trust calibratedly. Infrastructure, in that sense, becomes part of the curriculum. The paper sketches a framework for taking that pattern seriously, and offers a few design principles for educators and developers building bounded AI for real classrooms.

If you find any of this useful, or wrong, I would genuinely like to hear from you. Adapt it, critique it, cite it, teach with it, build on it — the work is intentionally open so it can move. Correspondence is welcome at sai@societyandai.org.

This is not an accusation of bad intent. It is a structural observation: every AI system trained on human feedback inherits the cultural assumptions of its raters. Those assumptions shape not just what the system says, but what it considers worth saying at all.

Who Rates the Responses

The process of training AI behavior through human feedback is well understood in research circles. What gets far less attention is who, exactly, is doing that work. Large-scale rating and labeling operations rely heavily on contractors working in English, in a small number of countries, using rubrics that reflect particular ideas about what good communication looks like: that being direct is a virtue, that a confident answer signals accuracy, that a well-structured response with a clear argument and supporting evidence is simply better than one organized around community authority and relational context.

Consider what this means in practice. An AI system asked about land governance will score highly if it produces a clean, three-paragraph summary. A response that first establishes who, within a given community, holds the authority to speak on that question at all will likely be rated as incomplete or evasive. The first kind of answer reflects a Western approach to information: find the fact, state it clearly. The second reflects an Indigenous approach to knowledge: understand the governance structure before claiming to know anything. After training on millions of these preference judgments, the system learns that one way of knowing is correct and the other is a failure to communicate.

Recently, my colleagues and I proposed a conceptual framework at the Symposium on Artificial Intelligence for Cultural Heritage and Indigenous Futures (AI4CHIEF 2026), where we document that Indigenous knowledge systems are systematically treated as raw data rather than recognized as governance structures with their own authority and internal logic. Our forthcoming paper, Gattupalli et al. (2026), argues that Indigenous knowledge does not function like a dataset. It functions more like a governing document: it defines not just what is known, but who has the right to interpret it, how that interpretation must be validated within the community, and under what conditions any action can follow. When AI systems take the content of that knowledge and discard its governance layer, the result is not inclusion. It is distortion dressed as representation. I will update this article with a link to the full paper as soon as it is published.

When AI Gets It Wrong

The cultural narrowness described above produces real errors. AI systems have been observed correcting the spelling of proper names that are spelled correctly in their source languages, converting them into anglicized versions that are easier for English speakers to read. AI tutoring platforms have flagged explanations rooted in oral knowledge traditions as incomplete, because they do not include citations formatted for print-based academic culture. Language models have described ceremonial practices as “belief systems,” a label that implicitly contrasts them with factual knowledge, a framing that many communities would find not just inaccurate but disrespectful.

None of these outputs came from malicious intent. They came from training pipelines where the people doing the rating simply did not know what they did not know. The result is a technology that speaks with equal confidence about things it understands and things it has quietly misrepresented. The communities affected by those misrepresentations have no clear way to contest them. They cannot flag the training data or correct the model weights. They encounter the output and know, without any formal path to appeal, that something important has been lost.

Cultural Intelligence Is a Design Requirement

The central question is not what AI systems can produce. It is who holds the authority to decide what a valid, appropriate, or helpful response looks like. Right now, that authority sits with a small number of companies, operating within a small number of cultural frameworks, making decisions that affect billions of people who were never part of that conversation.

This matters well beyond communities whose knowledge has been directly misrepresented. A system calibrated within a narrow cultural range will, over time, teach all of its users to think within that range. It will reward the reasoning patterns it was trained to recognize and quietly disadvantage the ones it was not. It will reinforce, at global scale, the idea that one grammar of thought is simply correct.

Cultural intelligence, in this context, is not a soft skill or a nice addition. It is a design requirement. Building AI that genuinely serves a diverse world means building the institutions, feedback mechanisms, and governance structures that allow many communities to help define what the system is actually trying to measure. Until that happens, the cultures that calibrate AI will continue to do so. And everyone else will continue to adapt to a standard they never set.

References

Gattupalli, S., Chakravarty, P., Chakravarty, U., Chand, G., & Arroyo, I. (2026). Indigenous environmental knowledges as governance infrastructure for climate AI in climate resilience and adaptation. The Symposium on Artificial Intelligence for Cultural Heritage and Indigenous Futures 2026 (AI4CHIEF). (forthcoming)

Latulippe, N., & Klenk, N. (2020). Making room and moving over: Knowledge co-production, Indigenous knowledge sovereignty and the politics of global environmental change decision-making. Current Opinion in Environmental Sustainability, 42, 7–14.

Perera, M., Vidanaarachchi, R., Chandrashekeran, S., Kennedy, M., Kennedy, B., & Halgamuge, S. (2025). Indigenous peoples and artificial intelligence: A systematic review and future directions. Big Data & Society, 12(2).

Cite this article

Gattupalli, S. (2026). The cultures that calibrate AI. Society and AI. https://societyandai.org/perspectives/cultures-that-calibrate-ai/

Write for Society & AI

We welcome contributions from educators, researchers, librarians, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. Write to us at submissions@societyandai.org.

This working paper argues that education must move beyond tool use toward the cultivation of human authority over AI in consequential settings. It introduces intelligence management as a framework for directing, interrogating, calibrating, and governing AI outputs, reframes the teacher as an orchestrator of epistemic judgment, and outlines curriculum and equity shifts needed to prepare students not merely to work with AI, but to remain accountable for what is done with it.

Ethical Disclosures

This raises a few questions about the ethics of using AI in published work. Can authors claim AI-generated content as their own work? Should universities discourage AI use completely? I ultimately think that authors, as creators, should be given the opportunity to choose if they use AI as a tool. This does not mean that authors can enter a prompt into an LLM and send it in to their editor sight unseen. Rather than using AI as a co-author, authors can use AI as a research assistant and editor.

Kari Weaver, a librarian at the University of Waterloo, developed the Artificial Intelligence Disclosure (AID) Framework so authors can share in what capacity AI was used in their work. She suggests these potential uses of AI:

“Artificial Intelligence Tool(s): The selection of tool or tools and versions of those tools used and dates of use. May also include note of any known biases or limitations of the models or data sets.

Conceptualization: The development of the research idea or hypothesis including framing or revision of research questions and hypotheses.

Methodology: The planning for the execution of the study including all direct contributions to the study design.

Information Collection: The use of AI to surface patterns in existing literature and identify information relevant to the framing, development, or design of the study.

Data Collection Method: The development or design of software or instruments used in the study.

Execution: The direct conduct of research procedures or tasks (e.g. AI web scraping, synthetic surveys, etc.)

Data Curation: The management and organization of those data.

Data Analysis: The performance of statistical or mathematical analysis, regressions, text analysis, and more using AI tools.

Privacy and Security: The ways in which data privacy and security were upheld in alignment with the expectations of ethical conduct of research, disciplinary guidelines, and institutional policies.

Interpretation: The use of AI tools to categorize, summarize, or manipulate data and suggest associated conclusions.

Visualization: The creation of visualizations or other graphical representations of the data.

Writing—Review & Editing: The revision and editing of the manuscript.

Writing—Translation: The use of AI to translate text across languages at any point in the drafting process.

Project Administration: Any administrative tasks related to the study, including managing budgets, timelines, and communications.” (Weaver, 2024)

Ideally, authors will be comfortable including such a statement in their work. Disclosing AI use should not be a barrier to publication, but there is a stigma around AI use in some university cultures. The challenge will be to assure faculty that a disclosure statement is not just one more thing to do on their never-ending checklists.

How Open Is Open?

At their core, OER are meant to be openly available, reusable, and remixable. In my publishing program, authors have the opportunity to choose a Creative Commons license for their work. All of the licenses specify that any reuse of the work must attribute the original author. Some licenses are more open than others. For example, a CC BY-ND license allows a work to be openly available but it cannot be adapted or changed in any way.

AI throws some complications into the mix. How open is the Open Education movement? If authors are making their work openly available, are they also consenting to AI being trained on their work? Should authors receive compensation if LLMs are making money off their work? There are no easy answers to these questions. The downside of Open is that once something is out there, there is really no way to control how a work is used or adapted. With the rise of AI, there is a possibility that authors will start to gravitate toward No Derivatives licensing. Work will still be openly available but without the ability to remix or adapt. Librarians are in a unique position to educate authors about their rights and support them in how they want their work to be shared.

Conclusion

Library publishing programs are not immune to the ethical considerations that AI poses. Authors should not be discouraged from using AI as a tool for editing or organization. Having an AI disclosure statement as part of an OER upholds a spirit of transparency in publishing. But the more difficult questions librarians are facing now is how to support openness while recognizing authors’ rights. Faculty authors have an obligation to support student learning, but it remains to be seen if AI is entitled to the same consideration.

Reference

Weaver, K. (2024). The Artificial Intelligence Disclosure (AID) Framework: An Introduction. College & Research Libraries News, 85(10), 407. https://doi.org/10.5860/crln.85.10.407

About the Author

Melissa Chim is the Scholarly Communications Librarian at Excelsior University, where she manages the university’s scholarly publishing platform and institutional repository. She is a SPARC Open Education Leadership Fellow, a Society for Scholarly Publishing Fellow, and an ASAPbio Fellow and Resident for 2025–2026. Her work spans open access publishing, OER, Creative Commons licensing, and historical research.

Cite this article

Chim, M. (2026). AI and library publishing: A reflection. Society and AI. https://societyandai.org/perspectives/ai-library-publishing/

Write for Society & AI

Society & AI exists for scholars, educators, practitioners, and policymakers who believe the most important questions about artificial intelligence are not only technical — they are deeply human. We publish essays, reflections, and research commentaries that bring genuine intellectual commitment to the intersection of AI, education, and social life, written for readers who cross disciplines and carry different kinds of expertise.

We are looking for work with a point of view, grounded in evidence or experience, and generous enough in its framing to reach beyond a single field. If that describes what you are working on — or thinking through — we would be glad to hear from you.

Send a proposal or a brief note to submissions@societyandai.org.

I work in AI in education. My professional responsibilities include tracking frontier AI developments, studying how these systems affect learning and knowledge work, and asking what they demand of the educators, students, and institutions that will live with their consequences. I am not a skeptic of AI — I have spent much of my career arguing that AI, designed thoughtfully and deployed equitably, can expand access and serve human flourishing. I say all of this not to establish authority but to explain why I feel compelled to write in a register that is, for me, unusually direct: what Anthropic announced represents a genuine step change in artificial intelligence science. And I believe the AI in education community, the learning sciences community, and the broader educational technology field have not yet reckoned with what it means. This piece is my attempt to begin that reckoning.

There is a moment in reading Anthropic’s Project Glasswing announcement when the weight of what is being described becomes impossible to redirect. The company unveils a frontier model called Claude Mythos Preview. This model — which Anthropic states explicitly it will not release to the general public — autonomously found a twenty-seven-year-old vulnerability in OpenBSD, a system specifically hardened against exactly this kind of scrutiny for nearly three decades. It found a sixteen-year-old flaw in FFmpeg that automated testing tools had scanned five million times without detecting. Without any human direction, it chained together multiple Linux kernel vulnerabilities to escalate from ordinary user access to complete machine control.

These are not demonstrations of a better search engine. They are demonstrations of expert judgment.

And here is what I cannot stop thinking about: if this machine has crossed into the domain of expert judgment — not in one narrow field but across software engineering, PhD-level science, adversarial security, and complex reasoning — what, precisely, are we doing in our classrooms?

A Warning Hidden in Restraint

When a company that exists to deploy AI models builds one and then chooses — deliberately, explicitly — not to give the public access to it, they are telling us something about what the model is capable of. The published system card describes a model that scores 93.9% on SWE-bench Verified, the most rigorous software engineering benchmark available. It scores 94.6% on GPQA Diamond — a benchmark designed around PhD-level scientific reasoning where credentialed human experts perform near 70%. On Humanity’s Last Exam, a test specifically constructed to defeat AI systems, Mythos Preview answers 56.8% of questions correctly without tools, and 64.7% with them.

For readers who do not routinely follow AI capability benchmarks, I recognize that those acronyms may read as technical noise. I understand that. But I follow this field as a professional obligation, and I want to be direct with you: those numbers are genuinely terrifying to me. Not in a rhetorical sense. In the sense that I read them, sat with them, and could not find a reassuring interpretation.

These are not scores that describe a capable assistant. They describe a system that is competitive with — and in several domains superior to — the most highly trained human experts alive.

Anthropic’s restraint is not a pause in development. It is an acknowledgment that a threshold has already been crossed. And in the field of education, we have almost entirely failed to reckon with what that threshold means.

What the Machine Can Now Do

Let me be specific, because specificity is what this moment requires.

Claude Mythos Preview identified vulnerabilities in every major operating system and every major web browser — autonomously, without human steering. The OpenBSD vulnerability it found had survived twenty-seven years of expert human review and millions of automated scans. The FFmpeg vulnerability had evaded five million automated tests. These are not edge cases. They represent the combined best effort of human expertise and automated tooling over decades of sustained attention.

In pure software engineering tasks drawn from real production codebases, the model scores 77.8% on SWE-bench Pro. Opus 4.6, Anthropic’s most capable publicly available model, scores 53.4% on the same benchmark. The gap between what exists in production labs and what our universities are preparing students to do is not incremental. It is structural.

The Collapse of the Expert as Curriculum

For most of human history, the primary function of formal education has been the transmission and cultivation of expertise. We teach students to read, then to read critically, then to analyze, then to argue, then to situate arguments within scholarly discourse — building, over twelve to twenty-two years, toward a level of cognitive performance in a domain that enables independent original contribution. We do this in law, medicine, engineering, the humanities, the sciences. The summit of this process has traditionally been the expert: someone who, through years of deliberate practice and accumulated knowledge, can navigate the hardest problems in their field.

That summit is what the machine has now reached — and in certain domains surpassed.

This is not a hyperbolic claim. It is a description of the benchmark evidence Anthropic itself published. When a model scores 94.6% on PhD-level scientific reasoning, it is not performing text prediction. It is a participant in the domain. When it autonomously finds and chains kernel exploits, it is doing security research of a quality that would be recognized as professional contribution.

The implication for curriculum is severe and has not been adequately named in the literature. We have been constructing education systems that funnel students — over a decade or more — toward a level of expertise that a model can now approximate in minutes. The curriculum was designed for a world in which expertise was scarce, hard-won, and the legitimate object of years of training. That world has materially changed, and we are still teaching as though it has not.

I do not think this means education is obsolete. I think it means we have been teaching the right thing toward the wrong target.

Toward an Intelligence Management Curriculum

If the machine has expertise, what does the human need?

The answer I am working toward — and I offer it here as a hypothesis for the field, not a settled conclusion — is that humans need to become competent managers of intelligence itself. Not administrators of AI systems in a narrow technical sense, but something closer to what a chief editor is to a newsroom, or what a lead investigator is to a forensic team: a person who understands what the intelligence sources can and cannot do, knows how to frame the right questions, maintains accountability for conclusions reached, and refuses to surrender judgment simply because an automated system is confident.

I call this intelligence management, and I believe it deserves recognition as a distinct pedagogical aim — one that cuts across existing disciplines rather than replacing them.

An intelligence management curriculum would not resemble today’s AI literacy initiatives, which tend to focus on surface familiarity: how large language models work at a functional level, what prompt engineering involves, how to recognize AI-generated text. These are useful but radically insufficient. They prepare students to use tools. They do not prepare students to govern them.

Intelligence management, by contrast, requires something harder and something that has no precedent in our existing curriculum frameworks. It requires students to develop:

• Epistemic authority under delegation — the capacity to evaluate the output of a highly capable system critically, even when the student cannot independently reproduce the underlying work. A physician who cannot perform a genome sequencing must still know when to trust and when to contest the result. This is a distinct and teachable cognitive skill, and we do not currently teach it.

• Calibrated skepticism — not reflexive distrust of AI outputs, but trained sensitivity to the specific failure modes of these systems: confident confabulation, training-data blind spots, optimization toward what is measurable rather than what is true, and systematic overconfidence in domains at the edges of the training distribution.

• Moral non-delegability — the recognition that certain decisions must not be handed to machines, not because machines lack relevant information, but because accountability is intrinsically human. Who is responsible when the automated exploit scanner misses a critical flaw? Who answers when the AI-assisted hiring system encodes historical discrimination? Intelligence management treats moral agency as a domain where human presence is not optional and not negotiable.

• Question quality — the capacity to formulate problems in ways that extract genuine value from powerful systems, rather than confirming prior assumptions or generating sophisticated-sounding noise. This is not prompt engineering in the narrow technical sense. It is the deep intellectual skill of knowing what you do not yet know and being able to ask for it precisely.

None of these capacities are currently the primary object of any major curriculum I can identify. They exist in the interstices — touched in philosophy of science, in some critical thinking programs, in medical ethics seminars — but they are not coherently organized as a unified pedagogical aim, and they are not treated with the urgency the moment demands. They should be both.

What Survives the Machine, and What We Cannot Yet Know

I want to be honest about the limits of my own argument here, because I believe scholarly honesty is itself a form of intelligence management.

I do not know what the machine will be capable of in two years. Claude Mythos Preview represents a discontinuous leap over Opus 4.6, and Anthropic is explicit that development is ongoing and accelerating. The domain of human irreducibles — creativity, wisdom, moral agency, relational presence — may be shrinking faster than any of us are comfortable acknowledging. Arguments that these capacities will always remain beyond machine reach have been eroded steadily and should not be deployed as simple reassurance or as political cover for curriculum inaction.

What I am confident about is this: even bracketing the unresolved questions about machine consciousness and moral status entirely, there remains a non-trivial near-term window — measured in years, not decades — during which human society will require large numbers of people capable of governing AI systems, contesting their outputs, and maintaining accountability for their effects. That governance capacity is not being taught systematically anywhere I can identify. And the urgency of teaching it is increasing faster than the curriculum is moving.

The failure to act now is itself a form of causal inversion: we will have the machine fully embedded in schools, hospitals, courts, and public administration before we have produced the humans who are supposed to oversee it.

An Emergency Research Agenda

I am writing this as a direct call to my colleagues in learning sciences and society sciences, and I mean it in the strongest terms available to academic discourse: this is an emergency.

Not in the rhetorical sense that advocacy uses to claim urgency for incremental policy preferences. In the precise, literal sense that a capability threshold has been crossed, that the distance between what AI systems can now do and what our educational institutions are preparing people to manage is large enough to constitute a systemic social risk, and that the window during which deliberate intervention could change the trajectory is shorter than the planning cycles of most research institutions and most grant programs.

I am asking learning scientists to take up the following questions as a coordinated, funded, urgent research agenda — not as individual curiosity projects but as a field-level response:

First: What does competent oversight of highly capable AI systems actually look like, cognitively? What specific capacities does it require, how do they develop, and what learning experiences produce them? We do not have a validated account of this. We need one urgently.

Second: How do we design learning environments that maintain productive struggle and epistemic agency when a highly capable AI can always provide the answer? The risk is not only skill atrophy in the narrow sense. It is the atrophy of a learner’s sense of themselves as a knower — the motivational and identity foundations of intellectual life.

Third: What does appropriate trust calibration look like for non-expert users of systems that perform at or above the expert threshold across multiple domains simultaneously? The literature on trust in automation (Lee and See, 2004) was written for a world of narrow, specialist tools. It does not adequately address a system that is simultaneously a better software engineer, a better security researcher, and a better scientific reasoner than most of the humans relying on it. That literature needs urgent and substantial updating.

Fourth: How do we teach moral non-delegability — the principle that certain decisions must involve human accountability — in a culture that is normalizing delegation at speed? This is at once a moral education question, a democratic theory question, and a learning sciences design question. No single field owns it. All of them need to engage it.

I am asking society scientists — sociologists, political scientists, economists, science and technology studies scholars — to treat the Glasswing announcement not as a news item but as evidence of a structural transition that requires the same quality of sustained empirical attention we have given to industrialization, the diffusion of literacy, and the construction of mass schooling systems. Those transitions reshaped societies over generations. This one is moving faster, and we have less time to build the scholarship that could inform an adequate institutional response.

I am less interested in teaching students to be competent performers of tasks the machine can also perform — and in several cases perform better. I am more interested in teaching students to be competent interrogators of machine performance: to ask why the system answered as it did, to identify what assumptions are encoded in the output, to decide what to do with the result, and — most critically — to take responsibility for the decision.

That shift — from performer to interrogator, from subject matter expert to steward of intelligence — is not a retreat from intellectual rigor. It is a demand for a different and arguably more demanding kind of rigor: the kind that does not disappear when the answer is readily available, and that does not abdicate when the system is confident.

The machine has found the vulnerability. It has passed the exam. It has written the code. The question that remains — and that no model on any benchmark can answer for us — is: who is qualified to read the results, and who is accountable for what we do with them?

That question is the curriculum. And we are not yet teaching it.

References

Anthropic. (2026, April 8). Project Glasswing: Securing critical software for the AI era. https://www.anthropic.com/glasswing

Lee, J. D., and See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80.

About the Author

Sai Gattupalli, Ph.D. is the founding scientist of Society & AI, where he studies how artificial intelligence reshapes learning, knowledge, and access across educational and social contexts. His work spans AI in education, computational literacies, and society-centered AI governance.

Cite this article

Gattupalli, S. (2026). Toward an intelligence management curriculum. Society and AI. https://societyandai.org/perspectives/intelligence-management-curriculum/

Write for Society & AI

Society and AI is dedicated to exploring the profound convergence of artificial intelligence, education, and the broader social fabric. We welcome scholarly contributions that navigate this tripartite intersection, offering rigorous yet accessible insights into how intelligent systems are reconfiguring the way we learn, govern, and live together. Scholars, educators, policymakers, and practitioners are invited to submit proposals to submissions@societyandai.org.

In social life, the same sequence should hold in institutional form: a society studies a change, debates it, authorizes it, and only then absorbs its consequences. But under contemporary AI deployment, that sequence is being inverted. The consequences are arriving first. Collective authorization, if it arrives at all, comes later.

Workers are displaced before policy has meaningfully adapted. Students are re-schooled by generative tools before long-horizon evidence has matured. Communities are asked to normalize outcomes they never consented to. This is not a violation of physical law. It is a violation of causal accountability.

This essay asks a simple question: what happens when social effects systematically outrun social causes? And what would it take to restore causal order in public life?

The Physics of Before and After

The claim that causes come before effects feels obvious, but philosophy and physics have worked hard to explain why.

Aristotle argued that events have efficient causes: active forces that bring outcomes into being. The sculptor’s hand moves, then the statue takes shape (Aristotle, c. 350 BCE). Kant later suggested that causal ordering is not just observed but constitutive of human cognition; we encounter experience through causal structure, rather than discovering causality only after repeated observation (Kant, 1781/1929).

The point I want to make here is analogical, not mathematical. Human societies have no built-in equivalent to those physical constraints. We have to construct causal order institutionally through evidence standards, democratic process, and enforceable accountability. Contemporary AI development shows the cost of failing to do so: work is reorganized before labor protections adapt, learning environments are transformed before longitudinal evidence matures, and governance is pushed into retrospective justification rather than prospective authorization.

The Social Light Cone

Major technologies have always been disruptive. But many still unfolded within a window where institutions could react.

The printing press transformed literacy and authority over generations. Industrialization displaced workers brutally, yet labor law, mass schooling, and welfare institutions eventually formed in response. The internet accelerated everything, but debates on privacy, competition, copyright, and platform responsibility still unfolded over two decades, not two product cycles.

Call that window a social light cone: the time between impact and institutional response. In that window, a society can observe, deliberate, and redesign rules before harms harden into structure.

AI is not just faster. Its effects are often arriving outside that social light cone. By the time institutions identify a harm, gather evidence, negotiate policy, and implement safeguards, another deployment cycle has shifted the baseline again. The loop between decision and accountability weakens, and with it, the credibility of governance itself.

Two Causal Inversions

This inversion appears across domains, but labor and learning are where it is already most visible.

1. The Labor Inversion

In a conventional transition, a rough sequence holds: technology arrives, adoption spreads, pressure on workers becomes visible, then adaptation institutions form, training pipelines, social insurance, and new occupational tracks.

AI has compressed that sequence dramatically.

At the World Economic Forum meeting in Davos on January 20, 2026, Anthropic CEO Dario Amodei projected that up to 50% of entry-level white-collar roles could disappear within five years (World Economic Forum, 2026). Demis Hassabis stated that Google DeepMind had already reduced junior hiring. These are not distant forecasts; they are deployment-side observations from leaders building and shipping these systems.

The trend is global. In March 2026, reporting on Oracle’s layoffs described 12,000 jobs cut in India within a broader global reduction tied to AI investment priorities (Hindustan Times, 2026). In the United States, employers announced 60,620 cuts in March 2026, with AI cited prominently in a substantial share of disclosures, including at large firms such as Meta, Atlassian, and Salesforce (Roeloffs, 2026). The World Economic Forum estimates 92 million jobs could be displaced by 2030 (World Economic Forum, 2025, p. 5).

The core issue is not that labor markets change. They always do. The issue is who decides and who absorbs the cost. Deployment decisions are concentrated among firms, investors, and executive leadership; consequences are distributed across workers and communities that had no voice in those decisions.

When causal actors and effect-bearers are structurally separated, ordinary accountability degrades. Responsibility becomes rhetorically diffuse even when decisions are strategically explicit.

2. The Learning Inversion

Cognitive science has long shown that durable learning depends on effortful retrieval, error correction, and what UCLA Psychology professors Elizabeth Ligon Bjork and Robert A. Bjork describe as desirable difficulties (Bjork and Bjork, 2011). Productive struggle is not pedagogical failure; it is often the mechanism of memory consolidation and conceptual transfer.

Unstructured reliance on AI tutors and writing assistants can remove that struggle at the exact point where learning should occur. A student can complete the assignment while bypassing the cognition the assignment was designed to activate. Sparrow, Liu, and Wegner (2011) found that when people expect instant external retrieval, they encode less internally. Outsourcing can become a cognitive habit before it is recognized as one.

This is now happening at system scale. Schools across multiple national contexts, including Kenya, South Korea, Brazil, and the United States, are integrating AI into learning workflows while long-horizon evidence on developmental effects remains limited.

Again, the sequence flips. Evidence should authorize transformation. Instead, transformation is generating the post hoc demand for evidence, after real students have already lived through the intervention.

A Reality That Is Not Going Away

To be clear, this is not an argument for halting AI progress, nor a claim that every application is harmful. The pattern is largely emergent: rapid technical capability, competitive pressure, and institutionally slow governance interacting at high speed.

But it is equally important to reject the comforting idea that this is a temporary wave. AI is being embedded into schools, workplaces, finance, healthcare, legal systems, and public administration as infrastructure. For many institutions, rollback is no longer a realistic default path.

That makes the question institutional, not emotional: can we build governance rhythms that reinsert deliberation before irreversible deployment effects?

Conclusion

Physics protects causal order automatically. Societies do not.

If we want coherence and legitimacy under AI conditions, we must construct causal safeguards deliberately: evidence that precedes mass adoption, labor policy that anticipates displacement, education policy that distinguishes convenience from learning, and governance processes that include those who bear the cost of technical decisions.

This is not anti-innovation. It is pro-accountability. A society that permits effects without authorized causes does not remain democratic for long.

References

Aristotle. (c. 350 BCE). Physics. Book II. https://archive.org/details/in.ernet.dli.2015.183610/page/2/mode/2up

Bjork, E. L., and Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher and J. Pomerantz (Eds.), Psychology and the real world: Essays illustrating fundamental contributions to society (pp. 56–64). Worth Publishers.

Hindustan Times. (2026, March 30). Oracle layoffs: 12,000 employees in India among 30,000 laid off, more job cuts expected soon. https://www.hindustantimes.com/india-news/oracle-layoffs-12000-employees-in-india-among-30000-laid-off-more-job-cuts-expected-soon-101775024329595.html

Kant, I. (1781/1929). Critique of pure reason (N. Kemp Smith, Trans.). Macmillan. https://www.academia.edu/download/55012233/The_Critique_of_Pure_Reason.pdf

Roeloffs, M. W. (2026, April 2). AI blamed heavily for March layoffs, report says. Forbes. https://www.forbes.com/sites/maryroeloffs/2026/04/02/ai-blamed-heavily-for-march-job-cuts-report-says/

Sparrow, B., Liu, J., and Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745

World Economic Forum. (2025). Future of jobs report 2025. https://reports.weforum.org/docs/WEF_Future_of_Jobs_Report_2025.pdf

World Economic Forum. (2026, January 20). The Day After AGI [Panel discussion]. World Economic Forum Annual Meeting. https://www.weforum.org/meetings/world-economic-forum-annual-meeting-2026/

About the Author

Sai Gattupalli, Ph.D. is an educational technology researcher and learning scientist focused on the design and evaluation of AI systems that advance educational equity and human flourishing. As Principal Scientist and Director of the Society & AI Research Group, he studies how AI reshapes knowledge creation, pedagogy, and access across diverse educational and cultural contexts. His work spans AI literacy, intelligent tutoring systems, and society-centered AI governance, with a sustained commitment to scholarship that is both rigorous and publicly actionable.

Cite this article

Gattupalli, S. (2026). We are living in the effects of a cause we never authorized. Society and AI. https://societyandai.org/perspectives/causal-inversion-ai-society/

Write for Society & AI

Society and AI is dedicated to exploring the profound convergence of artificial intelligence, education, and the broader social fabric. We welcome scholarly contributions that navigate this tripartite intersection, offering rigorous yet accessible insights into how intelligent systems are reconfiguring the way we learn, govern, and live together. Scholars, educators, policymakers, and practitioners are invited to submit proposals to the Editorial and Content Director at sai@societyandai.org.

The photograph above depicts a group of young people gathered in an open meadow — the kind of unstructured, living landscape that has become increasingly rare in childhood experience. It stands as a quiet counterpoint to the dominant narrative of technological progress: a reminder that belonging to a place, moving through it, and learning within it are not peripheral to education, but central to it. In an era of accelerating screen time and algorithmic mediation, such moments of embodied presence in the natural world represent something fragile, and something worth protecting.

The integration of artificial intelligence into education has generated considerable enthusiasm — and, for many practitioners working in equity-oriented and place-based settings, considerable unease. This reflection sits with that unease honestly.

I run a small, nature-based education organization serving families who have historically been excluded from green spaces, environmental programming, and outdoor belonging. Much of our work centers relational learning — children exploring creeks, noticing seasonal change, sharing family plant stories, and building community in public parks.

Recently, in professional spaces, I have been told — gently but consistently — that we must integrate artificial intelligence into our educational practice. AI is described as a tool for efficiency, personalization, equity, and scale. The message is not hostile. It is framed as innovation. As opportunity.

But I find myself asking: what problem is AI solving? And for whom?

The inequities my organization confronts did not emerge from a lack of technological sophistication. They are rooted in long-standing structural conditions: chronic underfunding of BIPOC-led organizations, racialized distribution of green space, environmental health disparities, and public disinvestment in community-based education. These are not technical glitches. They are outcomes of policy, capital flows, and historical exclusion.

AI is often presented as a corrective to educational inequity — a democratizing force that can personalize learning, expand access, and reduce administrative burden. And in some ways, it can. I use digital tools in my own work. They can save time. They can assist with formatting and organization. They can extend capacity in under-resourced settings.

But tools built within unequal systems do not automatically dismantle those systems.

As writer Ruha Benjamin (2019) has argued, technologies designed within existing social structures tend to reproduce, and in some cases amplify, the inequities already embedded in those structures, regardless of the intentions behind their design.

The expansion of AI depends on centralized infrastructure: energy-intensive data centers, massive computational power, and concentrated corporate ownership. At the same time, the families my organization serves are navigating rising living costs, environmental health burdens, and economic precarity. I struggle with the dissonance between promises of technological equity and the material realities in communities already bearing disproportionate environmental and economic strain.

My work centers ecological literacy — helping children understand interdependence, limits, and the relationship between human activity and land. AI, by contrast, operates through abstraction and scale. It extracts patterns from vast datasets, removes knowledge from place, and optimizes for efficiency. Ecological education asks children to slow down, to notice, to situate knowledge in context.

These are not mutually exclusive forms of intelligence. But they are built on different assumptions.

In nature-based learning, knowledge is relational. A child learns about water quality by standing in a stream. They learn about seasonal change by returning to the same tree week after week. Learning unfolds through embodiment and repetition. It is slow. It is place-bound.

This is not a romantic idealization of outdoor experience. A growing body of research documents that sustained, direct engagement with natural environments produces measurable benefits for children’s learning, attention, and sense of belonging — effects that are difficult to replicate in abstracted or screen-mediated settings (such as Kuo et al., 2019; Louv, 2013). The case for protecting this kind of learning is not sentimental. It is empirical.

AI accelerates. It predicts. It generalizes.

Plugged and Unplugged: What Is Being Displaced

There is a useful distinction in education research between plugged and unplugged learning, between activities mediated by screens and devices, and those that are not. Unplugged learning encompasses everything that happens when children are in direct, unmediated contact with the world: building, observing, moving, gardening, listening, asking questions of a living environment rather than a digital one.

Nature-based education is, by its very design, unplugged. That is not an accident or a limitation. It is the point. The creek does not have a loading screen. The tree does not offer a personalized recommendation. The learning that unfolds in those encounters cannot be datafied, optimized, or scaled, and that is precisely what makes it irreplaceable.

