The illusion of learning. You think you know more, you know less

Today it's the difference between feeling that you're learning and learning, which is one of those distinctions everybody accepts in the abstract and nobody applies to their own case. The thesis is uncomfortable: generative AI is the best machine ever built for manufacturing the feeling of having understood something, and that feeling guarantees nothing. Does that mean AI is no good for learning? Well, no, and in this text I'm going to have to concede a fair bit of ground to the other side. I say it from the inside: I've read summaries of books I thought I knew and, when I tried to explain them without looking, I got stuck on the second sentence. Let's see what the research says, which on this subject has moved a lot in the last year.

Feeling and learning have been drifting apart since the seventies

Any teacher recognises the scene. A student leaves class convinced they've understood everything; they underline, they nod, they take notes. Three weeks later, in the exam, they can't retrieve a thing. The feeling had lied.

Robert Bjork and Elizabeth Bjork, at the University of California, Los Angeles, have been documenting that gap since the seventies. Their central idea is counterintuitive: the study conditions that feel most effective are, often, the ones that produce the least lasting learning. And the other way round: the ones that feel slow, frustrating and full of mistakes (spacing out study, mixing types of problem, forcing yourself to recall without looking) are the ones that retain the most. They called them desirable difficulties.

In 2013, Robert Bjork, John Dunlosky and Nate Kornell summed up two decades of experiments in the Annual Review of Psychology. One of the most solid findings is that students prefer rereading to testing themselves. Rereading gives a feeling of mastery; self-testing gives a feeling of clumsiness. But on delayed tests, self-testing wins again and again. The judgement we make about our own learning rests on fluency: if something is processed easily, the brain concludes it's already known. And that, very often, is false.

A machine optimised for fluency

This is where AI comes in with an unprecedented capability. Everything it produces is polished to be easy to read. The summary is more fluent than the book. The explanation pitched at your level is more fluent than the textbook. The generated answer is more fluent than the one you'd have built yourself by searching, getting it wrong and going back.

If fluency is the signal we use to decide something has been learned, every interaction with a well-designed assistant pushes that signal upwards. Not because it deceives, but because it does its job well. That's the perverse part: the better the tool is as an explainer, the easier it is to mistake having read a good explanation for having understood.

Add to that the fact that the model doesn't convey its own uncertainty well. The shaky answer and the solid answer sound equally sure of themselves, so the user receives fluent content with no clue as to how far to trust it. I go into that more slowly in the article on suggestibility in front of the machine; here it's enough to note that metacognition, the ability to know what you know, is left without data to calibrate itself against.

A thousand Turkish students and a test with no safety net

The experiment that best illustrates the problem was published in PNAS in 2025. Hamsa Bastani, Osbert Bastani, Alp Sungu and colleagues worked with around a thousand secondary-school maths students in Turkey. Some practised with an interface similar to ChatGPT (they called it GPT Base), others with a tutor configured not to give the answers and to push towards reasoning (GPT Tutor), and others without AI.

During the practice sessions, results improved a lot: 48% with GPT Base and 127% with GPT Tutor. Then came the exam without the tool. The students who'd practised with GPT Base scored 17% lower than those who'd never had it. Those with the guardrailed tutor, on the other hand, ended up practically level with the no-AI group. The default interface, the one most people use, had made learning worse while making performance better.

Along the same lines, though much less solid, is the MIT Media Lab study led by Nataliya Kosmyna, Your Brain on ChatGPT. It's still a preprint (the second version is from December 2025), with 54 participants in the first three sessions and only 18 in the fourth. Its most-cited finding has to do with memory of one's own work: those who wrote with the model struggled to quote sentences from their own essay and felt the text was less theirs. It's a small study and I treat it as a hint, not as proof.

The meta-analysis that fell

Throughout 2025, the most-cited study on the optimistic side was a meta-analysis by Jin Wang and Wenxiang Fan published in Humanities and Social Sciences Communications, from Nature's publisher. It pooled 51 studies and concluded that ChatGPT had a large, positive effect on academic performance. It racked up around half a million views and more than two hundred and sixty citations. On 22 April 2026 the journal retracted it over «discrepancies in the meta-analysis» that undermined the validity of its conclusions. Critical analyses pointed to mislabelled studies, incorrect weightings and the inclusion of a paper that had already been retracted.

The same journal published another meta-analysis in 2026, by Xinning Wu and colleagues, with 35 experimental studies and 4,193 participants, which finds a moderately positive effect (g = 0.670). What it measures, like almost all of this literature, is performance with the tool in front of you or just afterwards. Bastani showed that isn't the hard question. The hard question is what's left when the tool goes away.

The best defence of the artificial tutor

Now the other side, which is strong and deserves more than a courtesy paragraph.

