Trace Pedagogy: Grade the Trace, Not the Guess

Suspicion is spreading faster than proof. At Mexico’s most prestigious tertiary institution, authorities suspected that many students had used AI on the national entrance exam. In a highly publicized case at Brown, a professor noticed that grade distributions looked wrong compared to past semesters. Neither case began with a confession but with statistics that flagged something off in aggregate learning.

If it were not for that historical data, authorities and teachers might be witnessing the rise of a generation of collective intelligence at a rate perhaps never observed in human history. Students who struggle academically would be submitting polished assignments unlike ever before. And this miracle would be gratefully credited to the arrival of artificial intelligence.

But intelligence, unfortunately, is the result of sustained engagement, and humanity has yet to find that magic potion. In Japan, a teacher asked a student who struggled with written English to reproduce the surprisingly accomplished piece they had submitted. The result was what everyone, including the student, expected: no learning whatsoever had taken place.

Why Policing Falls Short

Institutions and teachers are trying to rise to the challenge. At the University of California, San Diego, the Academic Integrity Office has staffed up and acknowledged the unpaid labor problem facing faculty who are being asked to redesign courses and assignments in response to AI. Individually, the same Brown professor highlights how administrations lack policies to streamline the reporting of growing academic misconduct, and some instructors have gotten creative by hiding directives inside assignment instructions to trap careless AI use.

But such cat-and-mouse measures only address the symptoms of a growing global malady. In fact, students need to learn what generative AI is and how to leverage it for their own good. Framing this breakthrough technology as a thinking partner is the real challenge, the fine line that administrators and teachers are walking, and need to overcome in the classroom.

Governments across the globe are aware of these trade-offs and have been drafting guidance accordingly. But policy does not move at semester speed, let alone AI iteration speed. Teachers, students and parents cannot wait for that gap to close.

So I would argue that each instructor has an increasingly pivotal role to play in teaching students how to use AI wisely. And the calibrated response should be informed by comparing grade distributions of current courses against past ones. Yes, the tyranny of averages can be ruthless toward outliers and it means little unless it is measuring the same, or very similar, variables over time.

Still, we cannot discount the value of structural breaks, or time series that no longer follow the same patterns, since they signal statistical anomalies worth investigating. Such investigations have revealed that colleges now have a problem of sorting their best students and that some of these students needed to retake their exams.

But the main question still lingers: how can teachers generate qualitative signals of use on the fly in their courses? And why is such a wealth of data essential to measuring real-time individual learning trajectories and to validating other grades?

Make Learning Visible

My direct recommendation is: make learning visible. Give every student chasing a learning goal the chance to leave a documented trace, open to everyone to see, like or comment on. Yes, it is radical and unsettling for some but so are the AI-fueled learning “miracles” we are observing today.

Curiously, radical transparency through traceable text is also an MIT initiative to make it easier to match AI training data with what it is meant to be used for so that the resulting models turn out better. The better the shared knowledge is documented, the more trusted and useful it becomes.

By doing this, embracing radical transparency and open feedback, instructors let genuine learning traces organically accumulate, showing the student, the cohort and the instructor how intellectual development is being approached and whether formative assessment matches summative results.

Formative assessments in my courses are those with few or no guardrails against external support. Students are allowed to use AI tools as they wish and to make mistakes while using them. But, with guidance, they learn from those mistakes and develop better practices and hopefully stop outsourcing their thinking to a machine.

I call these formative assessments Insights, which are part of the Trace Pedagogy I developed from my research on how ants coordinate autonomous acts through deliberate changes they make in a shared physical environment to indirectly communicate and build complex structures.

Who Thrives and Who Resists

Committed, process-oriented students tend to enjoy working on Insights because they embrace challenges and experience a pace of learning they have never felt before. They look forward to coming to my flipped classroom, where initial learning happens outside class, to demonstrate how deeply they have engaged with assigned and self-selected research material. These students join classes highly motivated to share, discuss and refine what they have learned with their peers.

Importantly, these intentionally designed collaborative, student-led conversations also signal to committed students, the cohort and the instructor whether a gap exists between what was shared before class and what is tested in class. They also allow the cohort to identify and judge for themselves the best practices of students who can expand on what they had shared as Insights before class. In this open environment, fabricated knowledge is short-lived while genuine effort becomes apparent for everyone to acknowledge and benefit from.

Two assigned Insights per week with clear deadlines add pressure to perform. This added pressure, some students share after the course, ends up correcting procrastination habits while instilling a strong work ethic. These students also appreciate the safe, controlled setting in which feedback is given in every class, where they are recognized for their sustained rigor and their ability to work well with their peers. Students come to understand that the capacity to collaborate with others is essential for succeeding in an increasingly competitive entry-level job market.

The opposite is true as well. Students who used to find comfort in working alone and being judged only on outcomes suddenly have nowhere to hide. The pressure to be accountable and to collaborate builds so quickly that some drop the course right away. When I started implementing this non-lecture pedagogy, nearly thirty percent of enrolled students dropped the course within the first two classes, when the guidelines were explained and expectations were set. This figure has since fallen to around ten to fifteen percent.