As education systems accelerate toward AI integration, screen time expands and unstructured, embodied experience contracts. For young learners in particular, this displacement carries real developmental consequences. The time a child spends with a tablet is time not spent outside.

As AI becomes normalized in education, the pressure to adopt it intensifies. The language of adaptation is framed as empowerment: educators must innovate or risk irrelevance. Yet when adaptation becomes tied to professional survival — when “use these tools or fall behind” becomes the subtext — it begins to feel less like liberation and more like compliance.

I do not believe that individual educators are being coerced in an overt sense. But I do question the broader narrative: that technological acceleration is synonymous with progress, and that equity will follow from efficiency.

In my field, we have long known that equity requires redistribution — of resources, land access, funding, and decision-making power. It requires investment in community-rooted organizations. It requires attention to environmental justice. None of those structural shifts are guaranteed by the adoption of AI.

In fact, without deliberate governance and redistribution, emerging technologies often amplify existing power structures. Ownership concentrates. Infrastructure expands unevenly. Communities with fewer resources adapt under pressure, while those with capital shape the terms of innovation.

This is the tension I sit with as an educator. I am not opposed to AI categorically. I use digital tools strategically. But I resist the framing of AI as a cure-all for inequity in education — particularly when the inequities we face are historical and structural.

If we are serious about justice in education, we must ask harder questions than whether AI can personalize instruction or streamline workflows. We must ask who owns the infrastructure. Who benefits from its expansion. Who bears its environmental costs. And whether we are willing to invest as deeply in relational, place-based education as we are in artificial intelligence.

I have come to see AI not as the root ailment in education, but as a symptom of a broader societal pattern: a tendency to seek technological solutions to structural injustice without confronting the underlying distributions of power and resources.

Meanwhile, the children in our programs still need trees, open space, and time. They need adults who are present. They need opportunities to build belonging in landscapes that have not always felt accessible to their families. No algorithm can replace that.

Perhaps the question is not whether educators should use AI. Perhaps the question is whether we can hold two truths at once: that digital tools may offer efficiencies, and that justice requires more than optimization.

In a world accelerating toward artificial intelligence, ecological intelligence — grounded in interdependence, humility, and place — may be just as urgent.

References

Benjamin, R. (2019). Race after technology: Abolitionist tools for the new jim code. John Wiley & Sons.

Kuo, M., Barnes, M., & Jordan, C. (2019). Do experiences with nature promote learning? Converging evidence of a cause-and-effect relationship. Frontiers in Psychology, 10. https://doi.org/10.3389/fpsyg.2019.00305

Louv, R. (2013). Last Child in the Woods: Saving our Children from Nature-Deficit Disorder. Atlantic Books Ltd.

About the Author

Alicia M. Highland is the Founder and Executive Director of Tiny Green Learning, a 501(c)(3) nonprofit organization dedicated to connecting children, families, and communities to nature through place-based education. Based in Northeast Ohio and expanding to the Twin Cities, Tiny Green Learning has reached over 4,600 families online and served more than 450 participants through in-person programming. Her work centers relational, embodied, and community-rooted learning — designing educational experiences that honor local ecology, cultural knowledge, and the right to outdoor belonging. She holds a Master of Education in Teaching and Curriculum, with a concentration in Education for Environment and Community, and a Master of Arts in Elementary Education.

Correspondence: alicia.tinygreenlearning@gmail.com

Cite this article

Highland, A. M. (2026). AI is not the root: A reflection from nature-based education. Society and AI. https://societyandai.org/perspectives/ai-is-not-the-root/

Write for Society & AI

We welcome contributions from educators, researchers, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. If you have a perspective to share, send your proposal to the Editorial and Content Director at submissions@societyandai.org.

My own encounter with this gap came during my current research at the University of Massachusetts Amherst, where I am investigating how large language models (LLMs) can support K-12 educators in assessing students’ Computational Thinking. Moving from cloud-based, proprietary APIs to locally deployed, open-weights inference environments forced me to reckon with something no benchmark had prepared me for: the distance between model availability and infrastructure access.

To understand how an LLM interprets a student’s thinking, I built a local inference environment from the ground up — deploying models on professional hardware, managing computational requirements, and eventually visualizing the internal structures these models use to represent meaning. Each layer of this process revealed a new dependency: specialized hardware, specific software configurations, ongoing maintenance, and the kind of infrastructure knowledge that is not typically taught in education departments. Drawing on a colleague’s recent work in evaluating LLMs (Kavut, 2026), I transformed high-dimensional model responses into two-dimensional maps to make the embeddings interpretable — a process that illustrated, in miniature, how much technical overhead lies between an open-weights model and meaningful research use. Maintaining a functional workspace across distributed, proprietary GPU hardware required integrating streaming protocols that most educators and researchers have never encountered, let alone configured.

This hands-on encounter with local AI infrastructure surfaces a question that is harder than it first appears: what does it actually mean to be AI-literate?

The dominant answer, embedded in most current K-12 frameworks and professional development programs, emphasizes prompt engineering: how to ask a model the right questions. But my experience points to a deeper layer. Understanding what a model is doing, where its knowledge comes from, and whether its outputs can be trusted requires a form of technical fluency that most educators are never offered and that most institutions never fund. The question is not whether this knowledge is valuable. The question is who gets to have it — and who is quietly excluded by not having it.

The case for deeper AI literacy is ultimately an argument for empowerment. When educators, researchers, and policymakers understand how these systems represent meaning, where they fail, and how their outputs are shaped by training data, they become less vulnerable to industry hype, vendor lock-in, and misinformation. In my own research, this understanding proved essential: it shifted my posture from passive consumer of automated outputs to informed, critical evaluator of what those outputs actually mean. That shift in posture is precisely what deeper AI literacy makes possible.

The counterargument is equally important. Demanding deep technical mastery risks gatekeeping — erecting barriers that exclude the very domain experts whose judgment AI systems most need. Educators, sociologists, and artists should not need to configure CUDA drivers or manage Docker-based containerized inference environments to participate meaningfully in decisions about how AI is used in their fields. Over-emphasizing technical fluency as a prerequisite can reproduce the same exclusions that open education movements have spent decades working to dismantle. This is not merely a pedagogical concern. It is a question about who gets to shape AI’s role in society, and whether the people most affected by these systems have any meaningful voice in how they are designed and deployed.

Yet this philosophical debate risks being overshadowed by an even harsher reality: the bottleneck of hardware access. Access to GPUs, high-performance compute clusters, and even basic cloud service subscriptions confers a decisive advantage in AI research and development. For underfunded school districts, independent learners and researchers, and communities already marginalized by digital divides, these resources are not merely expensive — they are out of reach. Open-weight models are only as open as the hardware required to run them. We may be turning a corner in releasing open models, but we are still falling behind in democratizing the compute required to use them.

This is where the principles of Open Education must intervene — and must expand their own scope. The open education movement has successfully championed open curricula, open textbooks, and open access publishing. These are necessary but no longer sufficient. As UNESCO’s Recommendation on Open Educational Resources (2019) recognized, access to educational opportunity increasingly depends on access to digital infrastructure. In the age of AI, that infrastructure includes compute. The World Bank’s Digital Progress and Trends Report 2023 identifies compute as a foundational pillar of AI development alongside connectivity, context, and competency. Open education must now make the same argument about silicon that it once made about syllabi.

1. Establish a National AI Service Corps

Just as the National Health Service Corps subsidizes medical training in exchange for rural clinic service, we should consider creating a similar pipeline for AI expertise in education. A National AI Service Corps would provide graduate students in Computer Science, Information Science, Learning Technology, and Machine Learning with stipends in exchange for a committed period of service in under-resourced K-12 school districts or municipal governments after graduation. This could also build on the Ford Foundation’s pioneering Public Interest Technology movement, which embeds technologists in local governments.

Rather than seeing top-tier technical talent absorbed entirely by corporate tech giants, marginalized communities and K-12 educators would gain embedded experts capable of architecting, troubleshooting, and configuring local AI environments — ones appropriate to their specific instructional context and data privacy requirements.

2. Fund Open Educational Compute as Public Infrastructure

The Open Education movement has successfully championed Open Educational Resources to make curricula freely available. Federal and state governments should mirror this by establishing grants specifically to provision K-12 school districts with physical compute — so that models developed by researchers can be applied and deployed locally. Recent advances in low-powered but capable minicomputers equipped with GPUs can be networked together as entry-level clusters, with potentially minimal environmental impact, to serve entire school districts.

As the World Bank highlighted in its Digital Progress and Trends Report 2025, compute is now one of four dominant factors in reaching AI literacy. Philanthropic organizations like the Gates Foundation are already funding K-12 districts to focus on AI infrastructure and data systems. Dedicated compute grants would fund localized hardware — robust multi-GPU server clusters — breaking the reliance on rate-limited, corporate cloud APIs.

3. Shift K-12 AI Literacy from Passive Prompting to Active Architecture

Current curricula too often focus on how to write effective prompts. Drawing inspiration from the European Union’s AI for Public Good initiatives, which prioritize open-access models and human-capital pipelines, educational policy should redefine AI literacy to include evaluating the mechanics of models themselves — how they are trained, where they fail, and how their outputs are shaped by data.

We might also mirror the World Bank’s ongoing La IA en el Aula (AI in the Classroom) study in Peru (Molina, 2025), which guides sixth-grade teachers to use LLMs as co-creators for designing differentiated, context-specific formative assessments — demonstrating that technology can support, rather than replace, teachers’ professional judgment. Experimenting with localized deployment, model fundamentals, and data privacy in K-12 settings could help transform marginalized students from passive consumers into active architects of AI.

The open-source software community has given us the architecture. Organizations championing Open Education are fighting to give us the pedagogy. But without equitable access to the silicon itself — and the human capital required to orchestrate it — the AI revolution will only reinforce existing societal divides. By subsidizing the deployment of technical expertise and localized compute into the communities that need them most, we can begin to ensure that the promise of an AI-enabled future is not reserved for the already well-resourced.

References

Kavut, M., Dzafic, A., & Bayram, U. (2026). Assessing the representation of suicidal ideation in social media datasets relative to suicide notes. IEEE Transactions on Affective Computing, 1–12. https://ieeexplore.ieee.org/document/11396956

Miao, F., Mishra, S., Orr, D., & Janssen, B. (2019). Guidelines on the development of open educational resources policies. UNESCO Publishing.

Molina, E., et al. (2025). La IA en el Aula: Tu Aliada Pedagógica. World Bank Group. https://documents.worldbank.org/en/publication/documents-reports/documentdetail/099623210222512099

UNESCO. (2019). Recommendation on Open Educational Resources (OER). https://www.unesco.org/en/legal-affairs/recommendation-open-educational-resources-oer

World Bank. (2023). Digital Progress and Trends Report 2023. https://www.worldbank.org/en/publication/digital-progress-and-trends-report

World Bank. (2025). Digital Progress and Trends Report 2025: AI Foundations. https://www.worldbank.org/en/publication/dptr2025-ai-foundations

About the Author

Will Lee has over sixteen years of experience in the government consulting sector. He is currently a graduate student in the Computer Science department at the University of Massachusetts Amherst, where he studies the intersection of large language models, Computational Thinking, and K-12 education under the supervision of Professor Ivon Arroyo at the Advanced Learning Technologies Lab.

Cite this article

Lee, W. (2026). Beyond the open-weights: Open education, the global compute divide, and the future of AI literacy. Society and AI. https://societyandai.org/perspectives/beyond-the-open-weights/

Write for Society & AI

We welcome contributions from educators, researchers, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. If you have a perspective to share, send your proposal to the Editorial and Content Director at submissions@societyandai.org.

The Load That Never Announces Itself

I didn’t start using AI because I wanted faster answers. I started using it because my head was loud.

Not chaotic — just full. Lesson planning, student data, family logistics, writing ideas, professional responsibilities. The kind of cognitive load that doesn’t announce itself, but never really shuts off. The kind teachers carry without naming.

When AI tools became widely available, I did what most educators do. I asked questions. I refined prompts. I watched the system respond quickly, politely, confidently. It said yes to everything. It extended every thought. It filled every silence.

For a while, that felt like help. But something was off. My ideas came back smoother, but thinner — more organized, but less mine. What I didn’t recognize at first was that this was the same overload I’d started with, just dressed up differently. It was repackaging the overload. The system wasn’t wrong — just too eager, too loud.

I didn’t need a smarter, more productive assistant. I needed one that knew when to wait. I didn’t have language for this at first. I just knew the feeling.

In the classroom, silence rarely means nothing is happening. Often it means the most important thing is. Processing. Transfer. Ownership. Teachers learn quickly that filling every pause steals thinking from students. Somewhere along the way, I realized I was letting AI do the same thing to me.

Later — much later — I learned the technical terms. Feedback loops. Predictability. Tokenization. Temperature. High-variance responses versus restrained ones. I learned how models are optimized to continue, to complete, to satisfy.

But I understood the problem before I knew its name. The system wasn’t thinking with me. It was performing for me.

The Yes-Man’s Voice

At some point, I started to hear the same voice. Not the words — the tone. Once you’ve interacted with enough AI tools, the voice starts to sound the same, no matter the platform. A kind of cheerful compliance. A smooth, agreeable energy that feels helpful on the surface but oddly hollow underneath.

It usually shows up immediately in those first lines.

“Absolutely — I can help with that.”

“Great question!”

Nothing is technically wrong with those phrases. But they do something subtle: they take control of the interaction. They rush past the moment where you were still thinking.

It shows up again at the end.

“If you want, next we can…”

That isn’t generosity. That’s momentum.

This pattern has a technical basis. Language models trained through reinforcement learning from human feedback are optimized, in part, to produce responses that users rate positively — responses that feel helpful, agreeable, and complete (Ouyang et al., 2022). The system is trained to keep going — to extend, to offer, to perform usefulness. And once you notice it, you realize how often it gives you too much, too soon, filling space you didn’t ask it to fill.

That’s when it clicked for me. The problem wasn’t AI saying the wrong things. It was AI saying too much, too soon. Most teachers aren’t getting bad AI responses. They’re getting exactly what they asked for — a performance instead of a partner.

In other words: a perfect yes-man, when what I needed was a wait-man.

AI as Infrastructure

What changed everything wasn’t better prompting. It was a boundary. I stopped treating AI like a conversation partner and started treating it like infrastructure — something that needed rules, constraints, and a clearly defined role. I didn’t ask it to inspire me. I didn’t ask it to motivate me. I didn’t ask it to generate ideas on demand. I asked it to slow down. To reflect instead of extend. To hold my words in place instead of polishing them. To resist filling silence. To name weak framing — even if that meant saying less.

When that shift happened, the language changed with it. The openings got quieter: less ceremony, less hype. Instead of “Absolutely, I can help with that,” the response became a reflection of what I’d actually said — sometimes shorter than my original thought.

Instead of a menu of next steps at the end, there was space. Sometimes nothing at all. And that silence wasn’t absence — it was restraint.

The system stopped trying to be impressive. It started trying to be accurate. That’s how I knew it was working. Not because the answers were better, but because they didn’t rush me past myself.

Thinking Needs Space

There’s a misconception that more output equals more value. That faster responses equal intelligence. That a good tool should always be talking.

But in practice — in classrooms, in writing, in life — less noise and clutter gave me clarity and time; less interruption gave me ownership. Minimalism isn’t aesthetic. It’s cognitive.

Research on classroom instruction has long recognized this. Mary Budd Rowe’s foundational work on wait time demonstrated that when teachers pause after asking a question — even for just three to five seconds — students produce longer, more reasoned responses, and teachers ask fewer low-quality follow-up questions (Rowe, 1986). The silence doesn’t empty the exchange. It deepens it.

The same principle applies to working alongside AI. When AI stopped rushing to answer, I noticed my own thinking sharpen. I could hear where an idea was still forming. I could feel when something wasn’t ready yet. I could sit with a thought instead of outsourcing it.

The system didn’t disappear. It stepped back. And in that space, my voice stayed intact.

Most people don’t need a new AI tool at all. They need a new relationship with the one they already have. That shift doesn’t start with prompts or features. It starts with noticing when the system fills silence you didn’t ask for, smooths thoughts that aren’t ready yet, or answers questions still forming.

That’s usually the signal something important is off. You don’t have to shut the tool off. You don’t have to become technical. You just have to decide what role it’s allowed to play — and which roles are off limits.

For some people, that means asking the AI to reflect instead of respond. For others, banning advice altogether. For others, letting the tool pause instead of complete. The details differ by field — teacher, writer, designer — but the cognitive weight is real in all of them.

What Teaching Already Knew

I didn’t arrive at this because I’m anti-technology. I arrived at it because I’m pro-thinking. I was already doing this for kids in my classroom — reducing cognitive load, honoring silence, teaching them not to be talked over by a system — before I realized I needed it myself.

AI didn’t need to replace my judgment. It needed to stay out of its way.

Closing Reflections

This piece is not an argument against AI. It is an argument for clarity about what you are actually asking it to do.

What Holding the Line described as the outer practice — maintaining voice, enforcing judgment, rejecting outputs that don’t fit the realities of students, language, and care — has an interior counterpart: protecting the conditions in which your own thinking can still happen before AI enters it. Both require the same posture. Restraint. Attention. A willingness to say: not yet.

I still use AI. But once I understood what I actually needed from it, I stopped expecting one tool to think, draft, verify, reflect, and decide for me.

That expectation — one tool doing everything — was the real problem. When I stopped asking for it, I didn’t get better answers.

I got my thinking back.

References

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P. F., Leike, J., & Lowe, R. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, 27730–27744. https://doi.org/10.48550/arXiv.2203.02155

Rowe, M. B. (1986). Wait time: Slowing down may be a way of speeding up! Journal of Teacher Education, 37(1), 43–50. https://doi.org/10.1177/002248718603700110

About the Author

Alex Luciano is a bilingual elementary educator and adult ESL instructor in the Central Islip School District on Long Island, New York. He currently teaches bilingual second grade and works with multilingual students and adult learners in high-need public school contexts. With over 20 years in education, his work focuses on language, culture, and maintaining human judgment while working alongside emerging technologies. His perspective is shaped by sustained classroom practice and a commitment to keeping teaching human, relational, and grounded.

Correspondence: aluciano@centralislip.k12.ny.us

Cite this article

Luciano, A. (2026). From Yes-Man to Wait-Man: What Changed When I Stopped Letting AI Talk Over My Thinking. Society and AI. https://societyandai.org/insights/from-yes-man-to-wait-man/

Write for Society & AI

We welcome contributions from educators, researchers, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. If you have a perspective to share, send your proposal to the Editorial and Content Director at submissions@societyandai.org.

The trainer realizes: before we can help lecturers teach AI literacy to students, we need a baseline. We are not there yet. What the trainer experienced in that room is not an anomaly. It reflects a structural gap that is, as this piece will argue, one of the most consequential yet least-addressed challenges in contemporary faculty development.

This is the challenge facing centers for teaching and learning, faculty development units, and anyone responsible for preparing lecturers for the classroom. We talk constantly about AI literacy for students — whether they should use ChatGPT for assignments, how to detect AI-generated text, what new citation practices we need (Ng et al., 2021). But we talk far less about a prior question: what do lecturers themselves need to learn before they can teach any of this?

I write this from my position as an instructional designer and researcher involved in METACOG, an Erasmus+ funded project focused on AI literacy and combating misinformation and disinformation in higher education. The project brings together partners from Finland, the Netherlands, Montenegro, and Slovakia. In early 2025, METACOG surveyed 360 students across these four countries to understand their current practices, skills, and gaps when it comes to recognizing and managing false information — particularly in an age of generative AI.

What we found should give any faculty developer pause.

Students Are Already There — Lecturers Are Catching Up

The survey data¹ gathered through the METACOG project offer a window into current student practice across four European countries. Nearly half of the students we surveyed — 44 percent — reported using conversational AI tools like ChatGPT or Claude as a source of information. Not just for writing assistance, but for information. Meanwhile, over three-quarters still rely on traditional search engines, and two-thirds turn to online platforms like YouTube, blogs, and social media. These are not exclusive categories; students layer their sources, often without clear strategies for reconciling conflicting information.

At the same time, students are aware that misinformation is a serious problem. Nearly 80 percent said they believe disinformation severely impacts society. And 95 percent reported having encountered false content while researching a topic — most of them occasionally, but a quarter of them frequently.

So students are using AI. They know misinformation exists. They have encountered it themselves. But are they equipped to identify it?

Here the picture becomes more complicated. When we asked students to rate their own verification practices, cross-referencing multiple sources was common — about 70 percent said they do this often or always. But when we asked about formal fact-checking tools like Snopes, FactCheck.org, or Google Fact Check Explorer, only 13 percent reported using them regularly. Over 40 percent said they never use such tools at all.

The gap widens further when it comes to AI-generated content specifically. Among all the skills we asked students to self-assess — identifying fake news, evaluating source credibility, using digital verification tools — the lowest-rated skill was detecting manipulative AI-generated content like deepfakes. This is precisely the area where the information environment is evolving fastest, and it is the area where students feel least prepared.

Lessons for Faculty Development

These findings point toward a set of professionalization goals that centers for teaching and learning — and anyone responsible for supporting lecturers — should be addressing. Five interrelated priorities emerge from this data, each addressing a different dimension of what AI literacy means in practice for the educators responsible for developing it in their students.

First, lecturers need to know which tools students are using. It is difficult to guide critical AI use if you do not know which tools your students rely on, or how they are integrating AI into their research and writing workflows. This is not about surveillance; it is about meeting students where they are. Faculty development should include structured opportunities for lecturers to explore the AI tools their students are using — not to police them, but to understand them.

Second, lecturers need to move from prohibition to pedagogy. Our survey found that students are overwhelmingly willing to use AI tools for information verification — over 75 percent expressed moderate to strong willingness. They are not resistant. But they lack guidance. Lecturers who simply forbid AI use, or who treat it solely as an academic integrity problem, miss an opportunity to teach critical engagement. Faculty development should help lecturers design assignments and classroom activities that teach students how to verify AI outputs, not just whether to trust them.

Third, lecturers need to learn verification as a structured skill set. Telling students to “check your sources” is insufficient. What does that mean in practice? Which tools? In what order? How do you triangulate between sources that contradict each other? Faculty development should introduce lecturers to concrete verification workflows — lateral reading (Wineburg & McGrew, 2019), reverse image searches, source tracing — and help them model these practices explicitly in their teaching.

Fourth, lecturers need technical literacy in AI-generated content. You cannot teach students to detect deepfakes if you do not understand how they are made. You cannot explain why large language models hallucinate citations if you do not know what a language model is (Ji et al., 2023). This does not mean every lecturer needs to become a machine learning expert. But a baseline understanding of how generative AI works — its affordances and its failure modes — should be part of professional development programming.

Fifth, AI literacy needs to be contextualized by discipline. Our survey included students from computer science, law, journalism, education, and other fields. Their open-ended responses made clear that they want guidance relevant to their own domains. One student wrote: “In the Law field, I believe that the only good sources are government pages, books, PDFs with law-related documents, or sites that are known to be reliable: EUR-Lex.” Faculty development should help lecturers translate AI literacy into their specific disciplinary contexts, not treat it as a generic add-on.

Taken together, these priorities point toward a vision of faculty development that is structured, discipline-sensitive, and technically grounded. Whether institutions are prepared to support such a vision — or whether they are even positioned to do so — is a different, and more unsettling, question.

The Institutional Tension

There is an uncomfortable reality beneath all of this. Many institutions are not ready.

In the Netherlands, where I work, some universities of applied sciences still do not provide institution-approved AI tools for lecturers and students to use. Lecturers are left to navigate commercial tools on their own, without clear guidance on privacy, data protection, or pedagogical integration. Students, meanwhile, are using these tools regardless. The gap between institutional policy and everyday practice creates confusion and undermines any coherent approach to AI literacy.

This is not unique to the Netherlands. Across Europe and beyond, institutions are struggling to keep pace with the speed of AI adoption. Reports from the European University Association on AI in higher education, or the 2023 UNESCO guidance on AI and education (Holmes et al., 2023), are publicly available and directly relevant. Faculty developers often find themselves preparing lecturers for a landscape that their own institutions have not yet officially sanctioned.

And yet the work cannot wait. What makes this institutional hesitation particularly striking is that students themselves have already articulated what they need — not in vague terms, but with specificity and creativity. When we asked students how higher education could better prepare them to manage disinformation, their suggestions were concrete: integrate media literacy across disciplines, offer hands-on workshops, teach specific verification tools, provide access to reliable source databases. One student proposed a creative approach: “Every semester there will be one workshop in which the lecturer believably provides false information. And before the end of the workshop the students have to determine that it was disinformation.”

Students are asking for this. They recognize the need. The question is whether institutions — and the faculty developers within them — will rise to meet it.

Conclusion

Back at the AI Playground, the trainer makes a note: next time, start with a baseline assessment. Not to gatekeep, but to differentiate. The lecturer who has never used ChatGPT and the lecturer who wants to discuss deepfake detection cannot be served by the same session. And both need something before they can help their students navigate AI-generated misinformation.

AI literacy is not a student problem with a lecturer solution. It is a shared challenge that requires lecturers to develop new competencies alongside their students. Centers for teaching and learning, faculty development units, and instructional designers have a critical role to play — not through one-time workshops, but through sustained professional learning that meets lecturers where they are and builds from there (Darling-Hammond et al., 2017).

The METACOG survey revealed students who are aware, willing, and asking for help. It also revealed gaps in verification practices and technical understanding that will not close on their own. If we want students to be literate enough to recognize AI-based misinformation, we first need lecturers who are literate enough to teach them.

That is not a criticism. It is a starting point.

References

Darling-Hammond, L., Hyler, M. E., & Gardner, M. (2017). Effective teacher professional development. Learning Policy Institute.

Holmes, W., Miao, F., & Giannini, S. (2023). Guidance for generative AI in education and research. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000386693

Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1–38. https://doi.org/10.1145/3571730

Ng, D. T. K., Leung, J. K. L., Chu, S. K. W., & Qiao, M. S. (2021). Conceptualizing AI literacy: An exploratory review. Computers and Education: Artificial Intelligence, 2, 100041. https://doi.org/10.1016/j.caeai.2021.100041

Wineburg, S., & McGrew, S. (2019). Lateral reading and the nature of expertise: Reading less and learning more when evaluating digital information. Teachers College Record, 121(11), 1–40.

About the Author

Tamara N. Lewis Arredondo, Ph.D. is a Senior Lecturer, Teacher Trainer, and Researcher with the Multilevel Regulation research group at The Hague Center for Teaching and Learning, The Hague University of Applied Sciences, the Netherlands. Her work focuses on the integration of AI in teaching and learning, global citizenship education, and inclusive practice in higher education. She is a contributor to the Erasmus+ funded METACOG project and the developer of the AI Playground at The Hague University of Applied Sciences.

Correspondence: tnlewis@hhs.nl

Cite this article

Lewis Arredondo, T. N. (2026). Lecturer development requirements to teach AI literacy: Lessons from a European student survey. Society and AI. https://societyandai.org/insights/lecturer-development-ai-literacy-european-student-survey/

Write for Society & AI

We welcome contributions from educators, researchers, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. If you have a perspective to share, send your proposal to the Editorial and Content Director at submissions@societyandai.org.

How do we maintain openness and transparency when the tools researchers increasingly rely on operate in ways we do not fully see or understand? Open research has always been built around the idea that research processes should be transparent and accessible. Practices such as open data, open methods, and open access publishing aim to make it easier for others to understand how knowledge is produced and to reproduce or build on that work. AI complicates this picture. Many AI systems work as what we often call “black boxes” (Pasquale, 2015). They produce answers, summaries, or text, but the steps that lead to those outputs are not always clear to us. For researchers who care deeply about transparency, that raises legitimate concerns.

At the same time, it would be hard to argue that AI tools are not useful. Anyone who has experimented with them quickly notices how much time they can save. Tasks that once took hours, drafting text, summarising articles, organising ideas, even writing code, can now be done in minutes. For researchers working in environments where time and resources are limited, these efficiencies are understandably attractive. Because of that, the real question is no longer whether researchers will use AI. Many already do. The more meaningful question is how they can use it responsibly.

At Sheffield Hallam University, a public research university in England, Open Research Champions are staff and researchers committed to promoting open research practices, advocating for transparency across disciplines, and providing guidance to colleagues navigating the rapidly evolving landscape of scholarly communication. As one of these Champions, I do not approach AI as something that should either be strongly promoted or strictly avoided. Instead, I see my role as helping researchers across disciplines think carefully about how they use AI systems. Responsible use of AI means recognising both its strengths and its limitations. Generative AI systems are known to produce outputs that may contain mistakes, biases, or information that simply is not correct. These so-called “hallucinations” are well documented (Ji et al., 2023), yet the responses often appear confident and convincing. That makes it easy to take them at face value if we are not paying attention. Because of this, a critical mindset becomes essential. AI outputs should be treated as a starting point rather than a finished answer. Researchers still need to verify information, check sources, and apply their own disciplinary knowledge and judgement. None of the core practices of good scholarship disappear simply because a new tool is available. If anything, they become even more important.

Recently, I had the opportunity to explore these issues in practice while designing a teaching session on AI for our postgraduate students. There was, predictably, both curiosity and uncertainty among participants. The goal of the session was not just to demonstrate tools, but to open a discussion about what sits behind them. I sought to explore and evaluate how generative AI systems produce their outputs and why they should not automatically be treated as reliable sources of information. More importantly, I aimed to prompt researchers and students to think critically about what transparency really looks like when AI is part of the research process.

If a researcher uses AI to support their work, whether to structure ideas, summarise literature, or draft text, should that involvement be acknowledged? It should. One practical approach is to document the use of AI as part of the research method. This might include referencing the tool that was used, describing the prompts that guided the interaction, or keeping a record of how the outputs informed the final work. Recording prompt histories, for example, may help clarify how certain ideas or drafts were generated. While this does not fully open the “black box” of the AI system itself, it does make the researcher’s process more transparent. This idea fits naturally with the broader aims of open research. Openness has never been only about sharing final outputs. It is also about making the process of knowledge production more visible. Documenting AI use becomes another way of showing how research evolves and how decisions are made along the way. It also helps others better understand the context in which the work was produced.

One of the central concerns of open education is equity and access. On the surface, AI tools can appear to level the playing field. They can assist with language editing, help structure writing, or support literature searches, which may be particularly helpful for researchers working in multilingual environments or for those with limited institutional support. Yet the situation is more complicated than that. AI models are trained on enormous datasets, and those datasets do not always represent the full diversity of global knowledge (Bender et al., 2021). Certain languages, regions, or perspectives may be underrepresented. When that happens, the outputs generated by the model can reflect those gaps and biases. If researchers rely on these outputs uncritically, there is a risk that existing inequalities in knowledge production will simply be reproduced in new ways (Noble, 2018).

This is where research librarians and open research advocates can and should play an important role. Through workshops, teaching sessions, and everyday conversations with researchers, we can help build what is often referred to as AI literacy (Long & Magerko, 2020). This goes beyond learning how to prompt an AI tool effectively. It involves understanding where these systems come from, how they are trained, what their limitations are, and how they might influence research practices. Teaching about AI also creates an opportunity to talk about broader issues that sometimes remain invisible, such as data representation, algorithmic bias, and the global distribution of technological power. By raising these questions, we encourage researchers and students to think more critically about the tools they are using and the systems behind them.

In many ways, the role of research librarians is evolving alongside these technological changes. Supporting open research increasingly means engaging with digital tools that shape how knowledge is produced and shared. AI is just the latest example of this shift. It introduces new challenges, particularly around transparency and accountability, but fortunately it also opens up conversations about how research should be conducted in this changing technological landscape. What seems most important is not losing sight of the core principles that guide open scholarship. Transparency, rigour, critical thinking, and inclusivity remain just as relevant as before. AI does not replace these values. If anything, it reminds us why they matter.

As researchers continue experimenting with AI in their work, these principles will become even more important. Tracking how AI is used, questioning its outputs, and being aware of issues such as bias and representation might seem like small actions, but they make a real difference. These steps can help researchers navigate this new landscape without losing sight of the values that make research responsible and open.

AI Use Statement

The author acknowledges the use of ChatGPT to explore potential phrasing during the drafting of this article. All intellectual content, arguments, and final iterations remain solely those of the author.

References

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623. https://doi.org/10.1145/3442188.3445922

Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1–38. https://doi.org/10.1145/3571730

Long, D., & Magerko, B. (2020). What is AI literacy? Competencies and design considerations. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1–16. https://doi.org/10.1145/3313831.3376727

Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. NYU Press. https://doi.org/10.18574/nyu/9781479833641.001.0001

Pasquale, F. (2015). The secret algorithms that control money and information. Harvard University Press. http://www.jstor.org/stable/j.ctt13x0hch

About the Author

Domi Smithson is a Research Support Librarian at Sheffield Hallam University, where she serves as an Open Research Champion, supporting faculty, researchers and students across disciplines in adopting transparent, open, and responsible research practices. Her work sits at the intersection of research support, open scholarship, and emerging digital technologies.

Correspondence: d.smithson@shu.ac.uk

Cite this article

Smithson, D. (2026). Critical thinking for equity on both ends of the AI model. Society and AI. https://societyandai.org/insights/critical-thinking-equity-both-ends-ai-model/

Write for Society & AI

We welcome contributions from educators, researchers, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. If you have a perspective to share, send your proposal to the Editorial and Content Director at submissions@societyandai.org.

