This page carries the detail behind our second focus area. For the argument in brief, start with Systems & Complexity.
Understanding Educational Systems: The Core Concepts
To design AI that strengthens rather than destabilizes education, we must first understand what makes educational systems complex. We ground our work in several foundational concepts:
Complex Adaptive Systems
Education is a complex adaptive system characterized by:
Multiple interdependent actors: Students, teachers, families, administrators, policymakers, technology vendors, community organizations, employers—each with their own goals, constraints, and agency. No single entity controls the system; outcomes emerge from millions of daily interactions.
Nonlinear relationships: Doubling funding does not double learning. Reducing class size by five students may have enormous impact in some contexts and negligible impact in others. Small changes can produce disproportionate effects, while massive interventions sometimes accomplish little.
Emergent properties: System-level behaviors arise from local interactions that cannot be predicted by examining components in isolation. School culture, for example, emerges from countless micro-interactions among people; it cannot be “installed” through policy.
Self-organization: Educational systems adapt and reorganize in response to pressures—sometimes in productive ways (teachers collaboratively developing better assessments), sometimes in counterproductive ones (teaching narrowly to high-stakes tests).
Path dependence: History matters. Current possibilities are constrained by past decisions. A district that invested heavily in one technology platform faces switching costs that shape future choices. Teacher expertise developed over decades creates capacity that new initiatives can build on—or that resistance to change can protect.
Feedback Loops: The Engines of System Behavior
Feedback loops are circular causal pathways where actions feed back to influence future actions. They are the engines of system behavior—determining whether educational conditions improve, deteriorate, or stabilize over time.
Reinforcing (positive) feedback loops amplify change. They create virtuous or vicious cycles:
Virtuous example: Teachers receive high-quality professional development → They implement more effective practices → Student learning improves → Teachers feel more efficacious → They invest more energy in improvement → Outcomes improve further.
Vicious example: Schools in under-resourced communities lose experienced teachers → Remaining teachers face higher workloads → Support for struggling students decreases → Outcomes worsen → More families with options leave → Resources decline further.
Balancing (negative) feedback loops resist change and maintain stability:
Example: Standardized testing creates pressure → Schools narrow curriculum to tested content → Short-term scores may rise → But deeper learning suffers → Eventually scores plateau or decline → Pressure increases → The cycle continues without addressing root causes.
AI can dramatically strengthen or weaken these loops. Intelligent tutoring systems that provide timely, actionable feedback can accelerate virtuous learning cycles. But algorithmic sorting of students into tracks can entrench vicious cycles of inequity. Understanding these dynamics is essential before deployment.
Stocks, Flows, and Delays
Educational systems can be understood through stocks (accumulated resources at a point in time) and flows (rates of change):
Stocks: Teacher expertise, student motivation, institutional trust, infrastructure quality, community social capital, curriculum materials, technological capacity.
Flows: Professional learning (adding to teacher expertise), burnout (depleting it); effective feedback (building student motivation), repeated failure (eroding it); transparent communication (strengthening trust), broken promises (destroying it).
Delays between action and consequence create governance challenges. Policy changes take months or years to reach classrooms. Teacher development programs produce measurable impact only after sustained implementation. Assessment redesign takes time to gain legitimacy. When these delays are ignored, systems oscillate between overreaction to early signals and premature abandonment of promising initiatives.
AI can compress some delays—providing real-time data on student understanding, for example. But it can also obscure others, creating the illusion of instant impact when actual learning happens over longer timeframes. Responsible deployment requires accounting for both.
Leverage Points: Where Small Changes Create Large Impact
Not all interventions are created equal. Leverage points are places in a system where small, well-designed changes can produce disproportionate positive impact. Drawing on Donella Meadows’ hierarchy, we distinguish:
Low-leverage interventions (easy to implement, limited systemic impact):
- Adjusting parameters: class sizes, budgets, schedules
- Installing new tools without changing practices
- Mandate compliance without changing incentives
High-leverage interventions (harder to implement, transformative potential):
- Information flows: What data is visible? To whom? How quickly? AI can create dashboards that surface equity patterns, enabling earlier intervention.
- Rules and incentives: What behaviors are rewarded? Assessment design determines what teachers prioritize; procurement policies shape what vendors build.
- System goals: What outcomes do we optimize for? Shifting from test-score maximization to holistic development changes everything downstream.
- Paradigms: The shared mental models that shape what people believe is possible. Shifting from “education as knowledge transmission” to “education as capability development” reorients practice fundamentally.
Our research maps where AI can strengthen high-leverage points—and where it risks entrenching low-leverage thinking at scale.
How AI Transforms Educational System Dynamics
AI does not simply add new capabilities to education. It restructures the system’s fundamental dynamics—often in ways that are subtle, delayed, and consequential.
Accelerating Feedback Loops
AI-powered adaptive learning platforms can shorten the feedback cycle between student effort and understanding. Instead of waiting days for graded assignments, students receive immediate, specific guidance. This can strengthen virtuous learning cycles—if the feedback is pedagogically sound, culturally responsive, and calibrated to actual understanding rather than surface performance.
But rapid feedback can also accelerate vicious cycles. When AI systems misdiagnose student needs, they can send learners down unproductive paths at scale. When algorithms optimize for engagement rather than learning, they can habituate students to shallow interaction. Speed without wisdom amplifies both good and bad pedagogical choices.
Creating New Information Flows
Dashboards that aggregate student data across classrooms, schools, or districts can surface patterns invisible to individual teachers. This visibility can enable responsive intervention—identifying struggling students earlier, revealing curriculum gaps, exposing inequitable resource distribution.