Of those who remain, another ten to twenty percent drop the course either after the results of the midterm summative assessment or shortly before the final summative assessment, a critical review of a relevant book of their choice, is due. These percentages have been slowly declining, though they vary from semester to semester because they depend on many other covariates, such as (i) the perceived workload of cumulative Insights, which has become less demanding every semester in terms of reading and writing load to adapt to students’ increasingly diminished metacognitive skills, including reasoning critically and independently; (ii) changes in formative assessments driven by advances in AI capabilities; (iii) the ability of a few influential students to sway the cohort and (iv) course coordinators’ support.

Nevertheless, one pattern has remained constant among the more than two hundred students who have learned under this highly collaborative, self-directed method over the last four years. They all say they have never experienced anything like it, for better or worse.

Common complaints from students include discomfort with having their Insights ranked or being public every class. Students also resent not receiving tailored feedback when their Insight is submitted after the deadline. And a small number of students just strategically decide to see the cup as half empty and judge the whole system by whichever part is easiest to attack.

But these experiences, the extreme positive and negative ones, are statistical outliers, yet they can significantly influence overall course satisfaction surveys. Trace Pedagogy focuses on elevating academic outcomes for the median student and is welcomed by learners willing to adapt and grind with or without AI, and to be tested without it. After all, these earnest learners not only understand the importance of developing AI literacy skills and of managing the cognitive risks inherent in AI use but also want to make the most of their education and avoid being mistaken for those looking to game the system.

More importantly, accumulated traces give instructors documented and observable, behavioral evidence of either authentic cognitive investment or repeated attempts to circumvent it. Such traces also allow instructors to give open, constructive feedback to the students who need and want it most so that everyone can learn from the productive and counterproductive practices of their peers.

Exams as Confirmation, Not Verdict

As for summative assessments, the controlled, high-stakes exams, they have become less essential, though still required. Controlled assessments come with well-documented caveats such as measuring test-taking performance and favoring a narrow scope of skills. High-stakes tests also put students from minority and vulnerable groups at a particular disadvantage because of the chronic stressors they experience outside the classroom. Still, the signals provided by summative grades give instructors residual, granular evidence to validate or disprove a hypothesis, and a basis for tracking the gap between formative and summative averages, flagging whether the two have diverged or converged, and in which direction the gap is trending over time.

Usually, the more the averages diverge, the more deliberate and decisive the corrective action must be. For instance, under the assumption that all students are using AI to complete their Insights, instructors must redesign formative assessments to capture real effort or provide clearer, more structured, tailored feedback that guides students in approaching large language models in ways that do not erode their own cognitive development.

Students who internalize learning through guided critical reasoning, regardless of how they instrumentally approach it, perform at or above expectations in summative assessments. However, students who are indifferent to feedback and believe that original thinking means prompting machines, which excel at stringing words together based on probabilities and patterns drawn from training data, to produce human-sounding responses find a wide gap between what they mistake for real effort, which they produce in formative assessments and, to some extent, reveal in class, and what true comprehension demands. Summative assessments expose this gap and show how wide it is.

But the real power of Trace Pedagogy, whose one-hour class typically requires three hours of instructor preparation for a cohort of around twenty students, is how it puts control of research and knowledge sharing in students’ own hands. It trusts and allows them to find the best ways to learn at their own pace with or without AI assistance, using the feedback given to and from their peers in every class and publicly sharing what they have learned before coming to class. In such a documented system, where transparency shapes how students are rewarded, accountable learners not only align more easily with ethical behavior but also rarely experience negative surprises in the form of low grades.

Summative assessments marked anonymously to mitigate marker bias consistently confirm, once names are revealed, that students who act on feedback and incorporate it into their conscious, honest effort to build a cumulative formative portfolio are significantly more likely to receive higher summative grades. For these students, the gap between formative and summative performance is either nonexistent or negligible.

Diligent students, like tireless ants independently modifying the common environment through their steady contributions, end up realizing that the whole is far greater than the sum of its parts. And that they were only able to achieve a sense of rapid, meaningful growth in a relatively short period of time because they were incentivized, rewarded and guided by the instructor and inspired by their peers to keep pushing the envelope from very early in, and throughout, the course.

And when it comes to AI assistance, students come to see their true potential through self-directed learning techniques that do not make them look like their peers, as the reckless use of AI does, but set their individuality, curiosity and promise apart by deepening their authenticity through judicious inquiry.

In the end, the trace is the proof and what gets measured, every class, is what keeps learning honest.

---

Cite this article

Starominski-Uehara, M. (2026). Trace Pedagogy: Grade the trace, not the guess. Society and AI. https://societyandai.org/perspectives/grade-the-trace-not-the-guess/

---

Write for Society & AI

Have a perspective on how assessment should change when AI can polish any final product? We want to hear it. Society & AI publishes in open access, free for anyone thinking seriously about these questions. Send your proposal to [email protected].

---

The two students and the scores in the interactive opening are illustrative, not real student data.