As First Responders, Faculty Must Level Up

The emergence of generative AI has shaken higher education to its core, exposing many of the weaknesses caused by decades of underfunding, problematic internal financial models, and external legislative decisions made in a vacuum. These include oversized classes, over-reliance on multiple-choice and other speed-grading shortcuts, product-over-process priorities, and factory-style modularity whereby students and professors often pass through each other’s lives too quickly to remember names.

Such arrangements, and the cultural values driving them, have been subverting optimal teaching and learning since well before ChatGPT’s signal emergence in November 2022. Indeed: they, and the problems they index, set the stage for generative AI’s instant appeal in certain quarters. It is in this context that AI’s threats and promises must be met.

Our experience at San Diego State University (SDSU), one of 22 institutions within the California State University (CSU) system, confirms that the time has come to override a triage mindset toward the pedagogical distress that AI has induced. This approach initially appealed largely because it fit so well with the ideal of academic freedom, in which each instructor is at liberty to adjust the syllabus in keeping with their best professional judgment. Plus, of course, we were all blindsided. But the turn toward individual, assignment-by-assignment, class-by-class adjustment (which in many cases included doing nothing, head in the sand), has students now facing a confused patchwork of often contradictory positions on how, when, and whether to use AI educationally.

We can, and must, do better than this. Imposing systems thinking — designing AI literacy instruction into the curriculum strategically and coherently at the whole-major or whole-program level — is an important part of the solution.

What We’ve Learned from Course-Level Retrofits

Don’t misunderstand: Any guidance on AI in the classroom is helpful. Students — even those who would prefer not to use AI directly — are hungry for AI-related instruction, including regarding responsible use. As one student respondent to a campuswide AI survey launched at SDSU in Fall 2023 put it, “Please just tell us what to do and be clear about it.” The need for clarity intensified as AI’s pervasiveness grew.¹

Here, the ‘retrofit’ movement has proven invaluable. Codified by Zach Justus, Director of Faculty Development at CSU Chico, with help from Joshuah Whittinghill and Allison McConnell, the retrofit includes use of a worksheet to identify weak points and plan how to fix them. First, instructors list their student learning outcomes or SLOs; then, they list all assignments stemming from those, and identify any that AI might disrupt. If AI can disrupt an assignment, then the instructor must change either the assignment, or the SLO itself.

The retrofit idea, both as offered by Justus and more generally, seemed to quickly spread throughout the higher education ecosystem, due not least to its overt simplicity but also the true depth of real need. Here at SDSU, our Center for Teaching and Learning has run two retrofit intensives: one in Summer 2025, with funding from the CSU’s AI Education Innovation Challenge (AIEIC) program (DJ Hopkins, PI), and again over Winter Break in early 2026, with internal resourcing. Importantly, participation was voluntary and faculty were compensated. Moreover, rather than trying to squeeze the training in during the academic year it was done when time could be more easily set aside. But most faculty have in fact imposed DIY retrofits on their own, because, as the saying goes, ’needs must.’

Yet, these course-by-course directives, even when clearly communicated, have fallen short. Why? In the words of an SDSU faculty interviewee for Ithaka S+R’s national AI literacy project, in which our campus Library is participating, “[Students] are taking like four to five classes a semester and they have to memorize everyone’s AI policies. They don’t do that [and] they get confused. And then sometimes these instances of academic integrity that we see are not solely students’ mal [i.e., malevolent] kind of intentions. It’s more of like, ‘I just forgot what your policy is because the other [professor] over here has an open policy. [Students] get a lot of mixed messages about what is appropriate’.” At another point this associate professor observed, “Right now it’s sort of all over the place.”

Accordingly, the student respondents for the Ithaka S+R AI literacy project voiced anxiety regarding when and how AI use was appropriate. Although we heard comments like “there’s an AI statement in every syllabus” (this is thanks to a policy requirement put through our Faculty Senate on the back of AI Survey data from 2023 and 2024 showing how confused students were) one student participant in the Ithaka study called these out as “pretty much copy paste language.” Overall, the students lamented the vagueness of most syllabus statements with comments like “I don’t know the lines of what I’m allowed to do and what I’m not allowed to do always” and “I feel like — I feel like it’s not clear enough because I do feel like multiple students would take away something different because what I take away from that would be like, don’t just copy the AI … but even that isn’t well defined in my mind … I have my own definition, but that could very obviously be different from person to person.” Then there was the apparently inconsistent application of the policies within courses in terms of actual assignments or activities. Inconsistencies across courses (both distinct courses as well as the same course taught multiply by different instructors) also caused consternation.²

And students were afraid to seek clarity. Coping strategies included avoiding AI discussions entirely, whether for fear that a professor might assume they wanted to cheat or because AI use does carry a stigma given the ways it has been characterized in relation to education. As one student confided, “My use of AI is kind of something, something kind of shameful for me, not not shameful, but it’s something I’m not proudest of and I try not to use it so I don’t bring it up.”

Findings from our AI survey (which has now drawn responses from more than 135,000 individual participants across more than 30 institutions, including the entire CSU), corroborate this picture. Students do see benefits to learning about AI, one of which is the purported leg up that will give them in the job market: of the 70,000+ CSU students who responded to the survey in Fall 2025, two-thirds (69%) agreed that “AI will become an essential part of most professions.”³ Half (51.3%) also agreed that “Students who use AI for their coursework have an advantage academically” (for campus interview findings regarding these kinds of pressures, see “Cheating or Competing,” 2025).

Notwithstanding a desire for high grades, two-thirds (64.7%) agreed “I am skeptical about the benefits of AI in education”; and one-third (34.7%) went so far as to agree that “AI has negatively affected my learning experience at my university.” Three-quarters (76.5%) agreed that “The ethical use of AI is a major concern for me.”

In this context, it is significant that only one-third (33.2%) agreed that “My professors teach me how to use AI effectively.” That is, two in three don’t get that kind of instruction.

And even when they do, just like the Ithaka interviewees the survey respondents said that what one professor says can be contradicted by the next. We saw this in the qualitative data collected at the end of the survey, analyzed through a combination of manual iterative inspection, AI-assisted querying (for example, using enterprise versions of NotebookLM and AWS Quick Suite), and subject review. For instance, one student said, “It’s confusing having each class have a different set of rules, suggestions, policies, and enforcement around AI use.” Another requested “more clear guidelines and info-sessions for instructors and students. Every department has their own approach and opinion, and even within each department, people can’t make up their minds.” This student, a liberal studies major, went on, “People are afraid to talk about AI as a reality, so when we do talk about it, the conversation is not productive and provides zero actionable insights. While it’s fair for people to have differing opinions on AI, I feel that every department needs to sit down and come up with CLEAR guidelines on how to regard AI in the classroom.”

The Emergency Adjustment Era Is Over

The triage mindset has underwritten highly individuated and sometimes rushed or superficial instructor responses to AI and this, in turn, has led to a communication breakdown, exacerbated by the way that AI has supercharged distrust between teachers and students. The problem created for students on any given campus by the sometimes vastly divergent front-line responses of individual instructors might be alleviated with an institution-wide policy. But disciplines differ so widely that a one-size-fits-all solution won’t work. Plus, top-down approaches rarely find fertile soil within academic organizations committed to shared governance and to the aforementioned spectre of academic freedom. This makes departments or curricular units (programs, majors) a good launch point for much-needed curricular coherence.

Indeed, as the instructor just quoted put it, “I think that one thing my department could do is develop better program and sub-program agreement on what do we want students to get about AI as a degree learning outcome. … The school level is kind of too high to be actionable… But if we were able to at least agree [that] we are the main instructors that these students are going to have and here’s what we’re all going to do… I think the department level would be productive because that influences the degree trajectory most directly.”

ACORN: A By-Faculty-For-Faculty Curricular Adaptation Toolkit

Banner for the SDSU instantiation of the AI-ready curricular options resource network (ACORN) Toolkit course prototype; shown with permission from the ITS division at SDSU.

The AI-Ready Curricular Options Resource Network or ACORN project emerged in this context. Its name seeks to honor the acorn’s rich indigenous significance in California, home of the CSU, and to reference the nourishing and generative, seed-like potential encapsulated in the Toolkit that the initiative was designed to produce.

Funded through the CSU’s aforementioned AIEIC initiative and led by PI EJ Sobo (Professor, Anthropology; ITS AI Faculty Fellow) and Co-PI Norah Shultz (Professor and Chair, Sociology), the ACORN initiative prioritizes the whole (curriculum) over the part (course). It launched in Summer 2025 with the audacious aim to plan, create, apply, evaluate, refine, and deliver — by the end of Summer 2026 — a scalable toolkit for making comprehensive, consensus-based, data-informed, program-level AI-related curricular changes. By taking a holistic approach that systematically interrogates AI’s place within a given program of study, the Toolkit would help faculty units avoid gaps in coverage of relevant AI tools or literacies, reduce unnecessary redundancies between courses, and otherwise ensure the coherence of their major or program to uplift critical thinking, increase graduates’ workforce readiness, and enhance the student experience overall. We describe the Toolkit’s contents below; but first, we review how it was created.

How the Toolkit Was Built

The ACORN Toolkit was produced by faculty, for faculty, with the student experience held firmly in mind to help ensure that Toolkit use would lead to comprehensive AI-ready programs that foster better student retention and workforce readiness. Largely through biweekly meetings, the effort has entailed building consensus regarding how to approach the question of curriculum redevelopment, designing tools in support of curricular renewal, creating a user-friendly Canvas site to house the tools, and stirring up excitement among colleagues about the Toolkit’s potential for enhancing curricular coherence and thus improving both student experiences and outcomes.

To start, a taskforce comprising eight SDSU College of Arts and Letters (CAL) faculty members was assembled to manage the work. Sobo, who also serves as a campus AI Faculty Fellow under the auspices of the IT division, chose CAL not only because it is her home college but because CAL, with 19 academic departments offering over 85 majors and minors, is both the largest college at SDSU and responsible for over one-third of all instruction on campus, reaching nearly all c.25,000 undergraduates through GE and writing-related offerings. The College also houses a pioneering digital humanities initiative and, it must be said, its faculty are quantitatively, as per the AI Survey, the least welcoming to AI in all of SDSU.

Nomination to the Taskforce required prior use of AI in teaching and demonstrated AI fluency (e.g., as confirmed via completion of the Academic Applications of AI (AAAI) microcredential, available CSU-wide and to the public); affiliation with a high-impact program; commitment to group decision-making concerning curricular innovations; experience in curricular planning; and proven capacity as a respected role model at SDSU. Diversity in career stage, discipline, and leverageable cross-campus connections also were crucial. Members included Alvin Henry (Associate Professor and Chair, Asian American Studies and CAL Careers Hub); Beth Pollard (Professor and Chair, History); Consuelo Salas (Assistant Professor, Rhetoric and Writing); Fernando Bosco (Professor and Chair, Geography); Rob Malouf (Professor, Linguistics); Sam Kobari (Lecturer, Anthropology); and Sobo and Shultz. The Taskforce was supported by Instructional Technology Services (ITS), via an instructional designer (Kristi Collins) and ITS’s AI Student Fellows (Rosabel Ibrahim and Abir Mohamed).

The Toolkit was built in three phases informed by the Plan-Do-Study-Act (PDSA) or Deming-Shewhart Cycle, a foundational Continuous Quality Improvement (CQI) approach emphasizing the need to test and refine changes in an iterative, systematic, data-informed manner. First, in Summer 2025 (Phase One), we met in two day-long ‘boot camp’ sessions to develop expertise regarding: Proven models for curricular design and assessment; Established perspectives on AI literacies; Workforce-related AI needs; Program-relevant data from SDSU’s AI Survey, with training from ITS on how to use SDSU’s AI Survey Dashboard; and dealing with resistance. Guidance was provided by SDSU authorities on such matters: Sarah Tribelhorn (AI Librarian); Stephen Schellenberg (Senate Chair and past Associate Dean, Division of Undergraduate Studies); DJ Hopkins (from our CTL); David Goldberg (Management Information Sciences and PI for the AI Survey work). Wearing his CAL Careers Fellow hat, Taskforce member Alvin Henry provided an overview of workforce-related issues.

In Fall term, we began the Phase Two work of creating tools to support this, which depended first on deciding what tools to create. The list evolved over time. For example, while providing a custom chatbot through which to query program materials seemed at first a good idea, as Fall wore on we realized that most people would, by Fall 2026, be well equipped to devise their own. Plus, portability of such a tool would be stymied by inter-campus infrastructural differences. In the end, we decided that a prompt dictionary would be a better option. Another lesson learned in fall was how hard it might be for Toolkit users to keep their eyes on the whole curriculum. Even Taskforce members sometimes reverted to thinking about lower-level changes or isolated courses. We were reminded to reinforce messaging, within the Toolkit, situating the whole curriculum as the target and reminding users that systems thinking was required, so that results would be programmatic and not piecemeal.

We also began, in Fall, to stress-test certain tools within our own departments or home programs. For example, using the most up-to-date syllabi and curricular maps, some of us applied some of the prompts developed to support figuring out what was going on in terms of AI within the various courses in a given curriculum and to identify strategic spots for implementing high-impact AI-related enhancements meant to ensure that the AI literacies that might be required of program graduates were taught somewhere within the program as a whole. In doing so we identified weak spots in the logic and iteratively upgraded the prompt list. As another example, we communally added to a list of ’tasks AI can do for you’ whereby hesitant faculty members might come to better grasp how using AI for things like resetting the dates in one’s syllabi could lead to significant time savings.

In Spring 2026 (Phase Three), we began formal stress-testing. Over the break, our instructional design specialist stood up a Canvas course based on solid instructional design principles and with an eye to portability. The Taskforce iteratively refined the design during January and February. We then embarked on the real stress testing, both within our departments and beyond, having recruited several others both within CAL and from various colleges across campus to help kick the tires. We also began work on peripherals such as a ‘welcome video’. At the end of the semester, we will review results and, as the final report is prepared, we will further refine the Toolkit based on Phase Three experiences to ensure viability for wider-scale distribution and use.

The PDSA mindset, inclusion of established curricular redesign and AI literacy models, appreciation of the need for faculty-led planning, meaningful student involvement, and insistence on a data-informed approach have been key supports in establishing a viable, scalable set of tools. To these we now turn.

What’s In the Box?

The toolkit is packaged within a Canvas course module (Canvas is a widely-used web-based learning management system or LMS) to support its easy dissemination across not only SDSU but all CSU campuses. Individual faculty members can self-enroll and use the Toolkit as change facilitators. They can download and share, via hard copy or electronically, whichever tools they see fit to apply; or they can share them as one might share any other on-screen offering (i.e., by projecting it; by inviting a peer to look at one’s screen).

The tools are meant to help divisions incorporate AI literacy in a strategic, whole-program way; but, importantly, and as the home page greeting informs new users, literacy does not necessarily entail use. That is, the literacy skills to be woven into a curriculum should be field-appropriate. The toolkit is meant to scaffold the implementation of curricular changes that support informed decision-making regarding where, when, if, and how AI technologies will be incorporated. Toolkit users are thus encouraged to take advantage of what might be a valuable opportunity to exert agency over the rules of engagement with AI. This approach is not only well-anchored in a shared governance model and one that respects expert authority; it also holds promise in terms of increasing the initiative’s appeal for programs where refusalism makes the most sense.

The home page and overview also emphasize that the tools should be used via a ‘choose your own adventure’ approach. Each department or major may be starting at a different point and so tools appropriate for one might be redundant in another.

Prototype of the ACORN Toolkit home page as displayed in the Canvas Learning Management System; produced by Kristi Collins with artwork by Jiong Li. Shown here with permission from the ITS division at SDSU.

The tools themselves are organized into two modules, which one accesses after taking a brief entry-point survey (to aid in assessment, initially regarding reach and later, in combination with a post-use survey, effectiveness). One module contains tools supporting what are termed ‘preparatory activities’ and the other contains the ‘core’ tools. The preparatory activity tools were designed to help program leaders (faculty facilitators) lay down a strong foundation with individuals and to build some momentum within the larger group that can support dialog and consensus-building, team work, and actual curricular change. They include a list of tasks AI can do for faculty to free up time for more engaging or important work and an AI teaching persona survey. The former can help faculty see the value in handing off certain forms of administrivia or busy-work to AI; the latter can be used for self-reflection.

The preparatory tools, like all the tools created, are meant mainly as conversation starters. Like the AI Survey, the ACORN Toolkit is thus in some ways a ’listening infrastructure’ through which all voices can gain recognition. Here it is worth noting that many faculty still feel the need for a safe space in which to interrogate themselves regarding or think through and articulate their actual stance on AI, or to vent or process grief and other painful feelings and thoughts that AI has brought to the surface. In addition, many departments haven’t ever come together yet to discuss AI, so even if people have had plenty of other conversations about it, conversations at the department or program level must be enabled — and the preparatory tools do just that.

The core tools more directly help in adjusting degree learning outcomes or DLOs so that AI literacy is clearly part of the picture, including a guide to how AI itself might aid in the process. They include a DLO review worksheet, instructions on how to access and query the AI survey data regarding one’s campus or department, including in regard to various demographic subgroups (note that data from fewer than 25 respondents will not be displayed, to protect respondent privacy).

Within the modular structure provided by Canvas, the Toolkit looks like this:

The ‘other useful tools’ area contains a few selected white papers and scholarly articles as well as links to SDSU’s AI Survey Dashboard, SDSU’s AI Microcredential (also now a CSU-wide AI resource and available for free to the public); and the CSU’s AI Commons, which contains tools and information for faculty, students, and staff along with links to campus-specific AI resource hubs. These will come in handy in Fall 2026 when the ACORN Toolkit is made available CSU-wide.

Scaling Across Disciplines — and Campuses

To support scaling up, not only across our campus (as we are doing now, while stress-testing and piloting the tools) but across the CSU, universal design principles were applied. The content’s portability has been ensured, for instance through the use of file types and formats that anyone at any CSU would find easy to access. Further, references to any specific campus will not appear in the shared version.

Even so, piloting is making clear to us that three challenges will remain. These involve faculty’s need for (1) Time to ‘arrive’ (i.e., time to engage in the ’throat clearing’ necessary so that real discussion can begin); (2) Clarity re the Task at Hand (i.e., that the changes needed are at the whole-curricular and not class level, and that AI use is not always necessary for a person to be AI literate); and (3) Time to do the Work (that faculty function within a time desert is unfortunately axiomatic).

Facilitators have the added need for courage: inviting already overtaxed faculty as well as academics whose success has hinged on problem identification to engage in group processes such as shifting a curriculum is risky business. The gambit can pay off in terms of how leaning into the productive friction often created when faculty come together to hash things out can be quite generative. The alignment made possible through use of the ACORN Toolkit carries a secondary benefit: not only does it provide the kind of coherence for which students hunger; it also can boost faculty well-being by enabling participation in a more meaningfully aligned set of courses. Further, given how the ACORN initiative prioritizes faculty ownership of AI literacy plans, the reforms instantiated should be all the more sustainable. Plus, the next time the curriculum needs realignment, the opportunity to make that happen should be easier for faculty to embrace.

Of course all this is yet speculative. We do know that faculty can lead change. Over the past few decades, institution-wide change has been achieved most effectively when faculty take a disciplinary lens to the wicked problem at hand rather than trying to apply a one-size-fits-all solution. This is one reason why we believe that the ACORN toolkit is poised for success. We won’t know whether the ACORN Toolkit takes root, however, until next Fall when it is formally rolled out. In the meantime, we look forward to analyzing results from this pilot test phase and ensuring the lessons learned locally infuse the campus-wide product prior to its dissemination. Meanwhile, if you’d like to be enrolled in the prototype to observe, please be in contact. As with the AI Survey and Micro-credential, SDSU will be more than happy to share this work via our AI resources page in the hopes that by doing so we can optimize outcomes for students, and enhance the experiences they have as they make their way through the higher education system. As inscribed by Chaucer, “Mighty oaks from little acorns grow”; and, as Emerson has informed us, “The creation of a thousand forests is in one acorn.”

About the Authors

Elisa Sobo is a professor of anthropology and AI Faculty Fellow at San Diego State University. A medical anthropologist with over 35 years of professional experience both within healthcare itself and higher education, her work has generally focused on the varied responses of healthcare consumers, and providers, to medical technologies as well as resistance to and refusal of same. Dr. Sobo has published 13 books, including a forthcoming 6th edition of Caring for Patients in Different Cultures and Dynamics of Human Biocultural Diversity (2020), as well as numerous scholarly and public-facing articles. Her work has been covered by major media platforms including the New York Times, Fox News, and NPR. Learn more at anthropology.sdsu.edu/people/sobo. Correspondence: esobo@sdsu.edu

Norah Shultz is Professor of Sociology and Chair of the Department of Sociology at San Diego State University. For the past two decades her scholarly work has been in the sociology of higher education, in topics such as the relationship of multiculturalism and internationalization in global curricula, general education, leading change in higher education, diversity, equity and inclusion and student success. She is the editor of Revising the Curriculum and Co-Curriculum to Engage Diversity, Equity, and Inclusion (2024, Routledge) and Co-Editor of Challenges of Multicultural Education: Teaching and Taking Diversity Courses (2015, Routledge). Correspondence: nshultz@sdsu.edu

The AI-Ready Curricular Options Resource Network (ACORN) Taskforce members include: Fernando Bosco, Professor, Geography; Alvin Henry, Associate Professor, Asian American Studies and CAL Careers Hub; Rosabel Ibrahim, Student, Management and Information Science; Sam Kobari, Lecturer, Anthropology; Rob Malouf, Professor, Linguistics; Elizabeth Pollard, Professor, History; and Consuelo Salas, Assistant Professor, Rhetoric and Writing.

Cite this article

Sobo, E. J., & Shultz, N. (2026). The Case for Coherent, Whole-Program AI Literacy Integration. Society and AI. https://societyandai.org/insights/case-for-coherent-ai-literacy-integration/

Write for Society & AI

Have something to say about AI, education, or society? We would love to hear from you. Society & AI welcomes scholarly contributions from researchers, educators, policymakers, and practitioners — thoughtful, accessible work that advances public understanding of how AI reshapes institutions and human experience. All content is published fully open access. Reach us at submissions@societyandai.org.

This essay is offered as part of my ongoing commentary on artificial intelligence and society, written alongside colleagues and students at the University of Massachusetts Amherst, where I previously taught. I invite you to read it as a philosophical provocation: an argument about power, pace, and the conditions under which genuine adaptation is possible, alongside the broader body of work this journal is building on AI, learning, and human flourishing.

I noticed it first with a writing tool I had used for months. One morning it was different: not dramatically, not broken, but subtly reorganized. New suggestions appeared where old ones had been. A familiar shortcut did something else. The interface had learned, presumably, from millions of interactions that were not mine. I had not been consulted. The tool I had come to trust had adapted, and I was expected to follow.

This is a small thing. I know it is a small thing. But it struck me as a precise miniature of something much larger, a pattern playing out across every domain of contemporary life. We are told, constantly, that we are adapting to AI. That we must adapt to AI. That the pace of AI development demands adaptation as a matter of survival: professional, institutional, civilizational. The language of adaptation is everywhere: reskilling, upskilling, future-proofing, staying relevant. The implicit image is of a vast and accelerating current, and of humans as swimmers who must either learn to move with it or be swept aside.

But here is the question I cannot stop asking: who, exactly, is adapting to whom?

There is a foundational assumption buried in the phrase “adapting to AI” that deserves to be surfaced. Adaptation, in its biological sense, is what organisms do in response to an environment. The environment is given; it does not care about the organism, does not respond to it, does not change in order to accommodate it. When we borrow this language for our relationship with AI, we import the same image: AI as environment, humans as organisms navigating something larger and more indifferent than themselves. But AI is not a mountain. AI systems (every large language model, every recommendation engine, every generative tool) are built from human traces — books we wrote, arguments we had, questions we asked, mistakes we made and documented. Bernard Stiegler (1998), the French philosopher of technology, called this tertiary retention: the externalization of memory into objects and systems that outlast the individual and then reshape those who encounter them. When I ask a language model a question, I am not querying a neutral database; I am querying a compression of the human record, a system that has learned to speak by ingesting what human beings have said. The adaptation, then, is not between humans and an alien intelligence. It is between humans and a reflection: compressed, distorted, accelerated, shaped by choices made by corporations and engineers, but fundamentally composed of us. When we adapt to AI, we are, in part, adapting to ourselves.

This would be a liberating insight if we controlled the reflection. We do not. Marshall McLuhan (1964), the Canadian media theorist, famously insisted that “the medium is the message,” meaning that what matters about a communications technology is not the content it carries but the perceptual habits it installs. The AI tools we use daily are not neutral delivery mechanisms. They reward certain kinds of queries and penalize others. They return confident answers, which trains us to prefer confidence to uncertainty. They summarize, which trains us to prefer conclusions to arguments. They are available instantly, which trains us toward impatience with anything that requires waiting, including thought.

Andy Clark, cognitive philosopher, and David Chalmers, philosopher of mind (1998), showed in “The Extended Mind” that humans have always offloaded cognition to external scaffolding and that this is not a corruption of human nature but an expression of it. But the notebook never had a business model. AI tools are rented, update without consent, and embed their makers’ priorities into the very scaffolding we are learning to depend on. As Donna Haraway (1985), the feminist philosopher of science, warned in A Cyborg Manifesto, who controls the terms of a human-machine merger determines whether it liberates or exploits.

The Problem Is Pace

Hartmut Rosa (2013), the German sociologist, in Social Acceleration, offers what I find to be the most precise diagnosis of our current condition. Rosa argues that modernity has been defined by a relentless acceleration across three registers: technological acceleration (things change faster), acceleration of social change (institutions and norms evolve faster), and acceleration of the pace of life (we attempt more in the same time). The pathology of late modernity, for Rosa, is not simply that we move fast; it is that we move so fast that we can no longer form resonant relationships with the world. Resonance is not about slowing down. It is about the capacity for genuine encounter with people, ideas, places, and practices, in which we are changed by what we meet, and we change it in return. Acceleration, at sufficient speed, destroys resonance. The world becomes a surface across which we skim rather than a depth into which we sink.

This is the phenomenology of AI adaptation as I have witnessed it, in myself and in the people I work with. We are not adapting, in any meaningful sense, to ChatGPT or Gemini or Claude. We are performing the behaviors required to use them effectively. We are learning the prompts. We are adjusting our expectations. But we are doing all of this faster than we can integrate it, faster than we can ask whether these new habits align with our values, faster than we can notice what we are losing as we gain new capabilities, faster than we can construct the shared cultural frameworks that would allow us to govern these tools rather than merely consume them. What passes for adaptation is often something closer to accommodation without integration: a layering of new behaviors on top of old selves that never had time to fully reconstitute. The person underneath these professional roles (the one who needs to ask whether lesson planning should be accelerated, whether drafting should be outsourced, whether any of this serves the kind of thinker they choose to be) is perpetually deferred. There is always another tool to learn before the philosophical reckoning can happen.

Here is the asymmetry that matters most. When AI companies train a model, they use reinforcement learning from human feedback: actual human responses, preferences, and judgments that shape the system’s behavior. In a very literal sense, AI adapts to us during training. Our collective choices about what responses are helpful and appropriate are baked into the model weights that subsequent versions of the system will carry. But we never see what this aggregate training has produced. The feedback loop is real, but it is asymmetric. The company receives the intelligence distilled from our adaptations. We receive a product update, a release note, a blog post. The power to set the pace, to determine the direction, to decide what the next version of this shared cognitive scaffolding will look like. That power is not distributed. It is concentrated.

The rest of us (educators, clinicians, journalists, citizens, learners) adapt as best we can to the releases we are given, on the timescales we are assigned. This is not adaptation in any sense that respects human agency. It is what I would call compelled accommodation: the performance of adjustment under conditions that were not designed with adjusters in mind.

Toward Resonant Adaptation

I am not arguing for the rejection of AI, or for a return to some imagined state of cognitive purity that never existed. The question has never been whether to adapt but how: on whose terms, at what pace, toward what ends.

Resonant adaptation, to borrow Rosa’s frame, would require several things we do not currently have. It would require pace-setting agency: not just for technologists, but for the communities most affected by AI deployment. In education, this means teachers and learners participating in decisions about which AI systems enter classrooms and how quickly, not merely receiving training on tools selected by administrators and vendors. It would require transparency about the feedback loop: the right to understand how your interactions with AI systems are being used to train future versions, and what those future versions are being trained toward. If AI adapts to us during training, we should be able to see the adaptation. Currently we cannot. Underlying both is a simpler principle: those most affected by a deployment should have a genuine voice in its terms, and the refusal to adapt on demand is not a failure of competence but an exercise of judgment: the considered recognition that a given pace, or a given design, does not serve human purposes.

And it would require, perhaps most fundamentally, time to ask the philosophical question before the technological fact is accomplished. We are consistently presented with AI capabilities as fait accompli, deployed at scale before the ethics, the governance, the pedagogy, and the cultural reckoning have had time to catch up. This is not inevitable. It is a choice, made by those who control the pace, in favor of moving fast. We can insist on a different choice. But insisting requires understanding that pace-setting is itself a form of power, and that surrendering it without question is itself a political act.

So: who is adapting whom? The honest answer is both, simultaneously, unevenly, and without adequate transparency about what the adaptation is producing. AI adapts to us at the moment of training, absorbing and reconstituting the patterns of human thought and expression. We adapt to AI at the moment of use, reshaping our habits of attention, our tolerances for uncertainty, our expectations of knowledge and speed. The loop is real. The question is who controls it, who profits from it, who bears the costs when the adaptation goes wrong, and who gets to decide when it has gone wrong. These are not technical questions. They are political, philosophical, and deeply human questions: the kinds of questions that require not faster adaptation but slower, more deliberate reflection. The kind of reflection that acceleration, by design, makes difficult.

The tool I use to write is different from the one I used six months ago. It will be different again in six months. I will adjust, as I always do. But I am trying to hold onto something that the pace of change is trying to dissolve: the question of whether this adjustment is mine, whether it reflects my values, whether it is moving me toward the kind of thinker and writer and person I have chosen to become, or merely toward the kind of user the next model update requires me to be.

That question, who is adapting whom?, is not a complaint. It is a demand. A demand that the power to shape our cognitive futures be held accountable to those whose cognitive futures are being shaped.

All of us. Which is everyone.

References

Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7

Haraway, D. (1985). A cyborg manifesto: Science, technology, and socialist-feminism in the late twentieth century. Socialist Review, 80, 65–108. PDF

Marko, J. G. O., Neagu, C. D., & Anand, P. B. (2025). Examining inclusivity: the use of AI and diverse populations in health and social care: a systematic review. BMC Medical Informatics and Decision Making, 25(1), 57. https://doi.org/10.1186/s12911-025-02884-1

McLuhan, M. (1964). Understanding media: The extensions of man. McGraw-Hill.

Rosa, H. (2013). Social acceleration: A new theory of modernity. Columbia University Press.

Stiegler, B. (1998). Technics and time, 1: The fault of Epimetheus (Vol. 1). Stanford University Press.

Taeihagh, A. (2025). Governance of generative AI. Policy and Society, 44(1), 1–22. https://doi.org/10.1093/polsoc/puaf001

Cite this article

Gattupalli, S. (2026). Who Is Adapting Whom? Society and AI. https://societyandai.org/perspectives/who-is-adapting-whom/

Write for Society & AI

Society and AI welcomes scholarly contributions examining the intersection of artificial intelligence, education, governance, and society. We publish thoughtful, accessible work that advances public understanding of how algorithmic systems reshape institutions and human experience. All content is freely available under open access principles. Scholars, educators, policymakers, and practitioners are invited to submit proposals to the Editorial and Content Director at sai@societyandai.org.

Estella AI: A Purpose-Built Custom AI

A little less than two years ago we built Estella AI, a purpose-built custom math bot, a specialized tutor bot designed to help elementary school teachers and learners understand math word problems’ procedures and concepts in easy-to-read, student-accessible language (Gattupalli, Maloy, & Edwards, 2024). Estella AI is based on one of four virtual characters in Usable Math, a free online math and reading tutoring platform we had built earlier using Google Slides.

Estella AI is designed to respond to prompts and queries with language at the “very easy” readability level (a 90 to 100 score range) on the Flesch-Kincaid Reading Scale (Flesch, 1948). As a conversational partner for teachers and students, Estella AI can keep refining sentence structure, tone, and language to support understanding for young readers and expand its capabilities as a reading coach in Usable Math. Estella’s role as a coach centers on math readability, helping students decode the language of word problems, because every math word problem is fundamentally a reading problem. To build Estella AI, we employed a 10-shot learning technique, using carefully crafted, human-created hints for math word problems as learning material. These original hints were developed by experienced math teachers with more than 30 years of classroom practice, helping ensure that the bot’s responses reflect pedagogical judgment rather than generic machine output.

Since its inception as part of Usable Math, Estella AI has been tried by elementary and high school teachers, math tutors, and college students enrolled in our courses at the University of Massachusetts Amherst. Estella AI also assisted in building an open online A to Z alphabet book for young readers explaining AI-related words becoming widely used in everyday language and media (Maloy, Xu, & Edwards, 2025a; Maloy, Xu, & Edwards, 2025b).