Yet new information flows also redistribute power. When administrators gain real-time visibility into teacher practice, professional autonomy can erode. When algorithms flag “at-risk” students, labels can become self-fulfilling prophecies. Transparency serves equity only when paired with trust, professional judgment, and accountability to those being observed.
Shifting Stocks and Flows
AI tools marketed as reducing teacher workload often shift rather than eliminate labor. Teachers spend less time grading but more time learning new platforms, interpreting algorithmic recommendations, and troubleshooting technical failures. Whether this trade-off improves or depletes the stock of teacher capacity depends on design choices that are rarely made with teacher input.
Similarly, AI tutors may increase students’ access to practice—but if that practice is procedural and decontextualized, the stock of deep understanding may not grow proportionally. Systems thinking demands we ask not just “Does AI increase X?” but “What stocks and flows matter most for learning, and how does AI affect them over time?”
Introducing Delays and Unanticipated Consequences
Technology adoption follows its own timeline, often misaligned with educational rhythms. Schools purchase systems before teachers are trained. Platforms update mid-semester, disrupting established routines. Data privacy concerns emerge only after student information has been shared.
These delays create vulnerability. Early enthusiasm for an AI tool may wane as hidden costs emerge—cognitive load on teachers, disengagement from students, algorithmic biases surfacing over time. Our research tracks these trajectories, documenting how initial promise gives way to reality.
Real-World Applications: Systems Thinking in Practice
Our systems perspective shapes how we approach AI deployment across educational contexts:
Mapping Before Intervening
Before recommending any AI tool, we map the existing system: Who are the actors? What are their goals and constraints? What feedback loops currently operate? What stocks are growing or depleting? What delays obscure cause and effect?
This mapping often reveals that the “problem” AI is meant to solve is actually a symptom of deeper system dynamics. For example, a district may seek an AI grading system to reduce teacher workload—but systems analysis might reveal that workload is driven by class sizes, bureaucratic reporting requirements, and lack of planning time. Automating grading addresses a symptom while leaving root causes intact.
Identifying Unintended Consequences
We model how interventions ripple through the system over multiple time horizons:
Immediate effects (weeks): Initial user reactions, technical functionality, first-order impacts on practice.
Short-term dynamics (months): Adaptation by teachers and students, emergence of workarounds, first signs of unintended consequences.
Medium-term patterns (years): Changes to curriculum emphasis, shifts in student motivation, impacts on equity gaps, effects on teacher retention.
Long-term trajectories (5+ years): Cultural shifts, path dependence from technological lock-in, cumulative impacts on learning outcomes, systemic effects on educational opportunity.
This longitudinal view routinely surfaces trade-offs that pilot studies cannot see. An AI system that raises short-term test scores may undermine deeper learning on a timescale no pilot is long enough to observe. A tool that increases efficiency may erode the relationships on which motivation depends. We regard the short evaluation window as one of the more consequential methodological weaknesses in this field, including in work we admire.
Designing for Adaptation
Because educational systems adapt around interventions, we design for graceful evolution rather than static implementation:
- Modular architectures: Components that can be adopted incrementally rather than all-at-once transformations
- Local configurability: Systems that educators can adapt to their context rather than one-size-fits-all mandates
- Transparent logic: Algorithms whose reasoning can be inspected and challenged rather than black boxes
- Exit strategies: Contracts and data formats that allow graceful disengagement rather than permanent lock-in
These design principles acknowledge that educational systems are alive—they will adapt, resist, and transform any technology introduced. Our job is to design technologies that support productive adaptation rather than forcing brittle compliance.
Centering Equity in System Design
From a systems perspective, inequity is not an unfortunate side effect—it is a stable pattern maintained by reinforcing feedback loops. Addressing it requires changing system structure, not just improving individual components.
Our equity-focused systems work identifies:
Reinforcing loops that concentrate advantage: Well-resourced schools attract experienced teachers → Students receive higher-quality instruction → Outcomes improve → Property values rise → More resources flow in → The cycle continues.
Balancing loops that resist change: Efforts to diversify curricula face resistance from dominant groups → Reforms are diluted → Representation remains limited → Marginalized students continue to feel alienated → Calls for change persist → The cycle continues without transformation.
AI can interrupt these patterns—or entrench them. Procurement policies that require equity audits before adoption are a high-leverage intervention. Algorithms that account for structural barriers rather than blaming students are another. But technology alone cannot overcome system dynamics; it must be paired with shifts in resources, rules, and paradigms.
How Society & AI Addresses This
Our approach to systems and complexity research reflects several commitments:
1. Mapping Over Assumptions
We do not assume we understand a system. We map it—documenting actors, relationships, stocks, flows, and feedback loops through ethnographic observation, network analysis, and participatory modeling with educators and communities. Where our map turns out to be wrong, the correction belongs in the published record, not only in our revised practice.
2. Multi-Scale Analysis
We study systems at multiple levels simultaneously—individual classrooms, whole schools, districts, national policy environments—because dynamics at one scale shape possibilities at others.
3. Longitudinal Tracking
We follow interventions over years, not months, because system effects unfold slowly. Pilot studies capture initial reactions; our research captures adaptation, resistance, and long-term transformation.
4. Participatory System Modeling
We bring stakeholders together to build shared system maps, making invisible dynamics visible. This process itself is an intervention—when people see the system they are part of, they act differently within it.
5. Public Knowledge Commons
We publish system maps, simulation models, and case studies openly so that educators, policymakers, and communities can use them to inform local decisions. Our goal is not to prescribe universal solutions but to build capacity for systems thinking everywhere.
Return to Systems & Complexity, or see the other focus areas: Knowledge & Intelligence, Educational Equity, Human Flourishing.