What Remains After Closing the Chat

Suppose I am studying for a mathematics exam. I get stuck on an equation, ask an AI for help, and receive a step-by-step explanation. I read it. Each operation makes sense. By the end, the problem is solved and I feel ready to move on.

Then comes the exam. The numbers change and the chat is gone.

Do I know where to start?

At a Turkish high school, nearly a thousand students participated in an experiment with GPT-4. The basic chat interface improved their practice performance. On the unassisted test, their scores were 17% lower than those of students who had practiced without AI (Bastani et al., 2025).

A school judging the tool solely by the practice exercises would have had reason to celebrate an improvement. The independent test showed that this result was not enough to establish what the student could do afterward.

Psychologists have studied this distance between performance and learning for decades. Nicholas Soderstrom and Robert Bjork brought that evidence together in a review published in 2015. Some conditions improve performance on a task without producing comparable gains in what we retain or apply later. The reverse can also occur: practice that feels harder can produce more lasting learning (Soderstrom & Bjork, 2015).

Return to the equation. Following a procedure someone has just explained can be very different from choosing that procedure when we do not yet know how to solve the problem. If I receive the full solution every time I get stuck, I may finish all the exercises without having practiced that choice enough.

The tool can make a difficulty less visible while leaving it unresolved.

Another group used a tutor that offered hints developed with teachers and avoided giving complete solutions. The disadvantage largely disappeared, with no detectable improvement over the no-AI group. The tests were immediate (Bastani et al., 2025).

That finding opens up a concrete possibility: changing how help is provided.

At Harvard, Greg Kestin and colleagues studied an AI tutor in a physics course involving 194 students. The system structured activities, incorporated solutions prepared by instructors, and provided feedback. Scores on the subsequent tests were higher than under in-person active learning, where students also solved problems and received guidance. The experiment, published in Scientific Reports in 2025, covered two lessons with immediate assessments (Kestin et al., 2025). Neither experiment establishes what happened to learning months later.

There are, then, reasons to develop these tools. The findings also show how much is lost when we describe such different experiences simply as “using AI.”

A chat that solves the exercise and a tutor that structures practice can use the same model. My reading is that we should pay attention to what the student does during the interaction. What they attempt before asking for help, how they review an error, and which decisions remain theirs to make.

This is a responsibility shared by those of us who teach and those who design these tools.

When we grade an assignment completed with AI, we observe what a person and a system can produce together. That combination can have real value. Education also seeks to expand what that person will be able to understand and do afterward. Evaluating a tool requires us to observe both.

A school or university could begin with one unit of a course and check what students can explain and solve without assistance. It should then assess them again, after some time has passed or the problem has changed. It would also be useful to evaluate how they work with AI, including the errors they detect and the decisions they can justify while using it.

Counting activated accounts or completed assignments leaves those questions unanswered.

The same demand applies to our assessments. If we assign a report and grade only its final version, we learn little about how it was produced. We can ask students to defend a decision, correct a flawed solution, or solve a variation of the problem. Doing so requires teachers’ time and clear criteria. A policy that only buys licenses will leave that work undone.

It is worth helping a student who gets stuck. A timely explanation can allow them to continue where they might otherwise have given up.

What happens next deserves our attention. If the student understands why an operation works, attempts the next step, and learns to recognize when to use it, that help can expand what they can do. We need to see what happens when the chat closes: a new problem appears, and the student has to decide, independently, where to start.

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), e2422633122. https://doi.org/10.1073/pnas.2422633122

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, 17458. https://doi.org/10.1038/s41598-025-97652-6

Soderstrom, N. C., & Bjork, R. A. (2015). Learning Versus Performance: An Integrative Review. Perspectives on Psychological Science, 10(2), 176–199. https://doi.org/10.1177/1745691615569000