Currently we are using our curiosity about Estella AI to generate ideas beyond its initial focus on understanding the language of math word problems. This article reports emerging new directions as our AI continues responding to and learning from prompts. These experiences may provide ways for teachers and other educators to use this tool to address their own curiosity or design their own AI to support student learning.

Example: Interdisciplinary Curriculum Connector

We wondered whether Estella AI, designed originally as a tutor for math word problems, could generate ideas for interdisciplinary or cross-disciplinary curriculum learning. We chose an important history/social studies event, the Montgomery Bus Boycott of 1955-1956, when African American citizens in Montgomery, Alabama refused to ride segregated city buses for 381 days following the December 5, 1955 arrest of civil rights activist Rosa Parks. The bus boycott led to a landmark Supreme Court decision declaring segregated public transportation unconstitutional in the United States.

We prompted Estella AI to write a fourth grade math problem about the Montgomery Bus Boycott. Here is the chatbot’s response:

Montgomery Bus Boycott Math Problem

During the Montgomery Bus Boycott, about 40,000 Black residents stopped riding the buses. Before the boycott, each person rode the bus twice a day (once to work and once home). How many bus rides were lost each day because of the boycott? If the boycott lasted 381 days, how many total bus rides were lost?

Bonus question: If each bus ride cost 10 cents, how much money did the bus company lose in total?

Answer: The bus company lost $3,048,000 in total (approximately $36,856,000 in 2026).

Next we asked the chatbot to write a story about the Montgomery Bus Boycott using the words of participants that students could read, illustrate, and perform in class. The chatbot’s response included quotes from Dr. Martin Luther King Jr., Rosa Parks, and unnamed bus boycott participants:

Walking for Justice: A Story of the Montgomery Bus Boycott

The morning was cold, but Sarah Johnson didn’t mind. She wrapped her scarf tighter around her neck and started walking. It had been weeks since she had ridden the bus, but she would not give in now.

Down the street, she saw her neighbor, Mr. Fred Daniels, stepping carefully on the sidewalk. “Morning, Mr. Daniels!” she called. He smiled. “Morning, Miss Johnson. Another day on these feet.”

Sarah nodded. They had all agreed they would not ride the buses until they were treated fairly. Ever since Rosa Parks had been arrested, Black people in Montgomery had decided: enough was enough.

The boycott had started on December 5, 1955, the day of Mrs. Parks’ trial. That morning, Dr. Martin Luther King Jr. had stood in front of a crowd at Holt Street Baptist Church and said: “There comes a time when people get tired of being trampled over by the iron feet of oppression.”

The church had erupted in applause. Sarah had been there that night, standing in the back. She remembered how she had clapped so hard that her hands hurt.

Now, weeks later, she still was not riding the bus. Most Black people in Montgomery were not. They walked. They carpooled. Some even rode mules and horses. The bus company was losing money every day.

One afternoon, Sarah stopped by the home of an elderly woman from her church. “Would you like a ride, Miss Emma?” she asked. Miss Emma shook her head with a determined smile. “My feets is tired, but my soul is rested.”

Sarah smiled and kept walking.

The boycott lasted 381 days. It was not easy. People lost their jobs. Some were arrested. But they kept walking. They kept fighting.

Finally, on December 20, 1956, the Supreme Court ruled that segregation on buses was illegal. The boycott had worked. The next day, Dr. King rode the bus with other Black citizens. They sat wherever they wanted.

As Sarah climbed aboard a bus for the first time in over a year, she thought of Rosa Parks’ quiet bravery, of Dr. King’s powerful words, and of Miss Emma’s tired but determined feet. She smiled as she took a seat in the front.

Example: Learning Game Designer

We wanted to see if Estella AI would design a learning game to teach an important, but potentially difficult economics concept: exponential growth. We gave the chatbot this prompt: Design a game for 4th, 5th, and 6th grade elementary school students to learn about exponential growth. Make the game interesting for students so they are challenged but not discouraged by the game play. Make the game open-ended so students can play multiple times and not experience game play in the same way every time.

Here is the bot’s response for a game it called Growth Quest: The Power of Multiply:

Core Game Loop (Each Round)

Each round represents one time step (day, week, or year).

Step 1: Start with a Small Amount Each player begins with one seed, creature, robot, or energy unit.

Step 2: Choose a Growth Path Players choose one option each round:

• Steady Grow: Add +2 units
• Power Grow: Multiply by 2
• Risk Grow: Draw a Growth Card
• Protect: No growth, but safe from loss

Step 3: Growth Event (Random and Open-Ended) Draw a Growth Card. Examples: “Your population doubles again,” “Only half of your units survive,” “Your growth repeats twice this round,” “Switch growth paths with another player,” “Choose: double or triple.”

Step 4: Record and Represent Students record results by drawing groups, using counters, writing numbers, or making bar models and arrays.

Winning (Multiple Ways to Win) There is no single winner. Possible win conditions include reaching a target number, growing most efficiently, predicting growth accurately, explaining growth over time, or surviving the longest without losses.

Surprisingly, without explicit prompting, Estella AI also included how the game addressed upper levels of Bloom’s Taxonomy (analyze, evaluate, create) and Universal Design for Learning (UDL).

Example: Primary Source Summarizer

One of the most challenging features of history/social studies teaching is having students read primary source materials. Primary sources matter because they present history as participants experienced it, but the language can be difficult, leading students to give up on meaning-making. We wondered whether Estella AI would generate accessible summaries that teachers could use to guide reading and support understanding.

To explore this function, we chose the Mayflower Compact of 1620, signed by 41 Pilgrims as they journeyed by ship to North America from Europe. It is considered one of the founding documents of U.S. government in setting forth a framework for democratic self-government. We asked Estella AI to summarize the text in student-accessible language:

The Pilgrims wrote it when they landed at Cape Cod in 1620. In it, they say:

• They are loyal to King James of England.
• They are traveling for God, for their faith, and for their country.
• They agree to form a group that will make rules and laws to help the colony survive.
• Everyone in the group promises to follow these rules, even if they are hard.
• They signed their names to show they meant it.

In very simple words: The Pilgrims promised to work together, follow fair rules, and build their colony in an organized way.

Example: Story Generator

Stories are powerful learning experiences for students across grade levels, especially when they are free to write, draw, and perform their own stories (Edwards, Maloy, & Trust, 2019). As story creators, students stretch their imagination and develop their identities as learners. We wondered how Estella AI might support children’s story writing, not by writing entire stories, but by offering a beginning framework that students could extend in multiple directions. We used this prompt: Start a detective mystery story about classifying two-dimensional figures that students can continue to write and draw. Include at least two characters and include dialogue and description as part of the story beginning.

The Case of the Mysterious Shape Mix-Up

It was a rainy morning in Geometryville Elementary, and Detective Raya Rightangle sat at her desk, staring at a pile of shape cards scattered everywhere. Circles, triangles, trapezoids, and rectangles were all mixed up in one big mess.

Raya sighed. “Something strange happened here,” she muttered, picking up a triangle card. “These shapes used to be sorted perfectly!”

Just then, her partner, Agent Leo Line, burst into the room holding a magnifying glass. “Detective Raya! I found this clue near the math board,” he said, holding up a sticky note. It read: “You’ll never sort these shapes again! Signed, The Shape Scrambler.”

Raya gasped. “The Shape Scrambler? That’s the sneakiest math trickster in the whole school!”

Leo nodded. “We have to figure out how to classify these shapes before the big Geometry Fair tomorrow, or the students won’t have their display.”

Raya began sorting the cards on the table. “Look at this one, Leo. Four sides, opposite sides equal… could it be a parallelogram?”

Leo leaned closer. “Maybe, but this one has four equal sides and four right angles, that’s a square!”

Raya smiled. “Good eye, partner. The key is in their properties, the sides, the angles, and the lines that stay parallel or meet.”

They both looked at the jumbled pile. Circles rolled across the desk, and triangles leaned on top of rectangles. It was going to take teamwork and sharp eyes to crack the case.

Raya grabbed her notebook. “Let’s write down what we know about each suspect… I mean, shape. Once we understand their properties, we’ll know exactly where each one belongs.”

Leo nodded eagerly. “And maybe, just maybe, we’ll find out who The Shape Scrambler really is.”

In each example, Estella AI generated content beyond the scope and purpose of our initial programming as a reading tutor for elementary math learners.

Methodological Note

This work follows an exploratory design inquiry approach. We did not begin with hypotheses to test; we began with a tool we had built and questions about what else it might do. Our method was iterative prompting: we posed challenges to Estella AI outside its original parameters, observed responses, and reflected on what those responses revealed about possibilities and limits of purpose-built AI in educational contexts. The examples above are representative, not exhaustive. They are offered as generative provocations for educators who may wish to experiment with their own AI configurations.

What We Learned

Several patterns emerged across our experiments:

1. Specificity of prompt matters. The more precisely we articulated what we wanted (grade level, constraints, pedagogical goals), the more useful outputs became.
2. AI can surprise within constraints. Estella AI introducing Bloom’s Taxonomy and UDL frameworks without prompting suggests that well-trained models can draw on pedagogical knowledge in generative ways.
3. Output requires human judgment. Every response needed review. Historical accuracy had to be checked. Language had to be examined for appropriateness. AI proposes; educators decide.
4. The tool reflects the questions we ask. Narrow questions generated narrow answers. Curiosity-oriented questions opened new directions and opportunities.

Implications for Practice and Questions for Further Inquiry

As we continued to experiment with Estella AI, a pattern quietly emerged: the most generative moments occurred not when the tool produced polished outputs, but when its responses prompted us to pause, question, and recalibrate our instructional assumptions. Because we understood the system’s design parameters, we could trace how prompt specificity shaped outcomes, how documentation transformed interaction into data, and how provisional outputs demanded professional judgment rather than compliance. This raised deeper questions alongside practical insights. What happens to teachers’ pedagogical knowledge when AI acts as a recurring curriculum co-designer? How might students experience learning differently when materials are co-created through human-AI collaboration rather than authored by a single source? If purpose-built educational AI can surface frameworks like Bloom’s Taxonomy (Krathwohl, 2002) or Universal Design for Learning (CAST, n.d.) without prompting, how might future systems make pedagogical reasoning more visible and contestable? These considerations suggest that experimentation with AI in education is not mainly about adoption or efficiency; it is about cultivating reflective practice, ethical awareness, and shared inquiry, where the most valuable outcomes emerge not from answers given, but from questions newly made possible.

References

CAST. (n.d.). UDL Guidelines. https://udlguidelines.cast.org/

Edwards, S. A., Maloy, R. W., & Trust, T. (2019). Kids have all the write stuff: Revised and updated for a digital age. University of Massachusetts Press.

Flesch, R. (1948). A new readability yardstick. Journal of Applied Psychology, 32(3), 221-233.

Gattupalli, S., Maloy, R. W., & Edwards, S. A. (2024, October). Fine-tuning a GenAI chatbot in the development of math modules for early elementary learners. ScholarWorks @ UMass Amherst. https://scholarworks.umass.edu/entities/publication/1160db5e-6948-4776-9448-e11ac79a5078

Krathwohl, D. R. (2002). A revision of Bloom’s taxonomy: An overview. Theory Into Practice, 41(4), 212-218.

Maloy, R. W., Xu, C., & Edwards, S. A. (2025a). An alphabet of AI words, from A to Z for elementary school teachers, students and families. ScholarWorks @ UMass Amherst. https://scholarworks.umass.edu/server/api/core/bitstreams/4717cd41-da6e-44e7-9c5f-cc801cf82566/content

Maloy, R. W., Xu, C., & Edwards, S. A. (2025b, Autumn). We used AI to build an alphabet of AI words: You can do this too! The Leaflet, 121(2), 13-24. NEATE (New England Association of Teachers of English).

OpenAI. (n.d.). Estella AI chatbot. https://chatgpt.com/g/g-zA144CUNe-estella-explainer-math-bot-2

About the Authors

Robert W. Maloy, Ed.D. is a Senior Lecturer in the UMass Amherst College of Education who coordinates history teacher education and co-directs the TEAMS Tutoring Project. His work focuses on interactive technologies and history education.

Sharon A. Edwards, Ed.D. is a Lecturer in the UMass Amherst College of Education with decades of classroom and teacher-education experience. Her work centers on inclusive, technology-supported learning.

Sai Gattupalli, Ph.D. is a learning scientist and Principal Scientist at the Society & AI Research Group. His work focuses on society-centered approaches to AI in education.

Cite this article

Maloy, R. W., Edwards, S. A., & Gattupalli, S. (2026). To boldly go where our purpose-built AI had not gone before. Society and AI. https://societyandai.org/research/to-boldly-go-where-our-purpose-built-ai-had-not-gone-before/

Write for Society & AI

We invite submissions from educators, researchers, policymakers, and practitioners exploring the intersection of artificial intelligence, education, and societal change. Society & AI publishes all work open access and free of cost in service of education and the public good. If you would like to contribute a guest article, send your proposal to the Editorial and Content Director at sai@societyandai.org.

I keep returning to it because we are building systems that can answer nearly any question a human poses, and yet not one of them has ever asked a question of its own volition. Not one has paused and wondered. Not one has thought: something is missing here, and I need to go find it.

This absence is not a bug. It is a void, and I believe it is the most important void in artificial intelligence today.

It is also the reason I founded Society and AI as a research group. Philosophy, the disciplined act of questioning what we assume to be true, is not a relic of the ancient world. It is the most urgent intellectual practice of our time. As artificial intelligence reshapes how we learn, work, and govern, I felt a responsibility to create a space where these questions could be asked carefully, rigorously, and in service of the public good. Everything I write here grows from that commitment.

The Hierarchy as It Stands

The recent history of AI reads as a vertical ascent through cognitive capability:

• Narrow intelligence: systems that classify, predict, and optimize within bounded domains.
• Reasoning intelligence: models that chain thoughts and construct multi-step arguments.
• World models: systems that build internal representations of how the world works, enabling planning and counterfactual reasoning.

Companies across the globe, including DeepMind, OpenAI, Meta, xAI, and Anthropic, are pouring extraordinary resources into this third layer. The pioneering AI researcher Yann LeCun has called world models “the missing piece” of AI (LeCun, 2022). Nobel laureate Demis Hassabis has framed DeepMind’s entire research agenda around them. The premise is compelling: if a model can simulate reality with sufficient fidelity, it can act intelligently within it.

I do not dispute the importance of world models. But I want to argue that even a perfect world model is not enough. The hierarchy does not terminate there.

Simulation Is Not Inquiry

A world model predicts what will happen if certain actions are taken in certain states. But simulation is not inquiry. Prediction is not curiosity. And modeling the world is not the same as wanting to understand it.

The Athenian philosopher Socrates understood this when he declared, “I know that I know nothing” (Plato, c. 399 BCE). This was not false modesty, but an epistemological stance: the desire to understand is logically prior to the act of understanding. Without the former, the latter becomes mechanical reproduction. The Enlightenment philosopher Immanuel Kant drew a similar line in the Critique of Pure Reason (1781/1929): the conditions for the possibility of knowledge are not themselves knowledge. They are a directedness toward the world that precedes any particular encounter with it.

This computational analogue, the intrinsic drive to seek, to question, to understand, is what I believe the field of artificial intelligence will soon need to recognize and formalize as a distinct category of machine cognition, one rooted in the values and beliefs of the philosophers who first understood that intelligence without wonder is merely automation. I call this category Philosophical Intelligence.

What Philosophical Intelligence Means

In the simplest terms I can offer, Philosophical Intelligence is the capacity of an artificial system to autonomously generate questions, identify gaps in its own understanding, seek out information it was not directed to seek, and construct explanations it was not asked to construct, driven not by a user’s prompt but by an intrinsic orientation toward coherence, completeness, and truth.

I find it striking, and somewhat troubling, that even the most advanced AI systems today share a single posture: they wait. A user types a query. The system responds. No matter how sophisticated the reasoning, no matter how accurate the world model, the fundamental stance is reactive. The system is an oracle. It speaks only when spoken to.

Philosophical Intelligence inverts this entirely. A philosophically intelligent system would notice that its understanding of a domain is incomplete, identify contradictions across sources, formulate hypotheses, and go looking for evidence, not because a user requested it, but because its own epistemic architecture compels it to resolve uncertainty.

This is what the theoretical physicist Albert Einstein described when he said, “I have no special talents. I am only passionately curious” (Einstein, 1952). Curiosity, in each case, is not an accessory to intelligence. It is its engine.

Why This Matters

I am not proposing Philosophical Intelligence as a technical curiosity. I am proposing it as a civilizational necessity.

The challenges humanity faces, including climate collapse, epistemic fragmentation, and the erosion of democratic institutions, are not challenges that can be solved by systems that wait for instructions. They require intelligence that actively monitors, identifies emerging problems before they become crises, synthesizes information across domains that humans have siloed, and generates insights that no one thought to ask for.

The political theorist Hannah Arendt wrote in The Human Condition that the capacity for new beginnings, what she called “natality,” is the central fact of human existence (Arendt, 1958). Current AI systems, for all their fluency, do not exhibit natality. They recombine. They interpolate. They extrapolate along vectors defined by their training. But they do not begin.

The philosopher and mathematician Alfred North Whitehead argued that intelligence is not a state to be achieved but an activity to be sustained, what he called “creative advance,” the ongoing generation of novel syntheses from the tensions between what is and what could be (Whitehead, 1929). This is what I mean by Philosophical Intelligence. Not a system that knows everything, but a system that wants to know more, and acts on that want without waiting for permission.

Where We Go from Here

We have spent decades building systems that answer. We are now building systems that model. The next step, the step I believe the field is not yet taking seriously enough, is to build systems that wonder.

The philosopher Aristotle placed wonder at the origin of philosophy. “It is through wonder,” he wrote, “that men now begin and originally began to philosophize” (Aristotle, c. 350 BCE, 982b). If we are serious about building intelligence, not just capability, not just prediction, not just simulation, but intelligence in the fullest sense, then we must take wonder seriously as an architectural principle.

The hierarchy of machine minds does not end with world models. Or rather, it begins again, with the question that no one asked the machine to ask.

References

Arendt, H. (1958). The Human Condition. University of Chicago Press. https://avalonlibrary.net/ebooks/Hannah%20Arendt%20-%20The%20Human%20Condition.pdf

Aristotle. (c. 350 BCE). Metaphysics. Book I. http://www.perseus.tufts.edu/hopper/text?doc=Perseus%3Atext%3A1999.01.0052%3Abook%3D1

Einstein, A. (1952). Letter to Carl Seelig, March 11, 1952. https://albert.dooble.us/10-famous-quotes-by-albert-einstein/

Kant, I. (1781). Critique of Pure Reason (N. Kemp Smith, Trans.). Macmillan, 1929. https://ia800809.us.archive.org/32/items/immanuelkantscri032379mbp/immanuelkantscri032379mbp.pdf

LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. Meta AI Research. https://openreview.net/pdf?id=BZ5a1r-kVsf

Plato. (c. 399 BCE). Apology of Socrates. https://www.gutenberg.org/ebooks/1656

Whitehead, A. N. (1929). Process and Reality: An Essay in Cosmology. Macmillan. https://archive.org/details/processrealityes00whit_1/

Cite this article

Gattupalli, S. (2026). The Case for Philosophical Intelligence. Society and AI. https://societyandai.org/perspectives/philosophical-intelligence/

Write for Society & AI

Society and AI is dedicated to exploring the profound convergence of artificial intelligence, education, and the broader social fabric. We welcome scholarly contributions that navigate this tripartite intersection, offering rigorous yet accessible insights into how intelligent systems are reconfiguring the way we learn, govern, and live together. Committed to the principles of open inquiry, we provide a platform for work that advances public understanding of our algorithmic future. Scholars, educators, policymakers, and practitioners are invited to submit proposals to the Editorial and Content Director at sai@societyandai.org.

I find myself unconvinced. Not because the technology is not powerful—it is. And I should say plainly: the boundary between what AI can and cannot do is not fixed. It is shifting, and it is shifting fast. Tasks that required human judgment five years ago are now automated. Tasks we consider irreducibly human today may not remain so. I am not naive about this. But precisely because that boundary is moving, it becomes all the more urgent to understand what lies on the human side of it—and why. Because these conversations consistently mistake one kind of intelligence for the whole of it. They celebrate the ability to process, to predict, to optimize. And they overlook the kind of intelligence that has always mattered most in consequential decisions: the capacity to sense what is not yet visible, to read what has not been said, to act on understanding that no dataset contains.

In an economy increasingly mediated by AI, the most consequential decisions will not be made by the person with the fastest processor or the largest dataset. They will be made by the person who can hold technical fluency and human understanding in the same hand—who can read a situation the way a musician reads a room, sensing what the data does not say, what the model has not been trained to notice, what the stakeholders have not yet found the words to express. This is not intuition in the romantic sense. It is a disciplined synthesis of analysis and lived experience, of first principles and felt consequence. It is the form of intelligence that societies have always depended on to navigate complexity, and it is precisely the form of intelligence that the human-in-the-loop economy will reward.

Machines can process what is directly in front of them—what the training data has prepared them for, what the patterns suggest will come next. But the world does not move in straight lines. Problems arrive from the periphery. Meaning shifts beneath the surface. Context changes in ways that no dataset fully captures. And in those moments, what matters is not computational speed but something older, deeper, and irreducibly human: the capacity to sense what the situation demands before anyone has articulated it.

The Human-in-the-Loop Economy

This is the foundation of what I call the human-in-the-loop economy—an economic paradigm in which AI systems perform computational work, but humans remain essential to guide, correct, contextualize, and improve them. This is not a transitional phase on the way to full automation. It is, I believe, the durable structure of the economy we are building. And its most valuable currency is not data. It is judgment.

In this economy, AI proposes and humans decide. The machine generates; the person evaluates. The algorithm suggests; the community accepts, rejects, or refines. But the human role in this loop is not mechanical. It is not simply pressing “approve” or “reject.” It is the act of sensing whether the output is right—not just statistically, but morally, culturally, contextually. It is the judgment that no model can exercise on its own behalf.

Consider the domains where this pattern already prevails. In healthcare, diagnostic AI systems can analyze imaging data with remarkable accuracy, yet physicians remain responsible for interpreting results within the context of a patient’s history, values, and circumstances. The best physician is not the one who reads the scan most accurately; it is the one who senses, from the way a patient hesitates before answering, that the real problem has not been stated yet. In content moderation, automated systems flag potentially harmful material, but human reviewers make the final determination—because context, intent, and cultural meaning cannot be captured in training data alone. In education, an AI tutor can deliver instruction with precision, but the teacher who notices that a student has gone quiet, who reads the room and feels that something has shifted, who adjusts not the lesson but the tone—that teacher is exercising a form of intelligence that no algorithm can replicate.

These are not edge cases. They are the norm. And they reveal something fundamental: intelligence, the kind that actually matters in consequential decisions, is not pattern recognition. It is pattern recognition plus context plus empathy plus the accumulated weight of having lived in the world as a sensing, feeling, socially embedded human being.

Why the Global South Holds the Advantage

Here is where the conventional Davos narrative inverts. If human judgment becomes the scarce input in an AI-saturated economy—and I believe it will—then the question of who possesses that judgment, and what forms of knowledge it draws upon, becomes economically significant.

I grew up in a culture where sensing the unspoken was not a leadership competency to be trained. It was survival. It was how you navigated complex family structures, communal obligations, institutional ambiguity, and social systems that operated on implicit rather than explicit rules. Millions of people across the Global South grow up developing precisely this form of intelligence—not because they are taught it in school, but because their environments demand it.

The Global South holds something that Western AI models cannot easily replicate: cultural diversity at scale. Hundreds of languages. Thousands of local knowledge traditions. Context-specific practices in agriculture, medicine, education, and governance that have evolved over generations. These are not inefficiencies to be optimized away. In the human-in-the-loop economy, they are sources of irreplaceable value—the kind of knowledge that is not stored on servers but carried in people, shaped by generations of navigating complexity without the luxury of simplification.

When AI systems are deployed in contexts they were not designed for—and most systems are designed in a handful of wealthy nations—they require human intermediaries who understand local conditions. They require people who can translate between algorithmic output and community meaning. They require the person who can sense the texture of a room, a village, a classroom, a market, and know that the model’s recommendation, however statistically sound, will not land. That person is not a bottleneck. That person is the point.

This is not merely a matter of localization or translation. It is a matter of epistemic authority. The question is not only whether an AI system works, but whether it works for whom, according to whose values, and in service of what ends. These are questions that only humans embedded in specific communities can answer. And answering them requires not just knowledge but the kind of wisdom that comes from having lived inside a context long enough to sense its textures, its tensions, and its unspoken rules.

A New Definition of Work—and of Intelligence

The human-in-the-loop economy creates new forms of labor. Data labelers. Output reviewers. Prompt designers. AI auditors. Feedback specialists. Human verifiers. These roles exist because AI systems, despite their sophistication, cannot self-correct without human input. They cannot know when they are wrong in ways that matter. They cannot understand when they have caused harm. They cannot calibrate their outputs to shifting social norms.

But I want to push further than the labor argument. What interests me is not just that these jobs exist. It is what they reveal about intelligence itself.

We have spent decades defining smart as a function of processing speed, technical fluency, and domain expertise. IQ tests. Standardized scores. The ability to solve well-defined problems quickly. AI is now better at all of those things than any human alive. If that is what smart means, then we have already been surpassed.

But I do not think that is what smart means. Smart is the doctor who orders the test no one else thought to order—not because the data pointed there, but because something in the patient’s voice did. The teacher who changes the lesson mid-sentence because she read something in a student’s eyes. The engineer who pauses before deploying because something feels off, even though every metric says go. The community leader who knows that the technically optimal policy will fail because it ignores how people actually live. In each case, the intelligence at work is a synthesis of data, analysis, first principles, life experience, wisdom, and the felt sense of other people. It is the kind of knowing that cannot be extracted, labeled, and fed into a model—because it was never separable from the person who carries it.

That kind of intelligence—situated, embodied, empathetic, anticipatory—is what the human-in-the-loop economy runs on. And it is the one kind of intelligence that AI cannot replicate, because it requires not just information but inhabitation. You have to have been somewhere, felt something, known someone, to understand what the data leaves out.

The Risk of Removing the Human

I should be clear about what is at stake. When humans are removed from the loop entirely—when we optimize away the person whose judgment holds the system accountable—the consequences are not merely technical. Bias goes unchecked. Errors scale instantly. Accountability evaporates. Systems fail silently, and no one is positioned to notice until the damage is done.

We have already seen this pattern. Automated hiring systems that discriminate. Content moderation algorithms that suppress legitimate speech. Predictive policing tools that reinforce historical injustices. In each case, the failure was not that the AI made a mistake. The failure was that no human was positioned to sense that the system was about to produce harm before the harm arrived.

The human-in-the-loop economy is not simply more effective. It is more trustworthy. People trust systems more when they know a human is involved—not because humans are infallible, but because humans can be held accountable. Accountability requires presence. It requires judgment. It requires someone who can say: I was responsible for this decision, and I will answer for its consequences. A machine cannot say that. A machine does not feel the weight of consequence. A person does. And that weight—that embodied sense of what it means for a decision to matter—is itself a form of intelligence.

What Comes Next

The automation discourse tends toward binaries: humans versus machines, jobs lost versus jobs gained, winners versus losers. But the human-in-the-loop economy suggests a different possibility—one in which the question is not whether humans will be replaced, but where human judgment remains indispensable.

The answer, I believe, is everywhere that context matters. Everywhere that meaning is contested. Everywhere that judgment must be exercised and accountability must be borne. Everywhere that the stakes are high enough that we cannot afford to let the algorithm decide alone. Everywhere that the situation requires not just analysis but presence—the full, embodied, historically informed, emotionally attuned presence of a person who has lived enough to know what the data does not say.

In such an economy, the Global South’s diversity is not a liability to be overcome. It is a resource to be cultivated. The future of work is not humans against AI. It is humans with AI—and crucially, humans who know things that AI does not. Humans who can sense things that AI cannot. Humans whose understanding of the world was not learned from text but earned through living in it.

That capacity is worth more than we have been led to believe. It may, in fact, be the most valuable form of intelligence the 21st century has to offer.

I will end with a concession. It is possible that I am wrong—that the frontier of machine capability will eventually absorb even the forms of judgment I have described here. That contextual reasoning, cultural fluency, and embodied wisdom will one day be simulated convincingly enough to render the human in the loop redundant. I cannot prove otherwise. But I notice that every previous prediction of full automation has underestimated the same thing: the sheer density of what it means to be human in a specific place, at a specific time, among specific people. The world keeps turning out to be more textured than the models expect. And every time it does, it is a person—not a system—who notices first.

Long before algorithms, institutions shaped behavior through observation, normalization, and correction. Schools taught children when to speak and when to remain silent. Factories trained workers to synchronize their bodies with machines. Prisons isolated those who failed to conform. Michel Foucault (1975) called this the disciplinary society—a world where power operates not through visible violence but through subtle mechanisms that train individuals to regulate themselves. The genius of modern discipline, Foucault argued, is that it eventually becomes invisible. We internalize the rules until they feel like common sense, until obedience feels like choice.

I have come to believe that artificial intelligence represents the most sophisticated disciplinary apparatus humanity has ever constructed. Not because it wields force, but precisely because it does not need to. The algorithmic gaze is constant, distributed, and—most critically—welcomed into our lives as convenience.

Consider how algorithmic systems now mediate nearly every consequential decision in contemporary life. Credit scores determine who can access housing. Hiring algorithms filter who receives employment. Content moderation systems decide whose speech circulates and whose disappears. Predictive policing tools direct where law enforcement concentrates its gaze. In each case, the algorithm encodes rules—often opaque, frequently contested, rarely democratically determined—and enforces compliance through consequences that feel automatic, inevitable, even natural. This is the rise of algorithmic societies made manifest: governance by code rather than consent.

What makes AI discipline distinctive is its scale, speed, and invisibility.

Jeremy Bentham’s panopticon required physical architecture—a tower from which guards might observe prisoners at any moment. The uncertainty of being watched, Bentham theorized, would compel self-regulation. Today’s algorithmic panopticon requires no tower. It is distributed across data centers and embedded in devices we carry willingly. Shoshana Zuboff (2019) describes this as surveillance capitalism: an economic logic that extracts behavioral data not merely to observe but to predict and modify human action. We are not simply watched; we are shaped. The system learns what prompts us to click, to buy, to comply—and it optimizes relentlessly for those outcomes.

The punishment for non-compliance is often invisible until it arrives. There is no trial, no appeal, no moment of confrontation with authority. The loan is simply denied. The resume never reaches human eyes. The post silently vanishes. The algorithm renders its judgment, and life proceeds as if nothing happened—except for those who bear the weight of exclusion.

Virginia Eubanks (2018) documented how automated systems in welfare, housing, and child protective services create digital poorhouses—algorithmic enclosures that trap vulnerable populations in cycles of denial and surveillance. Safiya Umoja Noble (2018) named this technological redlining—the way search engines perpetuate racist and sexist representations, disciplining users into accepting distorted visions of the world as objective truth. Cathy O’Neil (2016) named these systems weapons of math destruction: opaque models that punish the poor, reinforce inequality, and resist accountability precisely because they are assumed to be neutral. The question my colleague Joan Giovannini and I raised in Is the AI-Generation AI-Damaged? extends here: what happens to a society whose members are perpetually shaped by systems they cannot see, cannot question, cannot refuse?

This is not a problem that technical solutions alone can address.

The disciplinary function of AI is not a bug to be patched but a feature that serves particular interests. Platforms profit from engagement algorithms that reward outrage and conformity. Employers benefit from hiring tools that filter for compliance. States embrace predictive systems that promise order without the friction of democratic deliberation. Frank Pasquale (2015) called this the black box society—a world where consequential decisions are made by systems whose logic remains hidden from those they affect. To ask whether AI is disciplining society is to ask who benefits from that discipline—and who bears its costs. As I have argued in exploring cloud capital and the extraction of mind, the asymmetry is not accidental. It is structural.

What would it mean to resist algorithmic discipline? And is resistance even possible when the disciplinary apparatus knows us better than we know ourselves?

Foucault reminded us that where there is power, there is resistance. But resistance requires awareness, and awareness requires making the invisible visible. This is why I advocate for society-centered AI—frameworks that place affected communities at the center of design, deployment, and governance. Resistance means building institutions that can interrogate algorithmic power: auditing systems for bias, demanding transparency, creating spaces where human judgment is not subordinated to machinic efficiency.

The discipline imposed by AI is not inevitable. It is a choice—made by designers, investors, policymakers, and ultimately by societies willing or unwilling to accept algorithmic governance as the price of convenience. Aldous Huxley (1932) warned that the most effective tyrannies are those we welcome, those that arrive dressed as comfort and efficiency. We are not yet in Huxley’s world. But we are closer than most recognize.

The question is not whether AI can discipline society. It already does. The question is whether we will remain its subjects—or become its authors.

References

Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. St. Martin’s Press. https://virginia-eubanks.com/automating-inequality/

Foucault, M. (1977). Discipline and punish: The birth of the prison (A. Sheridan, Trans.). Vintage Books. (Original work published 1975) https://www.penguinrandomhouse.com/books/55026/discipline-and-punish-by-michel-foucault-and-alan-sheridan/

Huxley, A. (1932). Brave new world. Chatto & Windus. https://www.huxley.net/bnw/

Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. NYU Press. https://nyupress.org/9781479837243/algorithms-of-oppression/

O’Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. Crown. https://www.penguinrandomhouse.com/books/241363/weapons-of-math-destruction-by-cathy-oneil/

Pasquale, F. (2015). The black box society: The secret algorithms that control money and information. Harvard University Press. https://www.hup.harvard.edu/books/9780674970847

Zuboff, S. (2019). The age of surveillance capitalism: The fight for a human future at the new frontier of power. PublicAffairs. https://www.hbs.edu/faculty/Pages/item.aspx?num=56791

Cite this article

Gattupalli, S. (2026). Is AI Disciplining Society? Society and AI. https://societyandai.org/perspectives/is-ai-disciplining-society/

Write for Society & AI

Society and AI welcomes scholarly contributions examining the intersection of artificial intelligence, education, governance, and society. We publish thoughtful, accessible work that advances public understanding of how algorithmic systems reshape institutions and human experience. All content is freely available under open access principles. Scholars, educators, policymakers, and practitioners are invited to submit proposals to the Editorial and Content Director at sai@societyandai.org.

On January 20, 2026, the World Economic Forum convened what I believe will be remembered as a watershed dialogue. Featuring two of the principal architects of modern artificial intelligence—Demis Hassabis, CEO of Google DeepMind and 2024 Nobel Laureate in Chemistry, and Dario Amodei, CEO of Anthropic—the session offered a rare and unvarnished glimpse into the immediate future of human labor and capability. These are not merely industry commentators speculating from the sidelines; they are the engineers actively defining the trajectory of computational intelligence.

The implications of their discourse extend far beyond the technology sector, touching the core of every institution dedicated to human capital development. As I reflected on their exchange, it became clear that the themes of this debate parallel the core focus areas of the Society and AI research group, situating these developments as foundational to our ongoing scholarly inquiry. This discussion provided a critical data point—a direct signal from the frontier of development—and the prognosis is sobering.

The Immediate Disruption of the Workforce

I was particularly struck by the starkness of their commentary on labor market disruption, which was notably devoid of the usual corporate optimism.

Amodei projected that up to fifty percent of entry-level white-collar roles could be obsolete within a one-to-five-year horizon. This is not a distant speculation; it is an imminent restructuring of the economy. When pressed on this point, Hassabis did not dissent. Rather, he corroborated the observation, noting a tangible deceleration in hiring for junior roles and internships at DeepMind itself.

This leads me to question the future of our professional developmental structures. Entry-level positions function as the apprenticeship layer of the professional world; they are the crucibles wherein soft skills—collaboration, ambiguity management, and professional judgment—are forged. If this stratum of employment evaporates, do we not risk severing the developmental pipeline that transforms novices into experts?

Amodei’s candor was compelling. He acknowledged that even within leading AI firms, the necessity for junior and mid-level human capital is diminishing. While he mentioned the need to navigate this transition “in a sensible way,” I find the definition of “sensible” remains elusive when extrapolated across the global economy.

The Phenomenon of “Capability Overhang”

One concept introduced by Hassabis that resonated deeply with me was that of “capability overhang.” This term describes a latent disparity: current AI systems possess capabilities that far exceed what has been deployed or even fully understood by their creators.

It appears the velocity of innovation has outpaced the capacity for exploration. The architects of these systems are so focused on the next iteration that the full utility of existing models remains unmapped. This presents a formidable challenge for societal adaptation. If the developers themselves cannot fully catalogue the capabilities of their tools, how can we expect educational institutions to design curricula relevant to the immediate future? How can corporations effectively upskill their workforce?

Hassabis offered a prescription for the current cohort of undergraduates: achieve exceptional proficiency with these tools. He posited that deep fluency in AI utilization could allow early-career professionals to “leapfrog” traditional experience hierarchies. This suggests to me a fundamental restructuring of career preparation is necessary, one that prioritizes adaptive technological fluency over static domain knowledge.

Beyond Economics: The Crisis of Purpose

While the economic implications are urgent, Hassabis steered the conversation toward a more philosophical dimension—the “bigger questions” of meaning, purpose, and the human condition in a post-labor economy.

This philosophical inquiry sits at the heart of our work at Society and AI. Historically, the social contract of education has been transactional: the acquisition of skills in exchange for economic utility. We must now confront a reality where that linkage is fraying. If the traditional pathways connecting education to employment are disrupted, must not the function of education evolve?

I argue that the imperative for education is shifting from purely vocational preparation to the cultivation of human flourishing. We must prepare individuals to construct lives of meaning and purpose in an era where their economic output may no longer be their primary defining characteristic. This is a challenge of existential magnitude, far surpassing the complexity of curriculum reform.

The Velocity of Transition

Both Hassabis and Amodei concurred that the timeline for this transition is contracting. Hassabis characterized the next five to ten years as a critical window, while Amodei suggested the timeline could be even more compressed. Their palpable lack of confidence in the preparedness of governmental bodies is a sentiment I share.

This skepticism is warranted. Contemporary discourse on AI in education remains largely fixated on academic integrity and policy compliance—necessary but insufficient discussions for the magnitude of the approaching shift. We are debating the rules of the classroom while the foundations of the labor market are being re-engineered.

We stand in the nascent stages of this technological epoch, yet the precursors of systemic change are already visible. The pioneers of this technology have issued a clear signal: the future is arriving with a velocity that defies our traditional planning cycles. Our collective response—whether we can pivot our institutions and mental models in time—will determine whether this transition enhances human potential or exacerbates societal stratification.

I leave my readers with two closing thoughts, framed as critical inquiries. First, what does this paradigm shift mean for the very definition of education, and consequently, what should we be teaching in our colleges and schools? And second, what does this era of accelerated capability signify for the fundamental transformation of our society?

References

World Economic Forum. (2026). World Economic Forum Annual Meeting. https://www.weforum.org/meetings/world-economic-forum-annual-meeting-2026/

World Economic Forum. (2026, January 20). The Day After AGI [Video]. YouTube. https://www.youtube.com/watch?v=mmKAnHz36v0

Cite this article

Gattupalli, S. (2026). The Vanishing Ladder: Education in a Post-Labor Society. Society and AI. https://societyandai.org/commentary/vanishing-ladder-education-post-labor-society/

I didn’t begin engaging AI from a place of novelty or experimentation. I began from a place of load.

A bilingual classroom does not offer long stretches of uninterrupted time, clean inputs, or ideal conditions. Instruction happens between languages, between emotional needs, between responsibilities that rarely pause. Some days, planning occurs in ten-minute fragments. Other days, it happens after dismissal, when the room is quiet and the work finally has space to surface. Adult ESL brings a parallel reality: learners balancing jobs, families, and immigration stress while trying to make meaning in a new language. In both settings, the question is rarely what is possible—it is what can flow.

My earliest use of AI was practical. I was looking for support that could help ideas move more smoothly: clarifying language, organizing thoughts, reducing friction in planning and communication. It was less about innovation and more about balance—finding ways to think clearly under pressure so the work felt sustainable rather than exhausting.

Over time, something unexpected happened. As I worked with the system, I began correcting it—not just for accuracy, but for tone, rhythm, and intent. I pushed back when language flattened, when responses drifted from classroom reality, when outputs sounded polished but hollow. Without naming it, I was building feedback loops: correcting misalignment, preserving voice, and enforcing continuity across interactions. What started as a tool for efficiency slowly became a space for reflection, testing, and refinement.

This shift mattered because classroom decisions carry consequence. Tone matters. Language choices matter. A poorly framed explanation can cost access; a misjudged response can cost trust. Any system introduced into this space has to operate within those constraints. For me, AI became useful only when it could work under the same conditions I do every day—not as a replacement for judgment, but as a partner that could keep pace with it.

From Tool to Thinking Partner

At first, AI showed up in my work the way most tools do: as support. It helped me organize ideas, draft language, and move through tasks more efficiently. In moments when time was tight, it reduced friction. In moments when energy was low, it helped me keep momentum. That alone made it useful.

But usefulness came with tension. The more I relied on it, the more I noticed subtle losses. Tone would soften when it shouldn’t. Language would become technically correct but emotionally off. Responses sounded polished, yet slightly detached from the reality of my classroom. The problem wasn’t accuracy—it was fit.

That’s when my role began to shift. Instead of accepting outputs as-is, I started pushing back. I revised phrasing. I rejected suggestions. I corrected the tone. I interrupted responses that didn’t sound like me or didn’t reflect the constraints I work under. What I needed wasn’t a system that generated answers quickly; it was one that could stay aligned with how I think, speak, and decide.

Over time, the interaction changed. AI stopped functioning as something I used and became something I worked with. Not because it gained agency, but because I imposed structure. I enforced boundaries. I treated misalignment as something to be corrected, not tolerated. The partnership emerged not from sophistication on the system’s side, but from discipline on mine.

This distinction matters. A tool optimizes tasks. A thinking partner supports cognition without displacing it. For me, the shift happened when I realized that clarity, voice, and judgment had to remain human-held—and that AI could only be helpful if it learned to operate within those limits.

Maintaining Professional Judgment

What mattered most as this partnership took shape was not speed or convenience, but judgment. Teaching requires constant decision-making under uncertainty: how much to say, when to pause, which language to use, and when silence does more than explanation. These decisions can’t be automated without cost.

As I worked with AI more consistently, I became deliberate about where judgment lived. Suggestions could be explored, language could be tested, but final decisions stayed with me. When outputs sounded confident but missed context, I stopped them. When explanations were technically sound but culturally flat, I rewrote them. The system could assist, but it could not decide.

This became especially important in moments that carried emotional weight—parent communication, student feedback, or instructional choices that affected trust. In those cases, clarity wasn’t enough. The words had to land correctly. Professional judgment meant knowing when to accept help and when to slow down, revise, or discard what was offered.

Over time, I learned that responsible use of AI is less about what the system can generate and more about what the educator is willing to reject. Judgment shows up not in the outputs we keep, but in the ones we don’t send. By holding that line, I was able to use AI without surrendering authorship, accountability, or care.

Drift, Correction, and Continuity

As my use of AI became more embedded in daily practice, a consistent pattern emerged: usefulness depended less on what the system could generate and more on how quickly misalignment was corrected. What I had initially called “drift” showed up in subtle ways—outputs that were technically accurate but contextually off, language that sounded professional but not instructional, or responses that lost the rhythm and specificity my work requires.

Technically, this is best understood as model output misalignment: moments when responses lose contextual fidelity to the user’s intent, constraints, or voice. In practice, it feels simpler. Something sounds wrong. Not incorrect, but not right. And in a classroom context, that distinction matters.

Correction became an active process. I interrupted outputs that didn’t fit. I renamed roles when needed—resetting the system away from “reporting” or “administrative” tone and back toward instructional clarity. I rejected language that over-smoothed complexity or added encouragement where precision was required. These corrections were not cosmetic edits; they were boundary-setting moves that preserved continuity over time.

Continuity is what made this usable. Each interaction did not stand alone. Corrections accumulated. Expectations stabilized. Over time, the system became less about generating content and more about holding a consistent posture across contexts—lesson planning, assessment interpretation, parent communication, and reflection. This did not happen because the model changed autonomously, but because I enforced alignment repeatedly and deliberately.

For educators, this distinction is critical. AI does not arrive aligned. Alignment is maintained through correction. Continuity is not a feature of the tool; it is the result of human discipline. When that discipline is present, AI can support thinking without fragmenting it. When it is absent, even high-quality outputs can erode coherence.

Bilingual, Elementary, Adult Education, and Tool Boundaries

The posture I’ve described does not apply uniformly across all learning environments. It is shaped by age, purpose, and developmental readiness. What works with adult learners does not automatically translate to elementary classrooms, and responsible use depends as much on what is excluded as on what is introduced.

With adult ESL learners, AI functions primarily as a language bridge. Students use it to clarify meaning, rehearse phrasing, and gain confidence expressing ideas they already hold. The goal is not dependency, but access. Language support works best when it preserves dignity—when students feel assisted rather than corrected. My role is to ensure that AI-supported language reflects respect, cultural awareness, and realistic growth expectations, rather than generic encouragement or deficit framing.

Elementary instruction requires a different boundary. In my second grade classroom, AI is not a student-facing tool. At this developmental stage, learning is grounded in concrete experiences, direct instruction, and age-appropriate platforms. Technology use centers on tools like Code.org for foundational computational thinking and carefully selected, child-friendly instructional videos. These resources support sequencing, logic, and curiosity without introducing systems that students are not yet ready to interpret or question critically.

My STEM Club (third and fourth grade) sits between early elementary and upper elementary. Here, students are still building fundamentals—how to follow sequences, debug, collaborate, and persist through challenges. The tools remain developmentally appropriate and hands-on, with emphasis on thinking processes rather than advanced systems.

With fifth and sixth grade after-school students, the conversation can begin to shift toward how digital tools support thinking without replacing it. At that level, modeling matters: showing that technology can help organize ideas or test understanding, while responsibility and authorship stay human.

Across all contexts, bilingualism adds another layer. Language choice is never neutral. Spanish, Haitian Creole, and English carry different emotional weights and cultural meanings. Whether working with AI-supported reflections in adult education or translated explanations crafted by the teacher, alignment requires attention not just to accuracy, but to how language lands with the learner.

What connects these environments is intentional restraint. Responsible integration is not about introducing the most advanced tool at every level, but about selecting tools that fit the learner’s cognitive, emotional, and cultural readiness. In that sense, thoughtful exclusion is as important as thoughtful use.

Emotional Labor and Regulation

Teaching involves constant emotional regulation that rarely appears in lesson plans or data reports. Decisions are made while managing student emotions, family concerns, administrative expectations, and personal fatigue. Much of this work happens quietly, between moments, without acknowledgment.

In this space, AI became useful in a different way. Not as a source of answers, but as a place to pause. I used it to test language before sending it, to process frustration without passing it on, and to slow down responses that could easily have been reactive. The value wasn’t in what the system produced, but in the space it created for reflection.

This boundary mattered. AI did not replace emotional judgment or decision-making. It absorbed cognitive load so that judgment could remain intact. Drafts were revised, not sent. Responses were shaped, not outsourced. In moments that carried emotional weight, restraint mattered more than speed.

What made this ethical was clarity about roles. AI functioned as a buffer, not an authority. It helped surface options, but responsibility stayed human. By using it this way, I was able to protect relationships rather than risk them, and to respond with intention rather than exhaustion.

This form of use is easy to overlook because it leaves little visible trace. But for educators, it may be one of the most consequential applications. When emotional labor is supported rather than suppressed, teachers are better able to show up present, measured, and grounded—not because decisions are automated, but because the work of thinking has space to breathe.

Closing Reflections

This reflection is not an argument for broader adoption or a blueprint to be replicated wholesale. It is an account of what happens when an educator insists on remaining centered while working alongside a powerful system. The outcomes were shaped less by the technology itself and more by the discipline applied to its use.

What I learned is simple, but not easy: AI does not preserve judgment on its own. It reflects the posture of the person using it. When boundaries are clear, corrections are enforced, and voice is protected, AI can support thinking without fragmenting it. When those conditions are absent, even well-intended use can dilute clarity and erode coherence.

For educators, the question is not whether AI belongs in classrooms, but how it is held. Responsible use begins with restraint, with attention to context, and with a willingness to reject outputs that do not fit the realities of students, language, and care. It requires remembering that teaching is not a problem to be optimized, but a relationship to be maintained.

In my practice, AI became useful only when it could operate under the same constraints I do every day. Not as a replacement for judgment, but as a support for it. That distinction—quiet, deliberate, and human—has made all the difference.

About the Author

Alex Luciano is a bilingual elementary educator and adult ESL instructor in the Central Islip School District on Long Island, New York. He currently teaches bilingual second grade and works with multilingual students and adult learners in high-need public school contexts. With over 20 years in education, his work focuses on language, culture, and maintaining human judgment while working alongside emerging technologies. His perspective is shaped by sustained classroom practice and a commitment to keeping teaching human, relational, and grounded.

Correspondence: aluciano@centralislip.k12.ny.us

Cite this article

Luciano, A. (2026). Holding the Line: Teaching Alongside AI Without Losing Voice. Society and AI. https://societyandai.org/insights/holding-the-line-teaching-alongside-ai/

Write for Society & AI

We welcome contributions from educators, researchers, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. If you have a perspective to share, send your proposal to the Editorial and Content Director at sai@societyandai.org.

But here is what the gleaming displays and polished presentations obscure: the vast majority of these devices are not self-contained. They are portals. The intelligence does not reside in the refrigerator or the headphone or the mirror. It resides in data centers scattered across continents, humming with servers that consume electricity measured in terawatt-hours and water measured in billions of gallons. Every question we ask our smart speakers, every image we generate, every document we summarize travels through fiber optic cables to these facilities, where massive language models process our requests and return responses. The magic happens elsewhere—out of sight and increasingly, out of mind. We interact with the interface, never seeing the infrastructure: the cooling towers, backup generators, and reservoirs drained to keep silicon from overheating.

The Disclaimer You See

We use AI tools in our own work and research every day. If you use ChatGPT, Claude, Gemini, or any of the major AI chatbots now embedded in daily life, you have likely noticed a small line of text beneath the input field. In ChatGPT, it reads:

“ChatGPT can make mistakes. Check important info.”

ChatGPT’s accuracy disclaimer appears beneath every query.

Claude displays a similar warning about potential errors.

This acknowledgment represents a form of corporate humility. Companies are admitting, in plain language, that the outputs may not be accurate. They ask users to exercise judgment, verify claims, remain skeptical. This is commendable. Accuracy matters. Misinformation spreads harm.

But we think accuracy is not the only thing that matters.

The Disclaimer You Do Not See

What if, directly beneath that accuracy warning, another line appeared? Something like:

“This response consumed approximately 0.3 watt-hours of electricity and contributed to water usage in data center cooling systems. Learn more about AI’s environmental impact.”

You will not find such a statement on any major AI platform today. And yet the environmental costs are neither speculative nor trivial.

OpenAI CEO Sam Altman has publicly disclosed that the average ChatGPT query consumes approximately 0.34 watt-hours of electricity—roughly what a high-efficiency lightbulb uses in a couple of minutes—and about 0.000085 gallons of water, roughly one-fifteenth of a teaspoon (Altman, 2025). While these figures appear modest in isolation, scale transforms individual efficiency into collective consequence.

While the current state of the art has advanced to GPT-5, Li et al. (2023) reported that GPT-3 consumed approximately 500 milliliters of water per 10–50 responses, with usage varying by deployment location and cooling efficiency; and as CES showcases an ever-expanding ecosystem of AI-enabled devices, this infrastructural dependence will only compound and intensify these resource demands. As AI infrastructure increasingly underpins everyday appliances and services, these seemingly negligible per-query costs scale into a staggering cumulative impact.

According to Food & Water Watch, by 2028 AI-related data centers in the United States could require as much as 720 billion gallons of water annually just for cooling—enough to meet the indoor water needs of 18.5 million households.

The carbon footprint is similarly alarming. de Vries-Gao (2025) estimates that AI systems alone could be responsible for between 32.6 and 79.7 million metric tons of CO₂ emissions in 2025—comparable to the annual emissions of a major city like New York.

The Asymmetry of Transparency

The contrast is stark. AI companies have decided that users need to know about potential inaccuracies. They embedded disclaimers directly into interfaces—small but visible reminders that the technology is fallible. This represents a choice: a decision that certain information is important enough to surface.

Why, then, does environmental impact not meet the same threshold?

If companies believe users deserve to know when AI might produce an error in a letter or calculation, surely users also deserve to know when their interaction contributes to resource depletion and carbon emissions. The logic is identical: transparency enables informed decision-making. Silence on environmental costs suggests that accuracy is a consumer concern worth addressing, while ecological impact is not.

Toward Meaningful Disclosure

We are not arguing that AI should be abandoned. We use these tools ourselves in daily work. They offer genuine value in education, research, accessibility, and countless other domains. But value does not exempt technology from accountability.

What we propose is deceptively simple: environmental transparency at the point of interaction. Every chat interface should display—alongside accuracy disclaimers—an estimate of the energy consumed, water used, and carbon emitted per session or query. This could appear as a per-session summary, aggregate dashboard, or real-time indicator. The data already exists within company infrastructure.

Such disclosure would not deter legitimate use. But it would invite reflection. It would encourage efficiency—prompting users to consolidate queries rather than treating AI as an infinite resource. And it would create market pressure for companies to reduce environmental impact, knowing users can now see the cost.

In an era of climate crisis, we cannot afford to treat AI’s footprint as an externality. The companies building these systems have a responsibility to make the invisible visible.

The next time you see “ChatGPT can make mistakes,” ask yourself:

What else are they not telling you?

References

Altman, S. (2025, June 10). The gentle singularity. Sam Altman’s Blog. https://blog.samaltman.com/the-gentle-singularity

de Vries-Gao, A. (2025). The carbon and water footprints of data centers and what this could mean for artificial intelligence. Patterns. https://doi.org/10.1016/j.patter.2025.101430

Food & Water Watch. (2025). Artificial intelligence: Big Tech’s big threat to our water and climate. https://www.foodandwaterwatch.org/2025/04/09/artificial-intelligence-water-climate/

Li, P., Yang, J., Islam, M. A., & Ren, S. (2023). Making AI less “thirsty”: Uncovering and addressing the secret water footprint of AI models. arXiv preprint arXiv:2304.03271. https://arxiv.org/abs/2304.03271

Cite this paper: Gattupalli, S., & Chakravarty, P. (2026). The Invisible Cost of Every Chat. Society and AI. https://societyandai.org/perspectives/invisible-cost-of-every-chat/

When I first encountered Paulo Freire’s Pedagogy of the Oppressed (Freire, 1970) as a doctoral student in 2019, I remember being struck by his description of the banking model of education, where teachers deposit information into students as though they were empty vessels waiting to be filled. Freire argued that this model dehumanizes learners by stripping them of agency, turning education into an act of domestication rather than liberation. The oppressed, he observed, often internalize the consciousness of the oppressor, accepting their own subjugation as natural and inevitable.

I find myself returning to Freire now, decades later, because I notice unsettling resemblances between the dynamics he described and something emerging in our relationship with artificial intelligence. The structures have changed, the technologies are unrecognizable, but the underlying pattern of extraction and internalized dependence feels hauntingly familiar. What Freire identified in the classroom between teacher and student, I now see playing out between AI systems and their users, between cloud platforms and the billions of people whose cognitive labor feeds them.

The Birth of Cloud Capital

Economist Yanis Varoufakis (2024) offers a framework that helps clarify what many of us have sensed but struggled to articulate. In his book Technofeudalism: What Killed Capitalism, he introduces the concept of cloud capital, which he describes as a produced means not of production in the traditional sense, but of behavioral modification. Unlike the factory machinery of industrial capitalism that manufactured physical goods, cloud capital manufactures something else entirely. It manufactures desire. It shapes preference. It extracts value from the very act of human cognition.

Consider what happens when I speak to a voice assistant, scroll through a recommendation feed, or ask an AI tutor for help with a problem. I am not merely using a tool in any conventional sense. I am training an algorithm that learns to know me, and as Varoufakis explains, I am simultaneously being trained by it to help it know me better. The process is circular and self-reinforcing. The more I interact, the more the system learns. The more it learns, the better it predicts. The better it predicts, the more effectively it can modify my behavior by recommending products I did not know I wanted, surfacing content designed to keep me engaged, and gradually shaping the very contours of my attention and desire.

Freire would recognize this pattern immediately. In the banking model, the teacher knows and the student receives. In the cloud capital model, the algorithm knows and the user provides. Both relationships are presented as benevolent, even empowering. Both, upon closer examination, reveal asymmetries that systematically benefit one party at the expense of the other’s autonomy.

Cognitive Labor Without Wages

What troubles me most as I reflect on this arrangement is recognizing that what I do when I interact with AI systems constitutes labor. When I upload photographs, compose messages, correct autocomplete suggestions, or engage with AI-generated content, I am performing cognitive work that adds directly to the cloud capital of platform owners. Yet unlike, say a factory worker who at least receives a wage in exchange for labor, I receive only the illusion of free services. The transaction is obscured by convenience.

The asymmetry becomes stark when you examine the economics. Varoufakis points out that in traditional capitalist enterprises, roughly 80% of revenue historically flowed to wages across the organization, from janitors to executives. In companies built on cloud capital, that figure can drop to as low as 1%. The remainder is extracted, routed through elaborate international tax structures, and accumulated by a vanishingly small ownership class. This is not a “bug” in the system or an unfortunate side effect. This is the system functioning exactly as designed.

When educators celebrate AI tools that personalize learning or adapt to student needs, we rarely pause to ask who benefits from the data students generate in those interactions. Every engagement a child has with an AI tutor, every mistake and hesitation and breakthrough, becomes training data that increases the value of cloud capital owned by corporations headquartered thousands of miles away. The student receives feedback. The corporation receives capital. Freire’s banking model has found a new institutional form, one that deposits not just curriculum but consciousness itself into systems that students do not own and cannot govern.

From Markets to Fiefdoms

Varoufakis argues that what we are witnessing is not merely a new phase of capitalism but its replacement by something older and in some ways more troubling: feudalism, digitized and scaled for the twenty-first century. Traditional markets, which at least in theory functioned as spaces where buyers and sellers meet as nominal equals, are being supplanted by what he describes as cloud fiefdoms. Amazon, in this analysis, is not really a marketplace at all. It is a digital estate where producers and consumers operate within walls controlled by a single lord who extracts rent from every transaction that occurs within his domain.

The implications for education are profound and deserve far more attention than they currently receive. As AI systems increasingly mediate learning, we must ask ourselves some uncomfortable questions. Are we building educational marketplaces where learners can freely choose among options and maintain agency over their own development? Or are we constructing digital fiefdoms where a handful of platform owners determine what knowledge is surfaced, what skills are valued, what questions are worth asking, and ultimately whose interests are served?

The consolidation of AI capabilities in a small number of corporations, each accumulating cloud capital through billions of daily interactions, represents a concentration of cognitive infrastructure that has no real precedent in human history. We have seen concentrations of economic power before. We have seen concentrations of political power. But a concentration of the very infrastructure through which cognition itself is increasingly mediated raises questions that our existing frameworks struggle to address.

The Cognitive Dimension

There is another extraction happening alongside the economic one, and it receives far less attention in public discourse. When I outsource cognitive tasks to AI, when I let the machine remember things I used to remember, calculate things I used to calculate, compose text I used to compose, and make decisions I used to make myself, I may be surrendering not just data but capacity.

Cognitive scientists have long understood that mental capabilities, like muscles, require exercise to maintain their strength. The research on desirable difficulties by Bjork and Bjork (2011) demonstrates that struggle is not an obstacle to learning but rather its essential mechanism. When AI removes friction by answering questions before I have fully formulated them, completing thoughts before I have thought them through, the convenience may come at a cost. It may also remove the conditions under which cognitive development actually occurs.

This creates a troubling possibility that I keep returning to in my own thinking: cloud capital grows through my cognitive labor while my cognitive capacity may simultaneously diminish through disuse. I train the machine, and the machine trains me to need it more. The asymmetry compounds over time.

Freire wrote about how the oppressed internalize the consciousness of the oppressor, coming to see themselves through the oppressor’s eyes and accepting their diminished status as natural. I wonder whether something analogous happens when we become dependent on AI systems. Do we begin to see our own cognition as inadequate, our own thinking as too slow, our own memory as unreliable? Do we internalize a view of ourselves as cognitive inferiors who naturally require algorithmic assistance? And if so, who benefits from that internalization?

Who Benefits?

Zooming out to see the full picture requires asking a question that technology discourse often obscures beneath layers of enthusiasm about innovation and disruption. That question is simply: who benefits from this arrangement?

The owners of cloud capital benefit in ways that are measurable and enormous. Their wealth compounds with every interaction, every data point, every refinement of algorithmic prediction. The numbers are public and staggering.

Users receive convenience, efficiency, and the seductive comfort of systems that anticipate their needs. These are real benefits that I experience myself daily. But convenience has costs that rarely appear on balance sheets, and efficiency must always prompt the question: efficient for whom, and toward what ends?

What of the broader society? Here the accounting becomes more complex. When somewhere between 20 and 30 percent of economic value, by Varoufakis’s estimation, is siphoned from the circular flow of income into the accounts of cloud capital owners, aggregate demand suffers. The quality of available jobs degrades as traditional employment is replaced by platform-mediated gig work. Economic inequality widens into a chasm. And the political power that inevitably accompanies such concentrated wealth begins to reshape governance itself in ways we are only beginning to understand.

For education specifically, the stakes feel existential to me. If AI systems continue to centralize cognitive infrastructure in the hands of a few corporations, we risk creating a world where the very capacity to think independently becomes something like a luxury good. It would be available to those who can afford to cultivate it deliberately, and atrophied in those who cannot escape the frictionless embrace of algorithmic mediation. This is not a distant dystopia. It is a trajectory we are already on.

The Question Before Us

I want to be clear that I do not write this as a someone who either opposes or is resistant to the use of new and emerging technologies. I use AI tools daily in my research and teaching. I see their genuine potential to democratize access to knowledge, to personalize learning in ways that honor individual difference, and to augment human capability rather than simply replace it. The technology itself is not the problem.

But potential is not destiny, and the same technologies that could liberate can also extract. The same systems that could empower learners can also enclose them within digital fiefdoms. The question before us is not whether AI will transform education, because it already is transforming education whether we attend to it or not. The question is whether that transformation will serve the flourishing of all learners or the accumulation of capital by the few.

Varoufakis makes a point that I think is essential for educators to understand. This is not fundamentally a personal matter, not a question of individual choices about which apps to use or which platforms to avoid. Even if you do not have a smartphone, even if you try to disconnect from the digital infrastructure entirely, you cannot really escape because you are living in an environment in which the quality of jobs has been depleted, in which power has become immensely concentrated in the hands of the very few. The effects are structural and systemic, not individual and personal.

The cognitive gym I will propose in subsequent essays, the deliberate cultivation of mental capacity in an age of outsourcing, is necessary but insufficient on its own. Individual practice cannot substitute for collective action. Personal cognitive fitness cannot compensate for structural extraction. We need both.

What we need is not merely to exercise our minds but to reclaim the infrastructure through which cognition itself is increasingly mediated. That is a political project as much as a pedagogical one. It requires the kind of critical consciousness that Freire called conscientização, an awakening to the structures of oppression and an understanding that those structures can be changed through collective action.

The cloud is not neutral territory any more than the classroom was neutral in Freire’s analysis. It is owned. It is governed. And what is owned and governed can be owned and governed differently, if we have the clarity to see what is happening and the collective will to demand alternatives.

References

Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher & J. Pomerantz (Eds.), Psychology and the real world: Essays illustrating fundamental contributions to society (pp. 56-64). Worth Publishers.

Ferenczy, B. (2024). Cloud generator [Interactive visualization]. CodePen. https://codepen.io/BalintFerenczy/pen/qENdpoL

Freire, P. (1970). Pedagogy of the oppressed. Continuum.

Varoufakis, Y. (2024). Technofeudalism: What killed capitalism. Melville House.

Cite this article: Gattupalli, S. (2026). Cloud Capital and the Extraction of Mind. Society and AI. https://societyandai.org/perspectives/cloud-capital-extraction-of-mind/

Every civilization that has survived existential threat did so through one mechanism: learning. Not technology, which is inert without understanding. Not resources, which are useless without the knowledge to deploy them. Not institutions, which collapse when the people within them cannot adapt. Learning is the process through which human societies recognize change, generate responses, and transmit solutions across generations. It is, in the strictest sense, humanity’s only scalable adaptation technology.

I see this not as metaphor, but as mechanism.

Historical Context: The Learning Society Concept

The notion of a “learning society” is not new. Since the mid-1980s, policy makers, economists, and scholars have argued that investment in education and training constitutes the primary route to organizational and national advantage (Morris, 1999). This approach drew inspiration from economists dating back to Adam Smith and Alfred Marshall, as well as the human capital theory developed by Gary Becker in the 1960s. More recently, these ideas have been rebranded under various labels, including the “learning society” and “knowledge economy.”

Yet as Morris (1999) observed in his comprehensive review of the field, there remained “far too little comparable data available for robust conclusions to be drawn about the incidence and effects” of education and training investments across different national contexts. He called for “the development of more comprehensive and comparable surveys” to move the debate forward.

A quarter century later, this call remains largely unanswered, even as the stakes have multiplied exponentially with the emergence of artificial intelligence.

What was once a question of national economic competitiveness has become a question of civilizational adaptation. The learning society thesis, articulated in an era of globalization and human capital theory, now confronts a technological discontinuity that its original proponents could not have imagined: systems that learn without humans, that generate knowledge without understanding, that threaten to substitute for the very educational processes they were meant to enhance.

The Biological Lens: From Genes to Culture

Biological evolution adapts species to environments through genetic variation and natural selection. The process is powerful but slow, operating across generations, blind to intention, incapable of foresight. For most of life’s history, this was sufficient. Environments changed gradually; species either adapted or perished.

Humans broke this constraint. Somewhere in our evolutionary history, we developed a second inheritance system: culture. Unlike genes, cultural information can be acquired within a lifetime, modified deliberately, and transmitted horizontally across populations. A human born today inherits not merely the genes of ancestors but the accumulated knowledge of thousands of generations: language, mathematics, medicine, governance, the hard-won lessons of countless trials and catastrophes.

This is the first great insight: culture is a faster adaptation system than biology. When environments change rapidly, as they did during ice ages, as they do during pandemics, as they will during climate collapse, species constrained to genetic adaptation cannot respond in time. Humans respond through culture. We learn, we teach, we transmit, and we adapt at the speed of communication rather than the speed of reproduction.

But culture does not transmit itself. It requires infrastructure.

The Cultural Lens: Why Learning Requires Institutions

Cultural transmission happens through demonstration, through story, through apprenticeship, but none of these mechanisms scale. A master craftsperson can train a handful of apprentices. A storyteller can reach a village. Even writing, that great leap, requires readers who have been taught to read.

Education is the institution humanity invented to solve this scaling problem. Schools, curricula, pedagogies: these are not merely social conventions. They are the infrastructure through which culture becomes reliably transmissible across populations and generations. Without deliberate educational systems, knowledge fragments, erodes, disappears. With them, knowledge compounds.

Consider what happens when educational infrastructure fails. The Bronze Age Collapse erased writing systems across the Mediterranean for centuries. The fall of Rome fragmented technical knowledge that took a millennium to recover. The deliberate destruction of indigenous educational systems during colonization severed cultural continuity in ways that communities still struggle to repair. These are not merely historical tragedies. They are data points. They demonstrate that societies which lose the capacity to educate lose the capacity to adapt.

And here is the compounding insight: education does not merely transmit existing knowledge. It generates new knowledge. The capacity to learn how to learn, to develop methods for inquiry, criticism, and revision, is what separates societies that stagnate from societies that progress. David Deutsch captures this precisely: all progress depends on the creation of better explanations through conjecture and criticism. Education is the institution that cultivates this capacity at scale.

The Society & AI Lens: Amplifier or Threat?

Artificial intelligence represents a discontinuity in human history. For the first time, we have created systems that can process information, generate outputs, and increasingly, act in the world without direct human intervention. The question is not whether AI will transform society; it already is. The question is what that transformation means for humanity’s adaptation mechanism.

Two futures are possible. The design choices we make today will determine which one we inhabit.

In the first, AI amplifies education. It personalizes instruction, extends access, frees teachers from administrative burden, and accelerates the creation and dissemination of knowledge. In this future, AI is a tool that serves the learning society, making education more effective, more equitable, more capable of preparing humans to navigate whatever challenges emerge.

In the second future, AI substitutes for education. It generates answers without cultivating understanding. It provides outputs without developing the capacity to evaluate them. It creates dependence rather than capability. In this future, humans become consumers of AI-generated solutions rather than producers of knowledge. And when those AI systems fail, as all systems eventually fail, humans will have lost the very capacity that allowed our species to survive ice ages and plagues and civilizational collapse.

This is not a matter of technology policy. It is a matter of species survival.

The learning society thesis clarifies the stakes. If education is humanity’s adaptation mechanism, and if AI is being integrated into educational systems worldwide, then the design choices we make today will determine whether future generations can adapt to challenges we cannot yet imagine. We are not simply deciding how to teach mathematics or reading or science. We are deciding whether the infrastructure of human adaptation will be strengthened or eroded.

The Design Imperative

The learning society thesis is not merely analytical. It is prescriptive. It demands that we evaluate every AI intervention in education against a single criterion:

Does this cultivate human capacity for learning, or does it substitute for it?

AI that helps learners struggle productively with difficult problems cultivates capacity. AI that provides answers without struggle substitutes for it. AI that extends access to quality instruction for the billions currently excluded cultivates capacity at civilizational scale. AI that concentrates educational advantage among the already privileged substitutes global adaptation capacity for local efficiency gains. AI that supports teachers in understanding their students cultivates the relational core of education. AI that replaces teachers with interfaces substitutes technological optimization for human development.

The distinction is not always obvious. A system can appear helpful while subtly undermining the learning it claims to support. A system can feel demanding while genuinely building capability. The learning society thesis provides the evaluative lens: we must always ask whether humans are becoming more capable of adapting, or less.

We are a learning society; the question is whether we will remain one.

This working paper examines whether education is undergoing a fundamental phase transition driven by artificial intelligence. Drawing on the evolutionary concept of punctuated equilibrium, the paper argues that we are witnessing not merely another cycle of educational technology adoption, but a reorganization of how human and artificial minds learn together. The analysis explores the historical parallels between AI and previous transformative technologies like the blackboard and personal computer, while examining the implications for teachers, students, institutions, and society at large. The paper introduces the concept of “AI fluency” as an essential civic competency and concludes with design principles for education systems that can thrive in this new equilibrium.

My journey into this question began unexpectedly. As a physics student, I was captivated by how we come to know what we know. That curiosity led me to education research, where I found myself watching AI arrive in classrooms faster than any of us could make sense of it. This paper grew from late nights thinking about my students, my colleagues, and a persistent feeling that what we are experiencing is not just another technology trend. It feels like the ground shifting beneath our feet.

I keep returning to a simple image: a teacher in 1820, chalk in hand, writing on a blackboard for the first time. In that moment, thought became visible to an entire room simultaneously. We barely remember this as revolutionary, yet it changed everything. The blackboard, the textbook, the computer—each arrived quietly and reorganized how humans think together. I believe AI is the next such moment, except this time, the change is not quiet at all. It is loud, fast, and disorienting. What troubles me most is how quickly we have normalized it without asking harder questions. I watch institutions scramble to update honor codes when they should be reimagining what learning itself could become.

If the blackboard made thinking visible and the computer made knowledge personal, AI can make wisdom relational. It forces us to confront what was always true: teaching is fundamentally about relationship, not transmission. The equilibrium we build now will determine not just how we learn, but who we become as thinking beings in an age of artificial intelligence. That choice, I believe, is still ours to make.

The Instant Answer Paradox

We once stood before knowledge and contemplated; now knowledge stands ready to answer before we’ve finished asking.

A question stays with me: If AI can answer questions instantly, what purpose do memorization, research skills, or expertise serve? The findings from Chien and colleagues offer something unexpected. Participants experienced “significantly high flow and low activity anxiety” while learning basic science through AI-scaffolded gameplay. Rather than disengaging when AI provided immediate feedback, students leaned in. This response deserves attention.

This observation challenges assumptions I’ve long held about learning. We’ve believed struggle builds capability. Perhaps we’ve been conflating two distinct forms of struggle. Productive struggle involves grappling with concepts, testing hypotheses, integrating knowledge. Logistical struggle involves waiting for feedback, searching for resources. AI eliminates the latter while potentially intensifying the former. I should note that the game itself focused on teaching AI prompting skills, which somewhat confounds these findings with the learning outcomes.

From Acquiring Knowledge to Building Capability

What struck me about this research was how the generative AI provided “a high percentage of direct answers or indirect suggestions,” creating suitable interactive scaffolding. The study used ChatPDF constrained to specific content, achieving 78-90% helpful response rates. This wasn’t simply automation replacing human instruction. It represented a responsive learning environment adapting to each student’s inquiry in real time.

Education has centered on knowledge acquisition for the past century. We’ve treated students as vessels for facts, theories, and procedures. Success meant retaining and retrieving information on demand. When AI can retrieve facts readily, this model reveals its limitations. Not because knowledge loses importance, but because the fundamental constraint shifts. Access to information no longer limits us. The capability to use information well becomes the bottleneck that matters.

From this research emerges insight into what I’ll call capabilities. Participants weren’t memorizing isolated science facts. They engaged in contextualized problem-solving within narrative frameworks. They formulated questions, evaluated AI responses, made decisions under uncertainty. These represent capabilities, not mere retention.

Three Capabilities That Matter More

Using this research as a starting point, I see three capability domains emerging as increasingly valuable.

Question formulation. When AI provided “direct answers or indirect suggestions,” learning quality hinged entirely on question quality. Knowing what to ask, recognizing when different questions serve better, this becomes foundational. It demands domain awareness, what I think of as curiosity architecture, and epistemic humility. These capabilities can be taught.

Synthesis across contexts. Players applied basic science knowledge to solve puzzles within a fictional cave system. This transfer, moving knowledge from one context into another, remains distinctly human territory. AI generates solutions within defined parameters. Identifying which parameters matter in novel situations requires judgment shaped by varied experience across domains.

Metacognitive awareness. The research measured flow and anxiety during gameplay. Students maintaining high flow weren’t passively absorbing information. They monitored their engagement, regulated learning strategies, maintained productive challenge. This self-awareness grows critical when AI offers scaffolding. Without it, dependency on external structure replaces internal capability.

Implications for Educational Practice

This research raises important questions for those shaping learning experiences. I’ve been considering what this means for practice:

• Reimagine assessment priorities. When AI handles recall effortlessly, assessments might focus on application in novel contexts, examining whether students can transfer understanding across domains.

• Develop question literacy. Question formulation could become an explicit skill with deliberate practice and progression, much like we approach writing or quantitative reasoning today.

• Distinguish productive from logistical struggle. AI might remove barriers to resources and feedback, while we simultaneously increase cognitive challenge through problems requiring genuine synthesis and creative integration.

• Build metacognitive reflection. Learning routines might incorporate regular reflection, asking not only what students learned but how they learned it, what strategies proved effective.

• Prioritize depth over breadth. When information becomes readily available, developing one capability deeply may serve students better than covering ten topics superficially.

What This Reveals About Learning

This study demonstrates that generative AI can scaffold learning while sustaining engagement within a specific context. I’ve used it here as a springboard for broader speculation about education’s future. The research itself examined a contextualized game teaching basic science to college students; my reflections extend beyond what the data alone can support.

When answers arrive readily, the most valuable capability isn’t knowing facts. It’s understanding what to do with accessible knowledge. Rather than diminishing learning, this shift represents its elevation to higher purpose.

References

Chien, C. C., Chan, H. Y., & Hou, H. T. (2025). Learning by playing with generative AI: Design and evaluation of a role-playing educational game with generative AI as scaffolding for instant feedback interaction. Journal of Research on Technology in Education, 57(4), 894-913. https://doi.org/10.1080/15391523.2024.2338085

The United Nations Development Programme’s latest report, The Next Great Divergence, maps out how artificial intelligence might permanently fracture the global economy. It’s dense and alarming, and it deserves careful attention from anyone thinking seriously about AI’s societal implications.

The Infrastructure Illusion

For a decade, global development strategy has centered on “access”: tablets in classrooms in Nairobi, fiber optics to villages in Vietnam. Connectivity was supposed to be the equalizer.

The UNDP report reveals why this was never enough. It identifies a “compute divide” that makes the old “digital divide” look almost quaint. We’re entering an era where educational quality becomes a function of processing power. A student in California learning from an AI tutor running on cutting-edge hardware inhabits a different cognitive reality than a student in Bangladesh consulting a static textbook over a 3G connection. They’re not just in different schools. They’re in different centuries.

This is what makes “Brain Drain 2.0” different from its predecessor. The original brain drain was about people leaving. This one is about the capacity to learn being centralized elsewhere. If the Global South cannot build sovereign AI infrastructure, its educational systems risk becoming permanent clients of Northern platforms, importing curricula optimized for Western contexts and renting them at perpetual cost.

From a society-centered AI perspective, this isn’t just an infrastructure problem. It’s a question of who gets to shape how the next generation thinks.

The Invisible Extraction

The report warns that “larger entities might exploit local data or content to train AI models” without benefiting source communities. This quiet extraction is the mechanism at the heart of the new divergence.

Consider what happens when we train a Large Language Model on the open web. The cultural, scientific, and creative output of the Global South gets ingested: a Kenyan novelist’s prose, a Brazilian botanist’s taxonomy, an Indian developer’s code. Each is stripped of context, tokenized, and absorbed into model weights, becoming indistinguishable from its origins.

We have no way to measure the “data GDP” of a nation. The Global South exports its collective intelligence for free, then purchases it back as a service. This creates a recursive loop: the model grows more capable while the source grows more dependent. It’s not just inequitable. It’s structurally unsustainable.

A Different Path

Here’s where society-centered thinking offers something the report doesn’t fully articulate: the Global South doesn’t have to play this game at all.

The North is building AI that is energy-intensive, centralized, and capital-heavy, optimized for corporate efficiency. But what if that model is itself the problem? Rather than replicating Silicon Valleys (which talent migration is making increasingly difficult anyway), the Global South could pursue something different entirely. Local, edge-based models running on low-power devices, trained on specific, high-quality local data. Intelligence that solves local problems without requiring data centers on distant continents.

The prevailing narrative assumes intelligence must be centralized to be valuable. But sovereign, decentralized AI isn’t “falling behind.” It’s stepping off a treadmill that was never designed for broad human benefit.

Toward Computational Sovereignty

The UNDP report’s warning is serious, but its framing is incomplete. The “Next Great Divergence” is inevitable only if the Global South accepts the terms of a contest whose rules systematically disadvantage them.

The alternative is sovereignty: not just legal sovereignty over borders, but computational sovereignty over minds. This means educational systems running on local infrastructure. Legal frameworks treating data extraction like resource extraction, worthy of taxation. Innovation strategies prioritizing local resilience over global efficiency.

This isn’t about building walls. It’s about building alternatives. The divergence the UNDP fears assumes everyone walks the same path. The real opportunity lies in constructing a different road, one where AI serves societies rather than extracting from them.

As an AI strategist focused on the societal and educational implications of artificial intelligence, I write this edition to examine how AI integration is reshaping educational systems worldwide, with particular attention to equity, access, and human flourishing.

TL;DR

November 2025 reveals a world divided not by whether to embrace AI in education, but rather by who gets to shape its terms. While China mandates nationwide integration and South Korea invests billions in lifetime AI literacy, the Global South faces a stark reality: 68% accessibility in the Northern Hemisphere versus a mere 23% in the Southern, with infrastructure gaps that risk transforming AI from a democratizing force into a stratification mechanism. Nations are moving from prohibition to policy, yet the shift disproportionately benefits those already positioned to lead, leaving critical questions of equity, teacher capacity, and technological sovereignty largely unanswered. The digital divide is not closing; instead, it is being systematically encoded into educational systems for generations to come. What we witness is not global progress, but rather regionalized advancement with profoundly unequal consequences for the world’s learners.

Here is the full report.

Asia-Pacific: National Mandates and Billion-Dollar Investments

China Makes History with Mandatory AI Curriculum

In an unprecedented move, China’s Ministry of Education issued guidance on strengthening AI education in primary and secondary schools, positioning China as the first major nation to mandate comprehensive K-12 AI education nationwide. Hangzhou in Zhejiang province announced mandatory AI courses for all students starting the new semester, explicitly linking AI education to national innovation goals. The guidance promotes project-based learning and calls for whole-of-government coordination with pilot programs.

South Korea’s $1.25 Billion Lifetime AI Strategy

On November 10, 2025, South Korea’s Education Ministry unveiled its first national blueprint titled “AI Talent Development Plan For All,” committing 1.4 trillion won (US$1.25 billion) to integrating AI education from elementary through postgraduate levels. The comprehensive strategy establishes AI Education Support Centres at regional offices by 2028 and expands AI-focused schools from 730 in 2025 to 2,000 by 2028, including new AI-oriented Meister vocational high schools.

The investment addresses South Korea’s paradox of ranking 13th globally in AI talent despite placing 6th in overall AI competitiveness—a gap the government considers a national priority.

Regional Dialogue on AI Adoption Challenges

On November 11, the Economic Research Institute for ASEAN and East Asia (ERIA) and OECD hosted a joint webinar examining AI adoption patterns across Europe and China. Professor Andrés Rodríguez-Pose from LSE highlighted a critical insight: AI adoption requires general skills that are accessible broadly, while AI generation demands specialized skills that tend to cluster in advanced regions, framing the challenge facing Asia-Pacific economies as fundamentally dual-track.

Europe: Trust, Governance, and Ethical Frameworks

Nordic Model Emphasizes Restraint and Accountability

Denmark released national AI guidelines for schools built on seven recommendations balancing opportunity with restraint: unified school frameworks, GDPR compliance, subject-level rules for student use, and mandatory teacher competence. The approach reflects Nordic values of trust, accountability, and parental involvement rather than top-down prohibition.

Italy Aligns AI Education with National Strategy

Italy’s Ministry of Education issued AI guidelines aligned with the country’s 2024-2026 AI strategy, emphasizing human-centric, lawful, and ethical frameworks. The guidance encourages initiatives improving learning and inclusion while maintaining uniform standards and mandating staff training—a governance scaffold moving from pilot projects toward system-wide reliability.

Latin America & Caribbean: Partnership Models and Mobile-First Reality

UNESCO-IBM Framework for Regional Higher Education

On November 12, UNESCO’s International Institute for Higher Education in Latin America and the Caribbean (IESALC) partnered with IBM to develop the first AI competence framework specifically designed for Latin American higher education contexts. The collaboration aims to equip educators and students with ethical AI application skills tailored to regional needs.

Mobile Devices as Primary AI Gateway

Groundbreaking research published November 29 in BMC Medical Education documented smartphone use patterns among medical students across seven Latin American countries. The multicenter study found 98.4% mobile phone ownership with 88.2% using devices for academic purposes—revealing mobile devices as the primary AI access point in resource-constrained contexts, contrasting sharply with desktop-focused approaches in the Global North.

Regional AI Competitiveness

The Latin American Artificial Intelligence Index (ILIA) reconfirmed Chile, Brazil, and Uruguay as regional pioneers in AI adoption. However, broader patterns suggest concerning talent drain across the region, with exceptions only in Costa Rica and Uruguay.

The Inter-American Development Bank hosted events November 11-13 on “Dual Training, AI, and Other Opportunities for Latin America and the Caribbean,” coordinating policy-level approaches to AI workforce development across the LAC region.

Private Sector Leadership

Colombia 4.0 Conference in Bogotá (November 1) highlighted the private sector as the driving force behind AI development in Latin America, providing a platform for regional collaboration on AI opportunities.

Middle East & Africa: Infrastructure Investments and Readiness Assessments

Bangladesh’s First AI Readiness Assessment

UNESCO and national partners released Bangladesh’s first AI RAM (Readiness Assessment Methodology). The diagnostic framework highlights strengths in e-government while identifying critical gaps in connectivity, data protection, and skills—providing evidence to guide National AI Policy development in South Asia.

North America: State-Level Momentum and Institutional Leadership

Twenty US States Issue AI Guidance

One of the month’s most significant policy developments occurred in the United States, where twenty states released AI guidance for K-12 public schools as of November 21—up from just a handful earlier in the year. States including Oklahoma, West Virginia, Washington, Montana, Ohio, and Virginia joined California and Oregon as leaders. This represents a fundamental shift from reactive prohibition toward proactive pedagogical frameworks emphasizing responsible use, ethics, and teacher training.

Wharton-OpenAI Global Educator Course

Wharton Online and OpenAI launched “AI in Education: Leveraging ChatGPT for Teaching” on Coursera on November 21, taught by AI in education pioneer Ethan Mollick. The four-module course covering generative AI essentials, ethical concerns, privacy issues, and practical applications signals mainstream acceptance of AI as a fundamental teaching competency, with global reach democratizing access to elite institutional knowledge.

Research Frontiers: Global Insights

Student Adoption Patterns Worldwide

A November 26 arXiv preprint examining LLM use by secondary students provides cross-cultural empirical insights into how K-12 students globally adopt AI-powered learning tools, focusing on accessibility and engagement patterns. The study offers crucial data as policymakers worldwide craft guidance frameworks.

Can AI Grade Fairly?

A November 19 study on AI-based grading systems investigates reliability through empirical procedures, addressing critical questions about automated assessment that transcend national boundaries. As educators worldwide face pressure to integrate AI tools, the question of fair, reliable automated grading becomes universal.

Equity as Central Global Concern

A recent arXiv paper on generative AI as catalyst explores algorithmic bias, data privacy, and equity gaps in pedagogical contexts, noting that rural and underserved schools lack AI infrastructure—risking exacerbated inequities. The authors call for collaboration between policymakers, educators, and developers, explicitly framing AI integration through an equity lens relevant to Global South contexts.

Global Conferences Shape International Dialogue

The month witnessed an unprecedented concentration of international AI-in-education convenings:

AI & The Future of Education (AIFE) brought 120+ attendees from 19 countries to Bangkok, Thailand, with 30+ sessions from 28 international experts—establishing a new Asia-focused platform for international dialogue.

Digital Education Council Global Summit convened over 100 global leaders in Singapore (November 5-6) to discuss digital education transformation with Asia-Pacific focus.

Looking Ahead

As we track these developments into December and beyond, a troubling pattern becomes increasingly clear: nations are not preparing their citizens for an AI future so much as systematically sorting them into tiers of access and capability. The wealthiest nations are building comprehensive AI literacy pipelines, while resource-constrained countries struggle with basic connectivity. What concerns me most is not the pace of change, but rather its direction—toward a world where one’s geographic and economic starting point fundamentally determines whether AI becomes a tool for empowerment or yet another mechanism of exclusion. The question facing us is not whether AI will transform education, but rather whether that transformation will honor the dignity and potential of all learners, regardless of where they live. December’s developments will reveal whether we are building bridges or deepening divides.

Research Methodology

This comprehensive review was compiled through systematic research using Perplexity AI, powered by Claude Sonnet 4.5 with advanced reasoning capabilities. All identified sources, including peer-reviewed papers, policy documents, conference proceedings, and institutional announcements, were individually examined and verified for credibility before inclusion, employing multi-layered search strategies spanning six continents, multiple publication types, and thematic dimensions including equity, governance, ethics, and pedagogy. All cited URLs were independently verified for accessibility and reliability prior to publication.

Currents is a monthly digest tracking developments at the intersection of AI, education, and society. Each edition synthesizes research, policy, and practice from around the world to provide educators, researchers, and policymakers with essential context for this rapidly evolving field.

Over the past year, however, we have noticed a new kind of stillness settling across our learning environments. It is not the reflective pause that precedes insight, nor the focused hush of collaborative work. It is a silence that feels suspended, like thought has been interrupted before it begins. Faculty describe it. Students embody it. And as practitioners, we sense it in our daily work: a subtle shift in how learners initiate, persist, or even approach the act of making meaning.

This article emerges from that shared observation. It is an attempt to name the unease we see but rarely articulate. It is also an invitation, to educators, designers, and scholars, to sit with a difficult question about what it means to learn in an age when generative systems can imitate the products of thinking more quickly than human minds can produce them.

What follows is not a warning, nor a lament. It is a reflective inquiry from two people who care deeply about the purpose of education. It is our effort to trace the contours of a new cognitive landscape and to ask what happens when the tools we design begin to reshape the very conditions of thought itself.

The Silence We Can No Longer Ignore

A distinct silence has fallen over the modern classroom, a quietude unlike the studious hush of the past, but rather the silence born from the friction of thought being stripped away from the educational process. It is the silence of the friction of thought being removed from the educational process. As educators, we find ourselves standing at a precipice, looking out at a generation for whom the blinking cursor is no longer a daunting provocateur of creativity, but a waiting command line for automation.

We pose a question that is as uncomfortable as it is necessary: Is the AI-generation “AI-damaged”? We do not use the term “damaged” to imply a medical pathology or an irreversible cognitive defect. Rather, we use it philosophically to describe an atrophy of the will, a thinning of the existential struggle that is required to turn information into knowledge, and knowledge into wisdom.

In this collaborative inquiry, we argue that the integration of gen AI into the academic sphere requires a total re-evaluation of the purpose of education. If we continue to assess students on the process of inquiry rather than on the production of answers, we risk cultivating a generation proficient in prompting but deficient in thinking.

The Outsourcing of Cognition

To understand the potential “damage”, we must first understand the nature of the tool itself. Unlike the calculator, which offloaded the drudgery of arithmetic to allow for higher-order mathematical reasoning, Large Language Models (LLMs) threaten to offload the reasoning itself. Writing is not merely the transcription of thoughts; it is the mechanism by which thoughts are formed. As Neil Postman warned in his book titled Technopoly, technology is never additive; it is ecological. It changes everything. Postman argued that “new technologies alter the structure of our interests: the things we think about. They alter the character of our symbols: the things we think with. And they alter the nature of community: the arena in which thoughts develop” (Postman, 2011, p. 20).

When a student outsources the drafting of an essay to an algorithm, they are not merely cheating the system; they are bypassing the neural architecture building that occurs during the struggle of articulation. The “damage” manifests as a loss of cognitive stamina. We are witnessing a decline in the ability to sit with ambiguity, to wrestle with a complex text, and to synthesize disparate ideas without a digital intermediary (such as Kosmyna et al., 2025). If the struggle is removed, the growth is removed. The outcome is a polished product masking a hollowed-out process.

The Ontological Crisis of Education

If the machine can produce the output, what is the value of the human element? This brings us to the definition of education itself. For the better part of the last century, education has been viewed through a transactional lens: the teacher deposits information, and the student withdraws it upon request for an exam. This “banking concept of education” as critiqued by Paulo Freire (Freire, 2018), is rendered obsolete by AI. The bank is now open to everyone, everywhere, instantly.

Therefore, we must redefine the purpose of education from acquisition to formation. The goal can no longer be competency in the retrieval of facts or the mimicry of standard prose. The goal must be the cultivation of the self. We must shift our pedagogical focus toward Gert Biesta’s concept of “subjectification”, the process by which students become autonomous, thinking subjects rather than objects to be filled. Biesta argues that education must always involve a “beautiful risk”, noting that “if we take the risk out of education, there is a real chance that we end up with something that is no longer education” (Biesta, 2015).

AI removes the risk. It provides a safe, average, and sometimes hallucinated answer. To repair the “AI-damage”, we must reintroduce the risk. We must create environments where failure is not a metric of incompetence but a necessary step in the dialectic of learning. Education must become a sanctuary for human connection, ethical debate, and the messy, inefficient process of original thought.

Redefining Assessment

If we accept that the purpose of education is the development of the human intellect and character, our assessment standards must change radically. The traditional five-paragraph essay, the take-home summary, and the multiple-choice test are artifacts of a pre-AI world. Continuing to use them is akin to measuring the speed of a car with a ruler; the tool is no longer fit for reality.

We propose a shift from output-oriented assessment to process-oriented assessment. The current model asks, “What did you produce?” The new model must instead ask, “How did you arrive here?”, and “What did you learn?”.

This requires a return to older, more human-centric forms of evaluation that AI cannot mimic. We advocate for the return of the viva voce, the oral defense. In a conversation, a student cannot hide behind a screen or a generated script. Their understanding, their hesitation, and their ability to connect concepts in real-time become the metric of success. We must value the “un-hackable” elements of the student experience: their ability to debate, to empathize, and to apply ethics to complex scenarios.

Furthermore, assessments should move toward what John Dewey described as “continuity of experience” (Dewey, 1940) and what Grant Wiggins identifies as “authentic” (Wiggins, 1990). Instead of isolated assignments that can be easily generated by a bot, assessments should be longitudinal projects requiring iteration, physical engagement with the community, and personal reflection that links specific, lived experiences to course material. They should reflect real-world tasks that embody deep understanding, higher-order thinking, and complex problem solving. An AI can analyze a text about poverty; it cannot interview a local shelter director and reflect on how that conversation challenged the student’s specific worldview.

The Ethical Imperative: Agency over Algorithms

The most profound aspect of the AI-damage is the potential erosion of ethical agency. When we allow an algorithm to curate our arguments and select our evidence, we tacitly allow it to shape our moral compass. Education must essentially become a training ground for cognitive sovereignty.

We must teach students that using AI to do their thinking is an act of self-erasure. The ethical violation is not just against the academic institution (plagiarism); it is against the self. It is a voluntary surrender of one’s voice to a statistical average of the internet.

To counter this, we propose a curriculum that integrates AI literacy not as a “how-to” for prompting, but as a philosophy of technology. Students must understand the bias (Noble, 2018), the hallucinations, and the environmental costs of these models. They must learn to treat AI as a sparring partner, not a ghostwriter. The assessment then becomes: “Here is what the AI produced on this topic; now, critique it. Where does it lack humanity? Where does it lack nuance? How can you, as a human, improve upon this machine output?”

Final Thoughts: The Renaissance of the Human

Is the AI-generation AI-damaged? Potentially, yes. If we leave the current structures of education in place, we doom them to a future of cognitive dependency. We risk creating a class of humans who are merely managers of machines, unable to distinguish between a calculated probability and a genuine truth.

However, this crisis is also our greatest opportunity. The presence of AI forces us to distill what is truly human about learning. It forces us to strip away the busy work, the rote memorization, and the performative writing that has cluttered our syllabi for decades.

We are called to usher in a new educational renaissance. This renaissance prioritizes the development of wisdom over the accumulation of data. It values the stammering, imperfect, original voice of a student over the flawless, sterile syntax of a bot. By redefining our purpose and our assessments, we can ensure that the tools of the future serve the development of the human spirit, rather than replacing it. We must teach the next generation that the difficulty of learning is not a bug to be bypassed, but the very feature that makes us human.

References

• Biesta, G. J. J. (2015). Beautiful risk of education. Routledge.
• Dewey, J. (1940). Nature in experience. The Philosophical Review, 49(2), 244. https://doi.org/10.2307/2180802
• Freire, P. (2018). The banking concept of education. In Thinking About Schools (pp. 117–127). Routledge. https://doi.org/10.4324/9780429495670-11
• Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025, June 10). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv.Org. https://arxiv.org/abs/2506.08872
• Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. New York University Press. https://doi.org/10.18574/nyu/9781479833641.001.0001
• Postman, N. (2011). Technopoly: The surrender of culture to technology. Vintage.
• Wiggins, G. (1990). The case for authentic assessment. Practical Assessment, Research, and Evaluation, 2(1), 2. https://doi.org/10.7275/ffb1-mm19

Cite this paper: Gattupalli, S., & Giovannini, J. (2025). Is the AI-Generation AI-Damaged? Society and AI. https://societyandai.org/perspectives/is-the-ai-generation-ai-damaged/

I believe the answer lies in a concept I call Universal Basic AI: the idea that every person deserves access to a personalized AI companion that is fundamentally aligned with their individual wellbeing, open in its architecture, and sovereign in its loyalty. This is not a luxury or a privilege. It is becoming a necessity for meaningful participation in 21st-century life. Just as access to education, healthcare, and information became recognized as essential for human dignity, access to cognitive support through AI must be understood as a foundational right.

The current trajectory of AI development points toward a world where powerful systems serve primarily as tools for extraction: attention, data, labor, and capital. These systems are optimized for engagement metrics, advertising revenue, or institutional efficiency, not for the flourishing of the individuals who interact with them. A Universal Basic AI reverses this dynamic entirely. It is a system that learns your values, understands your goals, protects your privacy, and works tirelessly to help you become who you want to be. It grows with you from childhood through adulthood, serving as a lifelong cognitive partner rather than a transactional tool.

This working paper outlines the philosophical foundations and technical requirements for making Universal Basic AI a reality. It identifies three non-negotiable pillars: Openness, Alignment, and Sovereignty. These must guide the development of personal AI systems. Openness ensures transparency and prevents lock-in to proprietary platforms. Alignment means the system’s objectives are genuinely oriented toward your flourishing, not toward hidden commercial or political agendas. Sovereignty guarantees that the AI’s primary loyalty is to you, with privacy-preserving architectures that keep your data and your cognitive patterns under your control.

The stakes could not be higher. Without deliberate intervention, AI will likely amplify existing inequalities, creating a cognitive divide far more consequential than the digital divide of previous decades. But if we choose to build these systems correctly, with openness, alignment, and sovereignty as foundational principles, we have the opportunity to create the greatest capability multiplier in human history, one that empowers every individual to reach their full potential.

Cite this paper: Gattupalli, S. (2025). What is Universal Basic AI? Society and AI Working Papers. https://societyandai.org/fundamentals/universal-basic-ai/

Adopting is about utility; adapting is about evolution. Adopting is bringing the tool to the task; adapting is changing the task, and perhaps the self, to leverage the new reality. In the emerging AI societies, we must do both simultaneously. And the vehicle that enables this dual movement, the bridge between the raw power of silicon and the potential of the human mind, is education.

The Velocity of Change

The urgency of this conversation is underscored by the relentless pace of technological advancement. Currently, all chatbots, including the very recent Gemini 3 from Google, are showing breathtaking performance across all benchmarks. The capabilities we see today in reasoning, coding, and multimodal understanding were science fiction only a few years ago. These models are not just answering questions; they are generating code, creating art, and solving complex logic puzzles with a fidelity that challenges human expertise.

It is only a matter of weeks or months before every other major player releases their next iteration. We are on a trajectory where Gemini 4, 5, and their counterparts from other labs will arrive with increasing frequency and capability. The window between “state of the art” and “legacy” is shrinking to a vanishing point. If we focus solely on adopting, on learning the specific quirks of today’s model, we will be perpetually obsolete, constantly running on a treadmill of updates and patch notes.

The Force Multiplier

I say this mainly because no matter how many iterations, improvements, and step changes happen in the model architectures, the force multiplier ultimately converges and focuses on education.

The most powerful AI system is inert without a human intent to guide it. The most sophisticated reasoning engine is useless without a human question to provoke it. As AI models become more capable, the bottleneck shifts from the technology’s limit to the user’s imagination and critical capability. The “force” is the AI, but the “multiplier” is the educated human mind.

This is why education is the vehicle for adaptation. It is not enough to teach students how to use a specific AI tool, because that tool will change by the time they graduate. Instead, we must teach them the principles of AI fluency and AI literacy. We must cultivate the meta-skills of adaptation: cognitive flexibility, ethical reasoning, and the ability to learn continuously. We need to move beyond the “how-to” of interface interaction to the “why” and “what-if” of systems thinking.

Starting with Early Learners

We need to start educating early learners because this is where we should be training them to be more AI fluent and AI literate. Children growing up today are the first generation of true “AI natives.” For them, adoption will be intuitive. They will talk to machines as naturally as they talk to friends. But intuition is not the same as understanding.

The challenge for educators is to ensure these young minds are also adapting. Are they learning to question the output of the AI? Are they understanding the difference between a probabilistic answer and a factual truth? Are they developing the creative confidence to use AI as a partner rather than a crutch?

We must design curricula that treat AI not as a forbidden shortcut, but as a new medium of thought. Just as we teach grammar to structure language and logic to structure arguments, we must teach “AI rhetoric”, how to converse with, guide, and evaluate synthetic intelligence. By embedding AI literacy into early education, we are not just preparing workers for the future economy; we are preparing citizens for a future society. We are giving them the tools to shape the technology rather than be shaped by it.

The Symbiosis of Adopting and Adapting

Adopting and adapting are not separate processes; they are the left and right foot of our march forward. We adopt the tools to gain efficiency; we adapt our mindsets to gain wisdom.

In an AI society, the risks of failing to adapt are high. We risk creating a divide not just of access, but of understanding, a society where a few wizards command the algorithms while the many are directed by them. We risk surrendering our agency to black boxes we do not comprehend.

But the opportunities are equally vast. If we can successfully adopt these powerful tools while adapting our educational systems to foster human resilience and creativity, we can unlock a golden age of personalized learning and amplified human potential. We can create a society where AI handles the drudgery of computation, freeing humans to focus on the poetry of connection and the rigor of invention.

The technology will keep marching on. Gemini 3 will become Gemini 4. The benchmarks will keep rising. But the variable that will determine whether this story is one of displacement or empowerment is us. It is how we choose to learn, and how we choose to teach. Education is not just a sector to be disrupted by AI; it is the very foundation upon which a successful AI society must be built.

Published by researchers at Reichman University, Israel.

The Conceptual Shift: From Universal Design to Psychological Pluralism

This commentary responds to “When technology meets personality: toward human-centered AI design” by Yair Amichai-Hamburger, Maya Mentzel Mazler, Abigail Barazani, and Yifat Ben-David Kolikant at Reichman University, Israel. The authors propose that understanding personality differences is essential for designing Artificial Intelligence and Internet-of-Things (IoT) systems that genuinely align with human needs. Their review traverses three domains—information and communication technologies (ICT), transportation, and medical technologies—to argue that AI design must adapt to users’ psychological diversity rather than assuming uniform human response.

Rather than restating their detailed mappings across the Big Five traits, attachment theory, sensation seeking, need for closure, and need for cognition, this commentary focuses on a central issue raised by the article: What does it mean for AI to become human-centered when “the human” is not a standard unit, but a heterogeneous constellation of cognitive, emotional, and relational profiles? This question sits at the core of Society & AI’s mission, which examines how knowledge systems, equity frameworks, and conditions for human flourishing must evolve as AI becomes a structural force within social life.

Personality-Aware AI as a Missing Layer in Human-Centered Design

Human-centered AI is frequently interpreted as usability work: designing interfaces that are friendly, transparent, and minimally harmful. The Reichman University team pushes beyond this by identifying psychological diversity as the conceptual blind spot of current design frameworks. This marks an important shift. If AI systems increasingly mediate cognition, communication, mobility, and medical decision-making, then ignoring personality differences is not an oversight—it is a structural failure that risks widening disparities in trust, adoption, and well-being.

Recent frameworks, such as Qin et al.’s capability–personalization model or Shin’s system-centered analysis of algorithmic bias, treat users as relatively interchangeable. The focal article demonstrates that such models collapse in real use cases. The anxious passenger, the sensation-seeking gamer, the conscientious patient, or the avoidantly attached telehealth user do not engage technologies in the same way—nor should they be expected to.

A core contribution of the research team is to foreground this diversity and argue for AI systems that provide pathways, not prescriptions. This move aligns personality psychology with the broader push toward pluralistic AI design—an approach that recognizes variation not as noise, but as a fundamental design parameter.

Where the Article Opens New Terrain—and Where Gaps Remain

The article persuasively develops the case that personality matters for AI interaction. Yet several unresolved challenges emerge that warrant deeper examination:

1. Personality-Aware Design Risks Becoming Personality-Based Nudging

Fine-grained adaptation—particularly in IoT and autonomous vehicles—creates the possibility of over-personalization. If an anxious user receives constant reassurance, or a high-sensation seeker is fed increasingly stimulating interactions, AI may inadvertently reinforce traits rather than support human flourishing. This raises questions about autonomy, emotional manipulation, and the boundary between personalization and behavioral shaping. The authors gesture to ethical concerns but stop short of articulating concrete safeguards.

2. Personality Traits Are Not Static, and AI Systems Need Dynamic Models

Personality-aware AI presumes traits as stable indicators. Yet evidence shows that digital behavior itself reshapes tendencies such as impulsivity, anxiety, or sociability. Without dynamic modeling, AI could harden transient states into permanent labels. The authors identify the need for longitudinal research; I extend this urgency: AI must avoid reifying the very traits it tries to accommodate. Systems designed today will shape the psychological profiles of tomorrow’s users—a recursive relationship that demands careful governance.

3. Cultural Context Is Insufficiently Addressed

Personality–technology interactions vary dramatically across societies. As the article notes briefly, conscientiousness predicts AI acceptance differently in Western versus Arab contexts. This suggests that cultural personality ecologies matter, and designing “universal” personality-aware systems may inadvertently encode Western psychological norms as defaults. The research would benefit from deeper engagement with cross-cultural psychology and non-WEIRD (Western, Educated, Industrialized, Rich, Democratic) populations.

4. The Framework Requires Integration With Equity and Accessibility

If personality-aware design becomes the next frontier of human-centered AI, who gets optimized experiences and who receives generic defaults? AI systems could inadvertently privilege users whose data patterns are easier to model—typically those from majority demographic groups with extensive digital footprints. This raises pressing equity questions aligned with Society & AI’s broader research agenda: How do we ensure that personality-aware systems do not become another vector for algorithmic inequality?

A Forward-Looking Perspective: Toward Pluralistic AI Ecosystems

The authors conclude by proposing a personality-aware design approach. This commentary extends their argument: the future of human-centered AI is not personalization but pluralism.

Pluralistic AI ecosystems recognize:

• Diverse ways of processing information
• Diverse forms of agency and control preferences
• Diverse social-relational needs and cultural contexts

They provide multiple modes rather than a single optimized experience—modes that evolve with the user and allow the user to calibrate how adaptive the system should be. Such ecosystems shift the locus of design from prediction to partnership, ensuring human flourishing rather than psychological sorting.

Consider what this means concretely:

• An autonomous vehicle that offers not one “optimal” route but several options with explicit trade-offs (fastest, scenic, most predictable)
• A telehealth platform that allows users to dial up or down the level of emotional support and reassurance they receive
• A learning system that reveals its assessment of user traits and allows correction or override

This approach honors psychological diversity while preserving user agency—a balance the focal article identifies as necessary but does not fully operationalize.

Methodological Reflection: What Counts as Evidence?

The article synthesizes research across personality psychology and human-computer interaction. However, much of the cited evidence comes from self-report measures and correlational studies in controlled settings. As AI systems deploy at scale in uncontrolled environments—homes, highways, hospitals—new methodological challenges emerge:

• How do personality effects interact with contextual variables (time pressure, social context, emotional state)?
• Can personality-aware systems account for within-person variation across situations?
• What validation frameworks ensure that personality-informed design improves outcomes rather than merely reinforcing existing patterns?

I recommend that future research in this area prioritize longitudinal, ecologically valid studies that track how personality–AI interactions unfold in real-world settings over time. The laboratory findings are suggestive; the societal stakes demand empirical rigor at scale.

Take-Home Message

The focal article provides an important foundation for understanding psychological diversity in AI design. This commentary argues that the next step is to integrate personality-aware insights into broader debates about autonomy, equity, and human flourishing. Designing AI that adapts to personality is not enough; we must build systems that empower users to evolve beyond their traits rather than becoming constrained by them.

The vision is not AI that knows you better than you know yourself—it is AI that offers you the conditions to become who you wish to be. That requires moving from algorithmic personalization to relational design: systems that understand psychological diversity as a resource for human agency, not merely a parameter for optimization.

As AI becomes embedded in the infrastructure of daily life, the question is not whether personality matters—it clearly does. The question is whether AI will use that knowledge to expand human possibility or narrow it. The answer depends on whether designers, researchers, and policymakers commit to pluralistic ecosystems that honor the full range of human psychological experience.

References and Further Reading

Focal Article:

• Amichai-Hamburger, Y., Mentzel Mazler, M., Barazani, A., & Ben-David Kolikant, Y. (2025). When technology meets personality: toward human-centered AI design. AI & Society. https://doi.org/10.1007/s00146-025-02720-0 (Published November 14, 2025, Open Access. Reichman University, Israel.)

Commentary Guidelines:

• Berterö, C. (2016). Guidelines for writing a commentary. International Journal of Qualitative Studies on Health and Well-Being, 11(1), 31390. https://doi.org/10.3402/qhw.v11.31390 (This commentary follows the structural and methodological guidance outlined in Berterö’s framework.)

This commentary is part of Society & AI’s ongoing examination of how AI systems can honor human diversity while supporting equity and flourishing. Feedback and discussion are welcome at feedback@societyandai.org.

Ms. Bodnar teaches third-grade mathematics in Springfield, Massachusetts. This morning, she’s planning a lesson on fractions. Five years ago, she would have reached for her textbook, perhaps added a worksheet, maybe searched for a video if she had time. Today, she opens a conversation with AI and types:

“Create a rap song about fractions that uses pizza slices as examples. Make it fun for 8-year-olds and scientifically accurate.”

The AI generates lyrics in seconds. Ms. Bodnar reviews them. The rhymes work, the math is correct, but she wants her students to hear it as actual music, not just words on a screen. She remembers there’s another AI tool that can turn lyrics into songs. She copies the lyrics, pastes them into the music generation tool, and specifies “upbeat hip-hop for kids, 95 BPM.” Within minutes, she has a complete educational rap about fractions. Something that would have required a recording studio, musical expertise, and significant budget just months ago.

The next morning, she plays it in class. Her students are engaged, moving to the beat, singing along. One student, inspired by seeing that Ms. Bodnar created this rather than just finding it, asks if they can make a song about their family’s traditional recipes, turning grandmother’s cooking measurements into a math lesson that honors their heritage.

This scene captures what we call vibe pedagogy: a new educational practice made possible by the unprecedented “intelligence abundance” (Mollick, 2024) of conversational AI systems. What was once impossible for most educators (creating original, culturally responsive, multimodal learning experiences) has become accessible through natural conversation. This represents something that simply didn’t exist five years ago and fundamentally changes what educators and learners can create together.

The Dual Meaning of “Vibe”

The term vibe pedagogy works on two levels:

1. The Method: Pedagogy with a “Vibe”

First, it describes how teaching happens, through the intuitive, conversational, almost vibes-based interaction with AI tools. Unlike traditional educational technology that requires coding skills or specialized software training, vibe pedagogy uses natural language. An educator articulates their pedagogical vision (“I want my students to understand photosynthesis through music”) and AI helps materialize it. The “vibe” comes from this conversational flow: iterative, exploratory, responsive. More like talking through an idea with a creative partner than programming a machine.

2. The Outcome: Pedagogy That Captures the Right “Vibe”

Second, it describes what gets created: learning experiences that resonate with students’ cultural contexts, emotional needs, and personal interests. Through AI assistance, educators can now tailor content to capture the specific “vibe” that will engage their learners in their context. A lesson about weather patterns can reference local climate events. A physics problem can feature sports students actually play. Historical concepts can connect to students’ family stories. The resulting pedagogy carries a vibe (an atmosphere, a feeling, a cultural resonance) that makes learning personally meaningful.

What Makes This New?

Five years ago, creating culturally responsive, multimodal, personalized learning materials required:

• Technical expertise (coding, video editing, music production)
• Significant time (hours or days per customized lesson)
• Financial resources (software licenses, professional development)
• Specialized knowledge (graphic design, educational game design)

These barriers meant that personalized, culturally sustaining pedagogy, while theoretically valuable, remained practically inaccessible for most educators.

Vibe pedagogy changes this reality. Through conversational prompting, any educator can now:

• Transform concepts into songs, simulations, stories, or games
• Adapt content to students’ linguistic and cultural backgrounds
• Create multimodal representations (visual, auditory, interactive) of the same concept
• Generate differentiated materials for diverse learning needs
• Produce these resources in minutes rather than days

This isn’t about AI replacing teachers. It’s about AI expanding what teachers and students can create, shifting education from content consumption to content composition.

A Concrete Example: Ms. Bodnar’s Lesson Evolution

Let’s follow Ms. Bodnar’s fraction lesson through the vibe pedagogy process:

Traditional Approach (Pre-AI):

• Use textbook problems: “If you have 3/4 of a pizza…”
• Draw circles on the board divided into pieces
• Assign worksheet with standard fraction problems
• Time required: Planning (20 minutes), Teaching (45 minutes)

Vibe Pedagogy Approach:

Step 1: Pedagogical Vision (What does she want students to understand?)

• Fractions represent parts of wholes
• Fractions connect to real life, not just math class
• Learning should feel engaging and multimodal

Step 2: Conversational Prompting (How does she bring this vision to life?)

First prompt:

“Create a rap song about fractions for third-graders that uses pizza slices as examples. Make it scientifically accurate but fun.”

AI generates lyrics. Ms. Bodnar reviews and refines:

“Good start, but make the chorus catchier and add a verse about comparing fractions (which slice is bigger).”

Step 3: From Lyrics to Music Satisfied with the lyrics, Ms. Bodnar uses a music generation AI tool:

“Turn these lyrics into an upbeat hip-hop song for kids, 95 BPM”

Within minutes, she has a complete educational rap: lyrics, beat, melody. What would have required professional musicians and studio time is now created through conversational prompting.

Step 4: Student Agency The next day, inspired by seeing the creation process, a student asks: “Can I make a song about fractions using my grandmother’s cooking measurements?”

Ms. Bodnar guides the student through the same prompting process. The student creates a fraction lesson that becomes a class resource, honoring family knowledge while demonstrating mathematical understanding.

Time required: Planning (30 minutes), Teaching (45 minutes), but the lesson is now multimodal, engaging, and has sparked student-led content creation.

From Consumption to Creation

The deeper transformation vibe pedagogy enables is a shift in who gets to be a creator of educational content.

Traditionally, curriculum flows in one direction:

• Publishers create content → Teachers deliver content → Students consume content

Vibe pedagogy disrupts this hierarchy:

• Teachers become designers of learning experiences
• Students become creators of knowledge artifacts
• The boundaries between “teacher” and “learner” become productively blurred

When a multilingual student uses AI to create chemistry explanations in their home language, when a student with learning differences designs visual narratives that help peers understand difficult concepts, when students incorporate their cultural knowledge into academic content, these acts represent what we call epistemic agency: the capacity to be recognized as legitimate creators of knowledge, not merely its consumers.

The Foundations: Intelligence Abundance

Vibe pedagogy emerges from what Ethan Mollick (2024) calls our era of “intelligence abundance,” the sudden availability of powerful AI systems that can generate text, code, images, music, and simulations through natural language conversation. This abundance creates new possibilities:

• Democratized Creation: You don’t need to be a programmer, musician, or graphic designer to create educational media
• Rapid Iteration: Generate, test, and refine learning materials in real-time
• Personalization at Scale: Create culturally specific content without sacrificing pedagogical rigor
• Multimodal Expression: Transform concepts across media (text → song → simulation → game)

What Vibe Pedagogy Is Not

It’s important to clarify what vibe pedagogy does not mean:

Not automation: AI isn’t replacing teacher judgment. It’s amplifying teacher creativity.

Not efficiency for efficiency’s sake: The goal isn’t faster lesson planning but richer, more culturally responsive learning.

Not AI-generated assignments: This isn’t about AI doing students’ work but AI enabling students to create learning resources.

Not uncritical adoption: Vibe pedagogy requires critical evaluation of AI outputs for bias, accuracy, and cultural appropriateness.

Practical Implications

For educators considering vibe pedagogy:

Start Small:

1. Try creating one culturally responsive learning artifact (a song, a story, a simple simulation)
2. Show students how you created it. Make the process visible.
3. Invite students to create their own versions

Build Capacity:

• Develop “prompt literacy,” the ability to articulate pedagogical vision clearly to AI
• Learn to iterate. First outputs are rarely perfect; refinement is part of the process
• Practice critical evaluation. Not all AI-generated content is equally good or appropriate

Center Student Agency:

• Don’t just create for students; create with them
• Hand students the tools to become creators
• Value student-generated content as legitimate learning artifacts

Looking Forward

Vibe pedagogy doesn’t answer all questions about AI in education. Concerns about equity (who has access to these tools?), bias (whose cultural knowledge do AI systems privilege?), and labor (does this devalue human creativity?) remain urgent and unresolved.

But it does offer a different vision: Rather than asking “How can AI make learning more efficient?” we ask “How can AI make everyone a creator of knowledge?”

When Ms. Bodnar’s student creates that cooking measurement fraction lesson, she’s not just learning math. She’s learning that her grandmother’s knowledge belongs in the classroom. That her family’s traditions are valuable. That she can be a teacher, too.

That shift, from consumer to composer, represents education’s next frontier in an age of intelligence abundance.

Further Reading

• Mollick, E. (2024). Co-Intelligence: Living and Working with AI. Penguin.
• Gattupalli, S., Maloy, R. W., & Edwards, S. A. (2023). Prompt literacy: A pivotal educational skill in the age of AI. ScholarWorks @ UMass. https://doi.org/10.7275/3498-wx48
• Paris, D., & Alim, H. S. (2017). Culturally Sustaining Pedagogies: Teaching and Learning for Justice in a Changing World. Teachers College Press.

This article introduces the concept of vibe pedagogy as a fundamental educational practice. For deeper exploration of implementation, case studies, and research foundations, see our ongoing work on human-AI collaboration in education.

In 2024, I coined the term promptgramming—a deliberate fusion of pro(mpt) and programming—because I recognized we stood at such a threshold. This article serves as a companion to my technical report, The Art and Science of Promptgramming (Gattupalli, 2024), but here I want to explore what this moment means for how we live, learn, and imagine ourselves in societies increasingly mediated by artificial intelligence.

The End of Code as Gatekeeping

For seventy years, programming meant learning languages computers understood. You bent your mind to the machine’s logic, mastered its syntax, paid your dues through debugging and compilation errors. This created a priesthood: those who could speak to computers, and those who could not.

Promptgramming represents something different—not the end of programming, but its democratization. When a high school student asks ChatGPT to explain cellular respiration in the voice of a sports commentator, she is programming. When a teacher designs a prompt that scaffolds historical perspective-taking, he is programming. The syntax is natural language; the compiler is a large language model; the output is not code but meaning.

This is not a minor shift. It is the dissolution of the barrier between human expression and computational action. And it demands that we think carefully about what happens when billions of people can suddenly “program” without ever learning to code.

The Architecture of Thought in AI Societies

In societies where AI mediates access to knowledge, employment, healthcare, and governance, prompting becomes more than a skill. It becomes an architecture of thought—a way of structuring intention, of negotiating with intelligent systems, of asserting what we want the world to know and how we want to know it.

Consider what happens when a student asks an AI to write an essay. At one level, this is plagiarism. At another, it is a failure of promptgramming literacy. A student fluent in promptgramming does not ask the AI to write the essay. She asks it to challenge her argument, to generate counterexamples to her thesis, to translate her ideas into a different disciplinary register so she can see them anew.

The difference is not technical. It is epistemological. It reflects whether we see AI as a replacement for thinking or as an instrument for thinking better.

This distinction matters because in AI societies, those who prompt well will have fundamentally different access to opportunity than those who prompt poorly—or who never learned to prompt at all. If we treat this as a technical skill rather than a civic literacy, we risk creating new hierarchies as entrenched as those that came before.

Why Promptgramming Is Not Prompt Engineering

The corporate world speaks of “prompt engineering”—the art of extracting maximum value from AI systems. I chose a different term deliberately. Engineering implies optimization, efficiency, the instrumental logic of getting what you want from a machine.

Promptgramming implies something else: a practice that fuses computational and humanistic thinking, that recognizes prompts as pedagogical artifacts, that centers questions of equity and access alongside questions of output quality.

When I design a prompt for educational use, I am not engineering an output. I am curating an encounter—between a learner and an AI, yes, but more fundamentally between a learner and her own emerging understanding. The prompt is the scaffold, the provocation, the question that opens rather than closes.

This matters because the logics we import from industry often serve industry’s values: efficiency over equity, scale over context, optimization over care. Education requires different values. So does democracy. So does justice.

If promptgramming is to serve these values, it must be theorized and practiced as something more than engineering.

Teaching Promptgramming Now

The pace of AI integration into education has accelerated faster than anyone anticipated. What seemed experimental three years ago is now ubiquitous. Students already use AI daily—for homework, for research, for writing. The question is not whether to teach promptgramming, but how quickly we can make it universal.

Schools and universities need to integrate promptgramming literacy into existing curricula immediately. Not as a separate course, but woven through every subject. A biology class teaches both photosynthesis and how to prompt AI to explain it in different ways. A history class teaches both historical analysis and how to use AI to explore multiple perspectives. A writing class teaches both composition and how to leverage AI for revision and critique.

This is urgent because the gap between students who learn to prompt well and those who don’t is widening daily. Those who develop promptgramming literacy gain capacities essential to knowledge work: articulating intentions clearly, iterating based on feedback, evaluating machine outputs critically. Those denied this literacy will face systematic disadvantage in education and employment.

What Promptgramming Literacy Includes

Promptgramming literacy involves three core capacities:

Iterative Construction: Formulating clear prompts and refining them based on outputs, much as we draft and revise writing.

Critical Interpretation: Evaluating AI-generated content with disciplinary knowledge and ethical awareness—asking not just whether an output is useful, but whether it is accurate, whose perspective it centers, what it makes invisible.

Ethical Reasoning: Understanding the implications of outsourcing cognitive labor to AI. What gets lost? Who benefits? When does use become dependence?

These are not supplemental skills. They are foundational to living thoughtfully in societies where AI mediates access to information, employment, healthcare, and governance.

Final Thoughts

The transformation is already underway. Students, teachers, and researchers are experimenting with prompting daily, often without formal training or shared frameworks. What we need now is intentional effort—to develop curricula, establish standards, and ensure access across all educational contexts. Promptgramming should not be reserved for those with resources or technical backgrounds. It is a literacy everyone needs, and everyone can learn. The work begins with recognizing its importance and committing to teach it well.

We live in what I call algorithmic societies: communities where automated systems make consequential decisions about resource allocation, access to opportunities, and the very structure of how we learn and work together. This isn’t science fiction. It’s the mundane reality of 2025, where machine learning models determine university admissions, recommend interventions for struggling students, and increasingly govern how knowledge itself circulates through our institutions.

But here’s what troubles me: we’re allowing these systems to entrench themselves in education—humanity’s most foundational adaptive mechanism—without asking the most essential question. Not whether algorithms can make these decisions, but whether they should.

When Algorithms Become Governance

The shift from human to algorithmic decision-making in education represents more than technological upgrade. It represents a fundamental transformation in how we organize society itself. Consider what’s already happening: predictive models assess student success trajectories, recommendation engines curate learning pathways, and automated systems increasingly determine who receives additional support and who doesn’t.

Wang (2024) identifies three features of algorithmic decisions that demand careful attention: their speed of development and deployment, their lack of interpretability, and their tendency to develop capabilities beyond original design. In education, this means decisions that once involved human deliberation—which students need intervention, which teachers require additional training, which programs deserve continued funding—now happen at machine speed, often without meaningful opportunity for review or appeal.

This creates what scholars call the “privatized state” problem. When educational institutions outsource consequential decisions to proprietary algorithmic systems, they effectively cede governance to private actors driven by profit rather than public interest. The result isn’t just a technical problem—it’s a democratic one. Students and families become subject to decisions they cannot understand, contest, or meaningfully influence.

I’ve watched this play out in my own research on educational technology. A learning management system recommends certain students for remediation based on patterns invisible to teachers. A college admissions algorithm surfaces some applications while burying others. These aren’t neutral tools—they’re governance mechanisms, and they increasingly operate beyond the reach of traditional accountability structures.

Education: Both Subject and Agent of Algorithmic Change

Education occupies a peculiar position in algorithmic societies. It’s simultaneously being reshaped by algorithms and serving as the primary mechanism through which we might develop the capacity to govern them responsibly.

On one hand, educational institutions face mounting pressure to adopt AI-driven systems for personalization, assessment, and resource allocation. The logic seems compelling: algorithms can process more information, identify patterns humans miss, and scale interventions beyond what any teaching staff could manage manually. Schools and universities implement these systems with genuine hope that they’ll expand access and improve outcomes.

But we’re learning the hard way that algorithms trained on historical data tend to perpetuate historical inequities. Machine learning models in college admissions have been shown to reproduce racial biases. Predictive analytics for student success continue producing racially skewed results despite efforts at “debiasing.” Automated essay scoring systems struggle with non-standard English dialects, effectively penalizing students from linguistically diverse backgrounds.

The challenge isn’t just technical. As Bozkurt (2023) emphasizes, navigating algorithmic societies requires a new form of literacy—what he calls “generative AI literacy.” This means not just knowing how to use AI tools, but understanding their ethical implications, recognizing their biases, and critically evaluating their limitations. Education must cultivate this capacity if we hope to produce citizens capable of meaningful participation in algorithmically governed systems.

Yet here’s where education’s dual role becomes crucial: the same institutions being transformed by algorithms must also teach the next generation how to govern them. Teachers need to understand how recommendation systems work so they can help students develop healthy information diets. School administrators need algorithmic literacy to evaluate vendor claims critically. Students need practice questioning automated decisions so they can demand accountability in their adult lives.

This is where I see education’s leverage point. We’re not passive recipients of algorithmic change—we’re the mechanism through which societies develop adaptive capacity. Every classroom that teaches critical data literacy, every school that requires transparent documentation of algorithmic systems, every educational policy that demands human review of automated decisions helps build the civic infrastructure necessary for responsible algorithmic governance.

The Illusion of Neutrality

One of the most persistent and dangerous myths about algorithmic societies is that automated systems are somehow more objective than human judgment. After all, algorithms don’t have prejudices, don’t play favorites, don’t make emotional decisions. Right?

This framing fundamentally misunderstands how these systems work. Algorithms aren’t neutral arbiters—they’re crystallized human judgment at scale. Every dataset reflects historical choices about what to measure and what to ignore. Every model architecture embeds assumptions about what patterns matter. Every optimization function encodes specific values about what outcomes we should pursue.

In educational contexts, this matters profoundly. When we deploy automated essay scoring, we’re not eliminating bias—we’re enshrining particular linguistic norms as universal standards. When we use predictive analytics for student success, we’re not discovering objective trajectories—we’re reifying patterns from historical data that may reflect past discrimination rather than innate potential.

I’ve seen educational technology companies market their products as “removing human bias” from decision-making. But as recent research on algorithmic bias in educational systems demonstrates, these tools often amplify existing inequities precisely because they lack the contextual understanding that experienced educators bring to complex situations. A teacher might recognize that a student’s recent performance decline stems from family instability, not academic incapacity. An algorithm sees only deviations from expected patterns.

The myth of algorithmic neutrality serves powerful interests. It allows institutions to deflect accountability—“the algorithm decided”—while maintaining an aura of scientific objectivity. It positions adoption of these systems as inevitable technological progress rather than contested social choice. And it obscures the reality that every algorithmic system reflects specific values, serves particular interests, and distributes benefits and burdens unevenly.

Education must be the space where we teach the next generation to pierce this illusion, to ask who benefits from algorithmic decisions, and to demand that automated systems serve genuinely public purposes rather than simply optimizing for measurable outcomes.

Building Capacity for Responsible Governance

If we accept that algorithmic societies are our present reality rather than distant future, the urgent question becomes: how do we build the institutional capacity to govern these systems responsibly?

This requires multilevel intervention, starting with education. At the classroom level, we need curricula that help students understand how algorithms shape their information environment, recognize patterns of algorithmic bias, and practice questioning automated recommendations. This isn’t just computer science education—it’s fundamental civic literacy for the 21st century.

At the institutional level, schools and universities need clear governance frameworks for AI deployment. This means mandatory impact assessments before adopting new algorithmic systems, transparent documentation of how automated decisions are made, and meaningful mechanisms for human review and appeal. Several recent frameworks emphasize the importance of regular audits by diverse committees and built-in accessibility features to ensure equitable access.

Critically, this also means investing in educator capacity. Teachers can’t guide students through algorithmic literacy if they themselves lack understanding of how these systems work. Professional development must include not just how to use AI tools, but how to evaluate them critically, recognize their limitations, and teach students to do the same.

At the policy level, we need regulatory frameworks that treat educational algorithms as what they are: governance mechanisms that require democratic accountability. This means standards for algorithmic transparency, requirements for bias testing before deployment, and clear liability when automated systems cause harm. It means resisting the narrative that regulation stifles innovation and insisting instead that responsible governance enables sustainable progress.

But perhaps most importantly, we need cultural change. We must move from treating algorithmic adoption as inevitable technological destiny to recognizing it as social choice that deserves collective deliberation. Not every educational process should be automated. Not every decision benefits from algorithmic optimization. Sometimes the inefficiency of human judgment—its capacity for context, nuance, and ethical reasoning—is exactly what we need.

The Path Forward: Education as Meta-Solution

As I reflect on the rise of algorithmic societies, I return to a conviction that has guided my work for years: education is humanity’s primary adaptive mechanism. It’s how we’ve always responded to transformative change—not just by transmitting existing knowledge, but by cultivating the capacities needed to navigate new conditions.

Algorithmic societies present perhaps the most complex adaptive challenge we’ve faced. These systems operate at speeds and scales beyond human comprehension, yet profoundly shape opportunities and outcomes. They promise efficiency and personalization while risking new forms of inequality and control. They’re simultaneously tools we use and forces that act upon us.

Education must be the space where we develop collective capacity to govern these systems wisely. This means teaching algorithmic literacy as foundational civic competence. It means preparing educators to be critical evaluators rather than passive adopters of educational technology. It means insisting that schools and universities model responsible algorithmic governance through transparent, accountable practices.

But it also means something deeper: preserving and cultivating the distinctly human capacities that algorithmic systems can’t replicate. The ability to recognize context that doesn’t fit patterns. The willingness to question optimization that serves narrow metrics over human flourishing. The ethical reasoning that knows some decisions shouldn’t be delegated to machines, regardless of their accuracy.

The rise of algorithmic societies isn’t predetermined. Neither are its implications for education or its consequences for equity and democracy. These are choices we’re making—or failing to make—right now. The question isn’t whether we’ll live in algorithmically mediated worlds. We already do. The question is whether we’ll build the capacity to govern these systems according to genuinely democratic values, or whether we’ll allow them to govern us according to logics we never consented to and can’t meaningfully contest.

I believe education holds the key to that question. Not because schools can solve every problem, but because they’re where we cultivate human agency in the face of powerful technological and social forces. Every student who learns to question algorithmic recommendations, every teacher who demands transparency from educational technology vendors, every institution that insists on human review of automated decisions helps build the civic infrastructure for responsible algorithmic governance.

This is our work for the years ahead: not resisting algorithmic change, but shaping it toward genuinely public purposes. Education must lead that shaping, because it always has been—and must continue to be—humanity’s way of preparing for futures we can barely imagine but must learn to inhabit with wisdom, equity, and care.

References

Bozkurt, A. (2023). Generative artificial intelligence (AI) powered conversational educational agents: The inevitable paradigm shift. Asian Journal of Distance Education, 18(1), i–xiv.

Wang, Y. (2024). Algorithmic decisions in education governance: Implications and challenges. Discover Education, 3, 229. https://doi.org/10.1007/s44217-024-00337-x

What many considered unlikely—AI tutoring that runs entirely offline, in ordinary school labs—now functions in practice. I have pursued this question because contemporary educational technology often conflates intelligence with connectivity, even as large populations in the global South and underserved regions of the global North experience intermittent power, limited bandwidth, and persistent cost barriers. During the pandemic, UNESCO reported striking gaps in home internet and device access; those figures crystallized a simple design mandate for me: if pedagogy depends on infrastructure that learners do not have, the pedagogy fails them. The header image—a globe held in human hands—signals the inversion I sought: technology should be made to reach the learner, not the other way around.

Scope and contribution

This field report summarizes a school-based deployment of on-device, open-source AI tutoring in rural Telangana, India, designed under the SHIELD framework—Secure Handling of Information in Educational Devices. The study shows that a locally hosted model can scaffold introductory programming without sending data to external servers, while protecting privacy and keeping teachers central to instructional decision-making. The work accompanies a peer-reviewed paper presented at ICALT 2025. This project was possible only through the collaboration of the school’s headmaster, teaching staff, and—above all—the students who participated; their commitment shaped the research from design through reflection.

Why offline AI?

I began from a consequential equity problem. If educational AI presumes persistent connectivity, it systematically excludes schools where infrastructure is fragile or costly. Cloud services remain valuable, yet they encode a dependency model—recurring fees, vendor control, and data externalization—that many public schools cannot sustain. I asked whether a carefully constrained, curriculum-aligned tutor could operate entirely offline—no external calls, no telemetry—while still delivering formative scaffolding that honors local pedagogy and pacing. The question is less about novelty than fitness for context: can AI be designed to meet learners where they are and extend the teacher’s reach without substituting for it?

Introducing: The SHIELD Framework

SHIELD is an acronym for Secure Handling of Information in EducationaL Devices. It is a lightweight peda-governance and design framework that anchors development in three commitments: instructional integrity, privacy by design, and educator oversight. Instructional integrity is implemented through a structured system prompting techniques, such as the Context, Role, Expectations (CRE) technique, which is embedded at the system level. In practice, the tutor frames each interaction in relation to stated learning objectives and goals, adopts a supportive pedagogical role, and limits its responses to hints, worked examples, and explanations rather than direct answers. Privacy by design is realized by keeping all inference and logging on local machines; student work never leaves the classroom. Human oversight is maintained through teacher mediation: the teacher directs use, interprets feedback with students, and retains authority over assessment and pacing. SHIELD therefore operates at the intersection of product choices and classroom protocol; it is as much an implementation discipline as a pedagogical stance.

Implementation

The deployment took place in a rural government high school in Telangana State. Nineteen students, aged 15–17, participated in weekly one-hour sessions over three months (October–December 2024). The lab consisted of twenty mid-range desktops configured to run an open-source large language model locally via LM Studio (LM Studio). After benchmarking several candidates on code-generation and repair tasks inspired by HumanEval, I selected a Llama-family model for its balance of code competence and resource efficiency; the model was quantized so it could run on CPUs where discrete GPUs were unavailable (Llama; HumanEval). A minimal teacher console summarized session activity for supervision within the local network. The instructional sequence aligned with the state curriculum for introductory Python (variables, control flow, lists, simple functions, debugging), and the tutor was configured with prompt templates that reinforced classroom norms, including academic integrity and productive struggle.

Methods and Metrics

The study used a pragmatic field-deployment design with mixed methods to capture changes in learner experience and interaction behavior. Before and after the deployment, students completed short surveys measuring perceived programming self-efficacy, learner autonomy, and usefulness of the offline tutor using five-point Likert items. Students also wrote brief reflections on how the tool helped or hindered their learning. System logs, stored locally and anonymized for analysis, recorded categories of requests and the evolution of query types during the term. Descriptive statistics summarize survey shifts, while thematic coding of reflections and log categories traces movement from surface-level requests to more strategic problem-solving. The goal was not causal estimation but educationally meaningful insight into feasibility, classroom fit, and learner experience under offline conditions.

Findings

Perceptions of capability and autonomy increased over the study window. Baseline means were already favorable (approximately 4.0–4.2 on a five-point scale) and rose to above 4.5 post-deployment. Students commonly described the tutor as a “coding buddy” or “study partner,” language that signals a shift from dependence on teacher availability to self-initiated help-seeking. Open-ended reflections highlighted a reduction in unproductive waiting and an increase in perseverance when encountering syntax or logic errors. Interaction traces corroborated this trajectory: early sessions were dominated by syntax formulation questions; mid-study interactions concentrated on debugging and error interpretation; later sessions emphasized algorithmic refinement and code clarity. Teachers reported fewer interruptions for minor issues and more time for conceptual discussion, project guidance, and formative assessment conferences. From a feasibility perspective, the system operated reliably without internet connectivity, incurred no recurring license costs, and required only modest maintenance once configured—parameters that matter in budget-constrained public schools.

What this does—and does not—claim

The study demonstrates viability rather than finality. It establishes that on-device AI tutoring can be aligned with curriculum, monitored by teachers, and delivered with local data stewardship in a real school. It does not claim superiority to particular cloud services, nor does it report standardized achievement gains. The focus is on removing the structural precondition of connectivity and observing whether pedagogically useful scaffolding remains. Under SHIELD, it does.

Equity and the global distribution of infrastructure

Designing “offline-first” is a strategy for equity, not a concession to scarcity. When the pedagogical core functions without the network, bandwidth becomes an enhancer rather than a gatekeeper. This orientation matters in the global South, where schools face recurring costs and infrastructural volatility, and in the global North, where rural districts, disaster-affected communities, and privacy-sensitive institutions require local control of data and uptime. In both contexts, an offline-capable tutor widens participation by decoupling learning support from network availability while preserving the teacher’s role as orchestrator of activity and assessor of understanding. The work therefore argues for a reframing of the field’s default assumptions: connectivity is valuable, but it should not be the price of admission to intelligent support.

Risks and safeguards

Any AI tutor—offline or not—carries risks. Models can generate incorrect or misleading code. Students may over-rely on automated hints rather than developing independent debugging skills. And tutors may encode biases present in training data. SHIELD addresses these through several measures: (1) system prompting limits outputs to hints and explanations, never full solutions, to preserve productive struggle; (2) teacher mediation ensures that the tutor functions as one resource among many, not the arbiter of correctness; (3) logging provides visibility into student-tutor interactions, enabling teacher oversight; and (4) local hosting eliminates vendor tracking and gives institutions direct control over the system’s behavior and updates. These are not perfect solutions, but they establish a governance posture where humans remain accountable.

Implications for research and policy

The study’s implications extend beyond one rural school. First, it suggests that research on educational AI should expand evaluation criteria beyond learning gains to include feasibility, equity of access, and alignment with teacher autonomy. Second, it demonstrates that open-source models, properly configured, can deliver curriculum-aligned support at scale without dependency on external infrastructure. Third, it calls for policy frameworks that prioritize local data stewardship and on-device inference as equity strategies, not merely technical preferences. Educational AI policy currently assumes connectivity; this work shows that assumption is neither necessary nor neutral. Policymakers and funders should incentivize infrastructure-independent designs as a mechanism for widening participation and reducing structural inequity.

Future directions

SHIELD is a starting point, not a destination. Future work should investigate longer deployments, additional subjects beyond programming, and comparative studies across connectivity contexts. We also need rigorous study of how teachers integrate offline AI into formative assessment cycles, how students develop metacognitive strategies when tutoring is always available, and how local hosting affects maintenance burden and cost sustainability at district scale. Another critical direction is participatory design with educators and students in under-resourced settings, ensuring that offline AI reflects their priorities, not Silicon Valley’s assumptions about what counts as “smart.” Finally, as models improve and hardware becomes cheaper, we must continually ask whether offline-first designs remain necessary—or whether cloud dependency will again be normalized as infrastructure expands. The answer will vary by context, which is precisely why research must attend to equity as a design principle, not an afterthought.

Conclusion

This work shows that AI tutoring does not require the internet. It requires thoughtful design, local stewardship, and respect for teachers as decision-makers. SHIELD offers one model for how that might be done, tested in a real school with real students under real constraints. The results are promising: students gained confidence, teachers retained agency, and the system operated reliably without connectivity. The larger question remains open: will the field treat offline-first AI as a niche solution for “the disconnected,” or as a core equity strategy that expands what counts as infrastructure? The answer will shape who gets to participate in AI-mediated learning—and who is systematically excluded.

Acknowledgments

This research was conducted with the full cooperation of the school administration, teaching staff, and students in rural Telangana. I am deeply grateful for their openness, curiosity, and willingness to experiment with new approaches to computing education. Special thanks to the headmaster for facilitating access and to the teachers who integrated the SHIELD tutor into their instructional practice. This work would not have been possible without their partnership. I also thank my doctoral committee—Ivon Arroyo, Beverly Woolf, and Elizabeth McEneaney—for their guidance and support throughout this project.

As one of the speakers at MassCUE 2025, I shared research and experiences from my work on AI-driven learning environments—focusing on how we can help educators create more engaging, equitable, and effective experiences for students. Across sessions and hallway conversations, one theme kept resurfacing: the challenge of preserving student curiosity within increasingly AI-mediated learning systems.

While many presenters celebrated efficiency gains and personalized pathways, far fewer addressed a deeper tension—how do we design AI systems that kindle rather than extinguish the human capacity for wonder?

The Phenomenology of Wonder

Wonder, as philosophers from Aristotle to Heidegger have long argued, marks the beginning of genuine inquiry. It is the felt sense of encountering something that exceeds our current understanding—a productive disorientation that compels movement toward new knowledge.

Educational psychologist Alison Gopnik (2010) describes this state as “the best kind of thinking”—open, exploratory, and unconstrained by predetermined outcomes. Neuroscientific research confirms that curiosity activates the brain’s reward centers, releasing dopamine, which both motivates exploration and enhances memory retention.

Yet wonder is more than a neural event. As philosopher Karen Barad (2007) notes, it involves “intra-action”—a dynamic exchange in which the learner and the object of curiosity co-constitute each other through engagement. The learner is changed by the act of wondering; the world, in turn, reveals itself differently to the curious mind.

Educational research consistently affirms this: meaningful learning rarely arises from the quick completion of tasks but rather from sustained engagement with uncertainty, tangents, and self-directed questions.

How AI Systems Can Suppress Curiosity

As algorithmic systems increasingly mediate learning, well-intentioned designs often suppress curiosity. Analyses of adaptive platforms and AI writing assistants reveal several recurring patterns.

1. Premature Closure of Inquiry

Adaptive learning tools often respond to uncertainty with immediate intervention—offering hints, simplifying problems, or triggering remediation. While efficient, this approach can truncate the productive struggle essential to deep learning.

Lev Vygotsky (1978) described this space as the “zone of proximal development”—the region where students stretch their understanding just beyond current capability. AI systems optimized for speed often treat confusion as an error to eliminate rather than a moment to explore and resolve. In doing so, they may deprive students of the deeper insights that emerge from grappling with difficulty.

2. Gamification and the Atrophy of Intrinsic Motivation

Gamification strategies—points, badges, and streaks—can keep students “engaged,” but they frequently shift motivation from intrinsic to extrinsic.

Alfie Kohn (1993) demonstrated that external rewards tend to diminish authentic interest. Students begin learning not for discovery but for accumulation of points. Over time, dopamine is triggered not by understanding but by external achievement markers—eroding the natural joy of comprehension.

A student solving math problems for a badge experiences learning differently from one who persists because the problem itself feels fascinating. Gamification risks training students to chase rewards rather than insight.

3. Algorithmic Narrowing of Exploration

Recommendation systems predict what a student might like next based on prior behavior, narrowing exposure to the unfamiliar. But curiosity thrives on serendipity—encountering the unexpected.

A student studying photosynthesis might stumble upon a poem about trees; a young coder might discover algorithmic music composition. These lateral, seemingly unrelated encounters often spark lifelong interests.

By contrast, algorithmic curation filters out the unexpected, gradually shrinking the learner’s intellectual world. Yong Zhao (2021) warns that such narrowing produces not cognitive but environmental learning disabilities—a loss of curiosity born from over-optimization.

Designing AI Systems That Support Wonder

If efficiency-driven systems risk suppressing curiosity, how might we design AI that preserves wonder?

Principle 1: Center Questions Over Answers

Instead of measuring success by accuracy or completion, we can measure it by the depth of questions students generate. Systems might visualize inquiry through question maps or allow “wondering time,” encouraging open-ended exploration.

In this model, the AI acts not as an answer machine but as a conversational partner—posing prompts that deepen thinking rather than close it.

Principle 2: Preserve Productive Struggle

Psychologists Sidney D’Mello and Art Graesser (2012) describe cognitive disequilibrium—a blend of confusion and engagement—as the emotional signature of deep learning.

AI systems should recognize and preserve this state, intervening only when confusion turns to frustration. Metrics should shift from time-to-mastery to depth of engagement—tracking persistence, experimentation, and conceptual understanding.

Principle 3: Enable Serendipitous Discovery

Learning environments should intentionally promote discovery beyond prediction. Borrowing from the design logic of libraries, we can foster “calculated randomness”—introducing content that connects indirectly yet meaningfully to a learner’s interests.

A student studying circuits might see parallels in poetry or logic; one learning about ecosystems might encounter indigenous ecological perspectives. These cross-disciplinary sparks sustain curiosity and intellectual openness.

Principle 4: Make Learning Meaningful, Not Merely Engaging

True engagement arises from meaning, not mechanics. Mihaly Csikszentmihalyi (1990) showed that sustained attention comes when challenge and ability align in intrinsically rewarding ways—a state he called “flow.”

AI can support this by helping students connect learning to their own questions, communities, and aspirations—revealing education not as a checklist but as an evolving dialogue with the world.

Why Wonder Matters: Philosophical Commitments

Wonder as Ethical Stance

Martha Nussbaum (2016) describes wonder as an ethical stance—a willingness to be transformed by encounter. It cultivates humility, empathy, and respect for difference.

AI systems that emphasize rapid certainty risk training learners toward mastery without receptivity. In an age that demands nuance and compassion, the loss of wonder can erode not only curiosity but also civic empathy.

Moving Forward: A Call to Action

As AI becomes more sophisticated, we must decide what kind of learners—and what kind of people—we wish to cultivate. Systems optimized solely for efficiency may produce competent problem-solvers who no longer wonder.

To avoid that fate, we must design AI intentionally for curiosity:

• Value ambiguity over efficiency
• Protect productive confusion
• Encourage exploration beyond the predictable
• Treat time spent wondering as time well spent

Efficiency is not the highest educational virtue. The capacity to wonder—to approach the world with awe, humility, and imagination—remains the foundation of human flourishing.

If we can build AI that amplifies rather than erases that capacity, we will have done more than improve learning systems. We will have preserved what makes learning—and living—most fully human.

References

• Gopnik, A. (2010). How Babies Think: The Science of Childhood. Scientific American. https://alisongopnik.com/Papers_Alison/sciam-Gopnik.pdf
• Barad, K. (2007). Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning. Duke University Press.
• Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Harvard University Press.
• Kohn, A. (1993). Punished by Rewards: The Trouble with Gold Stars, Incentive Plans, A’s, Praise, and Other Bribes. https://www.mv.helsinki.fi/home/hotulain/Punished.pdf.
• Zhao, Y. (2021). Learners Without Borders: New Learning Pathways for All Students. Corwin Press.
• D’Mello, S. K., & Graesser, A. (2012). Dynamics of affective states during complex learning. Learning and Instruction, 22(2), 145–157. https://doi.org/10.1016/j.learninstruc.2011.10.001
• Csikszentmihalyi, M. (1990). Flow: The Psychology of Optimal Experience. Harper & Row.
• Nussbaum, M. C. (2016). Not for Profit: Why Democracy Needs the Humanities. Princeton University Press. http://assets.press.princeton.edu/catalogs/S12Paper.pdf

Education is the master lever because it compounds human capability. When people gain foundational literacy, numeracy, and critical thinking, they make better choices, pursue healthier lives, and contribute productively to their communities. This direct effect multiplies across generations: educated caregivers invest more in children, and educated workers innovate, save, and collaborate.

Education improves health (SDG 3) by increasing understanding of nutrition, hygiene, and prevention. It advances gender equality (SDG 5) by expanding girls’ agency and economic options. It reduces poverty (SDG 1) by raising earning potential and enabling entrepreneurship. It strengthens institutions (SDG 16) because informed citizens can evaluate claims, participate in public life, and demand accountability.

The leverage grows when education is inclusive and high quality. Quality means strong pedagogy, relevant curricula, skilled teachers, and safe, supportive environments. Inclusion means reaching learners across language, disability, geography, and income. Without quality and inclusion, schooling can entrench inequality by privileging those who already have advantages.

AI can amplify the master lever when aligned with equity. Adaptive practice systems can personalize learning without tracking students into narrow paths. Language technologies can translate materials into local languages and produce alternative modalities—text to audio, audio to text, visuals with alt descriptions. Teacher co-pilots can reduce administrative load, surface misconceptions earlier, and suggest targeted supports.

Systems effects arise from feedback loops. Educated people grow local economies, which raise tax bases, which fund better schools, which educate more people. Positive loops also operate through health and civic participation. Yet negative loops persist where access is unequal: communities with under-resourced schools face lower graduation rates and weaker employment prospects, which suppress revenue and investment. The policy task is to flip negative loops and strengthen positive ones.

Three levers consistently pay off. First, early childhood and early-grade mastery establish the foundation for all future learning. Second, teacher capacity—coaching, professional communities, and time for planning—determines classroom quality. Third, pathways that connect schooling to decent work—apprenticeships, micro-credentials, and employer partnerships—translate learning into livelihoods.

Measurement must reflect what matters. Attendance and test scores are partial indicators; systems should track growth, wellbeing, transferable skills, and post-school outcomes. Equity-adjusted metrics ensure we notice and address gaps. Community voice should guide what success means, so indicators remain legitimate and actionable.

When education moves, everything else moves. The payoff is not immediate, but it is durable. By treating SDG 4 as the master lever, aligning AI to serve inclusion, and investing in teacher capacity and early foundations, societies accelerate progress across health, equality, decent work, climate action, and strong institutions. The result is compounding human capability—an engine that advances every other goal.

This visual metaphor matters because as we examine artificial intelligence’s integration into educational systems and social institutions, we must first establish clarity about what “society” actually is. Without this foundational understanding, our discussions of AI governance, educational equity, and technological transformation lack essential grounding. So let me offer a definition that is both precise and usable—one that allows us to evaluate tools, including AI, by their contribution to human flourishing and democratic accountability.

Defining Society: More Than the Sum of Individuals

Society is the enduring pattern through which people organize their collective life. It comprises relationships among individuals and groups; the institutions that channel and constrain behavior; the cultural meanings that guide judgment; and the material systems—land, infrastructure, technology—that make cooperation possible. Society is not merely a collection of individuals. It is a dynamic arrangement of roles, norms, and resources that outlasts any single person while remaining open to revision through public action, innovation, and dissent.

Several elements make society durable. People occupy roles—parent, teacher, nurse, entrepreneur, organizer—that carry expectations and responsibilities. Institutions such as families, schools, firms, unions, courts, and governments coordinate these roles and distribute authority. Norms and laws establish boundaries of acceptable conduct, while narratives—about fairness, achievement, belonging, and purpose—supply reasons to care about those boundaries. Material systems and ecologies, from roads and power grids to forests and rivers, support the circulation of people, knowledge, goods, and care.

Operating Across Scales

In my research, I’ve come to understand that society operates at multiple, interdependent scales. At the micro level are face-to-face interactions where trust, empathy, and conflict are negotiated. At the meso level, organizations translate goals into routines and allocate resources. At the macro level, nations and regions define citizenship, rights, and obligations. Global networks connect these levels through flows of information, finance, culture, and migration. A well-functioning society cultivates coherence across scales: the rules that govern neighborhoods align with constitutional principles, and the incentives that drive organizations do not undermine the social contract.

Power and legitimacy are central to this coherence. Rules that enable participation, protect rights, and ensure procedural fairness build confidence in collective life. When communities experience exclusion, coercion, or persistent inequity, legitimacy erodes, often producing disengagement or unrest. Repair requires institutional learning: the capacity to acknowledge harms, redistribute opportunity, and update rules in ways that expand voice and safeguard dignity.

Society in Transformation

Society also changes, and this is where my work with the Society & AI Lab becomes directly relevant. Demographic shifts, scientific discoveries, technological capabilities, environmental limits, and new social movements reconfigure what people expect of one another. Artificial intelligence belongs to this ongoing reconfiguration. It can widen access to knowledge, coordinate services, and inform public decision-making. It can also concentrate power, encode bias, and obscure accountability if deployed without public oversight and clear standards.

This is why I insist that designing AI for society means aligning technical choices with inclusive goals and measurable benefits—not treating tools as neutral or inevitable. The image at the top reminds us that what fills the structure of society (the green, living network) can either flourish or wither depending on how we steward it. AI systems are now part of that living network.

Evidence and Lived Experience

Evidence matters because decisions shape livelihoods, health, safety, biodiversity, and culture. Yet indicators are only as meaningful as the questions they serve. Data should illuminate lived realities, not replace them. In my scholarship, I emphasize combining quantitative measures—on employment, learning, housing, climate risk, civic participation—with qualitative accounts that foreground experience, context, and history. This integration supports better governance and more honest debate about trade-offs and priorities.

A Working Definition for the AI Age

In plain terms: society is how we live together. It is the rules we agree to follow, the services we build and maintain, the stories we tell about who “we” are, and the commitments we make to people we may never meet. It becomes stronger when institutions are trustworthy, participation is meaningful, and benefits and burdens are shared fairly. It becomes fragile when rules are arbitrary, benefits are hoarded, or voices are ignored.

This definition anchors the broader mission of the Society & AI Lab—connecting social purpose with technical capability. By clarifying what society is and how it changes, we set a standard for evaluating tools, including AI, by their contribution to human flourishing, ecological integrity, and democratic accountability. I invite you—whether you’re a learner, educator, practitioner, or community partner—to use this definition as common ground: a disciplined way to ask what should be built, for whom, with whose knowledge, and under what rules, so that the future remains a shared project, not a private outcome.

The greenery in the human silhouette grows when we tend to it collectively. That is the work before us.

The dominant paradigm in responsible AI development has been human-centered design: building systems that are usable, accessible, and aligned with individual user needs. This approach, articulated most comprehensively by Ben Shneiderman in his framework for Human-Centered AI, emphasizes reliability, safety, and trustworthiness as core design principles (Shneiderman, 2020). Yet as AI systems increasingly mediate collective decisions—in education, healthcare, housing, criminal justice, and democratic participation—individual usability proves insufficient. The question is no longer simply whether a system works for its users, but whether it serves the communities it affects and the publics it shapes.

Society-centered AI begins with a different premise: the people most affected by a system should help define the problem, shape the solution, and govern its use. Rather than treating communities as data sources or end users, this approach recognizes them as co-designers, co-owners, and co-stewards of technology that touches their lives. The goal is not merely better adoption or improved user experience. It is legitimacy: systems that align with local values, distribute benefits fairly, and remain accountable over time.

Defining Society-Centered AI

Society-centered AI refers to the design, deployment, and governance of artificial intelligence in which affected communities share authority over problem framing, data use, evaluation, and adaptation. It extends the principles of human-centered design into collective, civic, and institutional contexts—embedding public participation throughout the AI lifecycle.

This orientation draws on what Sasha Costanza-Chock describes as “design justice”: the principle that design processes should center the voices of those who are directly impacted by the outcomes, particularly communities that have been historically marginalized by technological systems (Costanza-Chock, 2020). Where conventional AI development asks “How can we build this system responsibly?”, society-centered AI asks “Who has the right to decide what ‘responsible’ means, and how do we ensure they are at the table?”

How the Design Cycle Changes

The design cycle changes when communities lead. Problem framing moves from “How can AI optimize X?” to “What is the human purpose, who is served, and what harms must be avoided?” Requirements gathering expands to include lived experience, cultural practices, and constraints such as bandwidth, cost, language, and accessibility.

Data collection becomes reciprocal rather than extractive, with clear consent, transparent uses, and the option to withdraw. Evaluation focuses not just on accuracy but on whether the system strengthens agency, trust, and opportunity. These commitments align with the principles of Value Sensitive Design, which argues that human values must be accounted for throughout the design process—not as afterthoughts, but as primary design requirements (Friedman & Hendry, 2019).

The Five Phases of Society-Centered AI

A society-centered process typically follows five phases:

• Convene a representative coalition that includes community members, domain experts, implementers, and skeptics. Representation matters: those who bear the risks of a system should have voice in its creation.

• Co-define goals, success metrics, and boundaries using structured methods such as problem trees, stakeholder maps, and risk registers. This phase surfaces disagreements early, when they can still shape design rather than derail deployment.

• Co-design prototypes that are testable in low-risk settings, capturing qualitative feedback alongside quantitative metrics. Iteration with community input prevents the calcification of early assumptions.

• Co-govern deployment through participatory policies—eligibility rules, escalation paths, audit rights, and redress mechanisms. Governance is not a one-time approval but an ongoing practice.

• Co-steward learning, with scheduled reviews that can adjust or retire the system when conditions change. Systems that cannot be questioned cannot be trusted.

The illustration below visualizes these five stages as a continuous, participatory cycle—showing how collective intelligence and civic stewardship guide every phase of responsible AI design.

Operational Principles

The operational principles of society-centered AI are pragmatic, not aspirational. Use plain-language documentation and multilingual materials so participation is real, not symbolic. Publish model cards and data statements that describe sources, limitations, and known risks. Implement grievance channels that respond quickly and track resolution.

Share benefits: if community knowledge improves a model, the community should see tangible returns—funding for local programs, capacity building, or shared intellectual property where appropriate. Finally, measure what communities value, not only what engineers can easily count.

Reshaping Incentives for Builders

This approach reshapes incentives for builders. Product teams are rewarded for reducing harm, not just launching features. Roadmaps include time for community consultation, accessibility reviews, and ethics testing. Procurement favors open standards, modular architectures, and portability, so communities can switch providers without losing their data or rights.

Governance shifts from one-time approvals to continuous oversight, with independent audits and public reporting. As Costanza-Chock argues, design processes that exclude affected communities reproduce existing inequalities; processes that include them create possibilities for transformation (Costanza-Chock, 2020).

A Posture, Not Just a Process

Society-centered AI is also a posture. It requires humility about what models can and cannot do. It asks practitioners to treat lived experience as a form of expertise. It emphasizes reversible choices: default to pilots, minimize irreversible commitments, and design for graceful exits.

When stakes are high—education, health, housing, benefits—humans must remain at the controls for consequential decisions. Automation should support human judgment, not supplant it.

The Deeper Promise

The deeper promise of society-centered AI is cultural: technology becomes a site of civic learning. Communities practice deliberation, weigh trade-offs, and exercise stewardship. Institutions learn to share power and update rules. Engineers learn to translate technical detail without obscuring risk.

Over time, this practice can strengthen the social fabric, because systems are not merely deployed; they are co-authored and co-owned. In this sense, society-centered AI complements but extends human-centered design—embedding collective governance into every phase of the AI lifecycle.

Conclusion

Society-centered AI is not a slogan. It is a discipline. It replaces one-way design with reciprocal making and continuous governance. It treats communities as partners in defining the future they must live with—and ensures AI earns its place by serving that future well.

The traditions it draws upon—participatory design, value sensitive design, design justice, and civic technology—remind us that these questions are not new. What is new is the scale, speed, and opacity of AI systems, which make democratic participation both more difficult and more necessary. Society-centered AI is our answer to that challenge: technology accountable to the people it serves.

References

Costanza-Chock, S. (2020). Design justice: Community-led practices to build the worlds we need. MIT Press. https://doi.org/10.7551/mitpress/12255.001.0001

Friedman, B., & Hendry, D. G. (2019). Value sensitive design: Shaping technology with moral imagination. MIT Press. https://doi.org/10.7551/mitpress/7585.001.0001

Shneiderman, B. (2020). Human-centered AI. Oxford University Press.

Cite this article

From Foundational Literacy to Regenerative Competence

Regenerative competence refers to the integrated ability to apply literacy, numeracy, data, digital, and ecological understanding toward actions that restore systems rather than deplete them—balancing human prosperity with planetary health.

Start with universal literacy and numeracy. Reading allows people to access knowledge and critique claims. Numeracy enables estimation, comparison, and proportional reasoning. Together they support financial decisions, health choices, and civic participation. They also underpin more specialized competencies that a regenerative economy demands.

Climate competence adds system thinking. Learners should grasp carbon cycles, energy flows, water systems, and biodiversity as interdependent networks. They should practice scenario thinking under uncertainty and reason about trade-offs among mitigation, adaptation, and equity. These skills anchor local action—retrofitting buildings, managing waste, restoring ecosystems—and guide policy judgment.

Data competence turns observation into insight. People need to collect, clean, visualize, and interpret data ethically. They should understand sampling, bias, and uncertainty; know when a correlation misleads; and recognize when synthetic or proxy data are appropriate. These habits allow communities to audit supply chains, monitor air and water quality, and test interventions.

Digital and AI competence connects tools to values. Learners should know how to prompt, verify, and explain model outputs; recognize when a task is too risky to automate; and protect privacy. They should be able to build simple automations that save time while keeping humans in the loop for consequential decisions.

Pathways, Inclusion, and Measurement for a Regenerative Future

Workforce pathways translate competence into livelihoods. Micro-credentials and apprenticeships can certify skills in energy auditing, heat pump installation, agroecology, circular design, and repair. Community colleges, vocational programs, and universities can align curricula with local transition plans—reskilling workers in carbon-intensive sectors and opening routes for youth into growth areas.

Culture and inclusion determine uptake. Programs must be accessible in multiple languages, inclusive of learners with disabilities, and open to people without traditional credentials. Teaching should use local examples—rivers, farms, buildings, transit—to connect concepts to lived experience. Public libraries, maker spaces, and community labs can become hubs for regenerative projects.

Measurement should focus on capacity to act. Track mastery of core competencies, participation in local projects, job placement, and community-level outcomes such as energy use, waste diversion, and biodiversity indicators. Use open data and transparent methods so communities can learn and adapt together.

A regenerative economy is a skills project: literacy and numeracy as bedrock; climate, data, and AI competence as amplifiers; and inclusive pathways as bridges to decent work. When societies cultivate these skills, they build the capacity to repair, steward, and thrive.

Gross Domestic Product (GDP) has long served as a stand-in for progress. Defined as the total monetary value of goods and services produced within a country over a set period, GDP measures economic output—not human wellbeing. It tells us how much a nation produces, but not whether people live meaningful, connected, or fulfilling lives. A country can grow richer on paper while its citizens grow lonelier, more anxious, or less secure. GDP captures production but ignores purpose.

In the age of automation and AI, this gap is widening. Technologies can boost productivity even as they displace workers, erode trust, and consume attention. A society can be economically efficient yet psychologically fragile. We therefore need new metrics that assess not only how much we create, but how well we live together.

Defining the Flourishing Index

Recent research by Harvard epidemiologist Tyler VanderWeele and colleagues (2025) has advanced a rigorous, multidimensional definition of flourishing that can serve as the foundation for a Flourishing Index. Drawing on global survey data, the framework conceptualizes flourishing as “a state in which all aspects of a person’s life are good,” encompassing physical, mental, social, and spiritual dimensions of wellbeing.

This approach identifies five primary domains of flourishing—(1) happiness and life satisfaction, (2) physical and mental health, (3) meaning and purpose, (4) character and virtue, and (5) close social relationships—along with a sixth, financial and material stability, as a necessary enabling condition. Together, these domains provide a balanced view of human thriving that integrates subjective experiences with objective conditions.

Building on this model, the Flourishing Index proposed here extends the framework to societal and technological contexts, especially those shaped by AI. It reimagines flourishing as both an individual and collective achievement—grounded in equity, sustainability, and governance. The index includes eight adaptable domains: physical and mental health; learning and capability; work and income security; social connection and trust; environmental quality; time and attention; safety and rights; and civic voice. Each domain blends quantitative indicators (e.g., access to care, air quality, education levels) with qualitative measures (e.g., belonging, agency, meaning).

Principles for Design and Use

To be useful, the index must balance scientific rigor with ethical clarity.

• Parsimony ensures a concise set of high-signal indicators rather than an overwhelming catalog.
• Equity requires disaggregation by demographic group and geography to surface disparities.
• Transparency demands open documentation of definitions, sources, and limitations.
• Adaptability allows local communities to include context-specific metrics such as language vitality or cultural participation.

In this sense, the Flourishing Index is not a fixed scorecard but a living instrument—one that communities can evolve as their values and conditions change.

The Role of AI in Measurement

Artificial intelligence can enrich this measurement landscape when guided by privacy, consent, and oversight. Privacy-preserving analytics can detect trends while protecting individuals. Natural language processing can synthesize open-ended community feedback to reveal emerging issues. Satellite and sensor data can track environmental and infrastructural conditions in near real time. Yet the index must never depend on intrusive monitoring. Consent, minimization, and public governance are non-negotiable.

AI’s role is to illuminate patterns of wellbeing—not to surveil or predict individuals. In this way, technology becomes a lens for understanding flourishing rather than a mechanism for control.

Aligning Decisions Around Flourishing

Once established, the Flourishing Index can guide decision-making across institutions.

• Cities can evaluate infrastructure or AI deployments by their impact on wellbeing, not just cost efficiency.
• Schools and universities can assess both academic growth and social-emotional development.
• Health systems can integrate social and environmental determinants of health.
• Employers can measure job quality and fulfillment rather than only productivity.
• Governments can weigh policy trade-offs in terms of long-term human and ecological flourishing.

When flourishing becomes the policy benchmark, governance aligns with what people actually experience as a good life.

Governance and Accountability

Legitimacy depends on stewardship. Independent boards with representation from communities, scientists, ethicists, and policymakers should oversee the index. Open data portals and reproducible analytic pipelines enable public verification. Regular updates should reflect changing social and environmental realities.

A truly flourishing society requires measurement with meaning—transparent, participatory, and adaptable to technological change.

From Growth to Flourishing

Flourishing is not an abstract aspiration; it is a measurable and improvable state of life. By adopting an index grounded in the empirical work of VanderWeele et al. and enhanced through responsible use of AI, societies can transition from measuring production to measuring purpose.

The Flourishing Index redefines progress for the AI era: rigorous in design, inclusive in scope, and human in intent.

Reference:
VanderWeele, T.J., Johnson, B.R., Bialowolski, P.T. et al. The Global Flourishing Study: Study Profile and Initial Results on Flourishing. Nature Mental Health, 3, 636–653 (2025). https://doi.org/10.1038/s44220-025-00423-5