In June 2025, Gregory Kestin, Kelly Miller and colleagues published in Scientific Reports a randomised trial with Harvard physics students. They compared an AI tutor designed on the same pedagogical principles as an active-learning class, one of the good classes, not a lecture. The students with the tutor learned more than twice as much in less time, going by the gain between pre-test and post-test, and they also felt more engaged.

The optimistic side's argument, then, is this: the illusion of learning is a design problem, not a technology problem. Bastani's own experiment proves it, because the guardrailed tutor didn't damage learning. A well-made tutor can build in desirable difficulties: ask before explaining, refuse to give the answer, force retrieval from memory. It can do it with infinite patience and at any hour, something no human teacher can offer thirty students at once. Distrusting AI in education because of the bad interface would be like distrusting books because of self-help books.

It strikes me as a very good argument, and I buy it almost entirely. Almost.

Why the objection doesn't close the case

The nuance is in the word «designed». The Harvard tutor and the Turkish GPT Tutor are tools built on purpose by researchers who wanted people to learn. What the vast majority of people use is something else: a general assistant, optimised to solve whatever you ask in the fewest steps, which is exactly what GPT Base did. Desirable difficulties are friction, and friction reduces usage. There's no clear commercial incentive to put it in by default.

There's also a methodological detail. Kestin and his colleagues measured learning in a test at the end of each lesson; it's a valuable result, but it doesn't say what's left a month later. The effect that worries Bjork is precisely the delayed one.

So my position ends up like this, corrected from the one in the original draft: AI can teach very well if someone designs it to teach and if the student agrees to use it in that mode. Outside those two conditions, which today are the exception, it manufactures illusion.

The exam with no safety net is no longer on the calendar

The school exam used to serve a function few people credited it with: it was a diagnosis for the student themselves. That's where they found out what they really knew and recalibrated their judgement for next time. When the work is prepared with AI, handed in with AI and sometimes examined with AI, that moment of reckoning disappears. The feeling of knowing can inflate for years without running into anything.

The exam doesn't actually disappear. It moves house. It turns up in a meeting where someone asks the question nobody planned for, in a job interview, in an emergency, in a class you have to teach unprepared. And it arrives without warning for someone who didn't know they had to prepare, because their self-perception told them they already knew.

A debt with no due date

There's an expression that helps to think about this: deferred cognitive debt. It's a debt because something has been produced without training the skill that something presupposed. It has no date, because nobody knows when or where it'll be collected. It doesn't show up on any balance sheet as long as the tool is still there. And it grows over time, because the longer it takes to be collected, the harder it is to pay off.

The comparison with money is telling. A loan has a contract, an interest rate, a date and a legal obligation to inform you of the risks. Cognitive debt is taken on without signing anything and without any interface being obliged to explain it to you. I'm not saying it as a proposal for a law, sorry, I'm saying it as a symptom: the problem doesn't yet have a name in any regulation.

The student who hands in good work and the professional who doesn't progress

Take a master's student who writes their final project with heavy AI help. The text is solid, it gets through the plagiarism checker, it gets a good mark. Three years later, in an interview, they're asked to defend the methodological decisions and to place their contribution within the literature. They can't. Not because they've forgotten, but because the work of connecting authors, contradicting them and deciding what to cite happened largely outside their head.

The adult version is quieter. A professional with five years' experience builds AI into her daily work. She delivers more and faster, and her reviews are good. What stops happening to her is learning from the work. The difference between five and ten years in a trade was always in the sediment left by problems solved with your own hands, with hits and misses. If the model solves them and she supervises, the sediment doesn't form. At ten years' seniority, she'll be someone with five years of experience repeated twice.

Neither of them was warned. They were told they were learning faster, and in part it was true, during the task. The feeling of learning is one of the cleanest pleasures of adult life, and nobody is immune to it; I'm not either. The problem starts when you can live for years on the feeling without accumulating any of the learning, and nothing in life points it out to you until the day it was needed.

Definitions

• Desirable difficulties: study conditions that make immediate learning more costly (spacing, interleaving, self-testing) and for that reason produce greater long-term retention. A concept from Robert Bjork.
• Fluency illusion: the tendency to judge as learned the material that is processed easily, whether or not it's retained.
• Metacognition: the ability to assess accurately the state of one's own knowledge, what one knows and what one doesn't.
• Retrieval practice effect: the improvement in retention gained by forcing yourself to recall content without looking at it, as opposed to rereading it.
• Deferred cognitive debt: the accumulation of skills left untrained through systematic delegation, whose cost shows up late and in situations where delegation is no longer possible.

References

• Bjork, R. A. & Bjork, E. L. — Making things hard on yourself, but in a good way. Creating desirable difficulties to enhance learning. In Gernsbacher, M. A. et al. (eds.), Psychology and the Real World (Worth, 2011).
• Bjork, R. A., Dunlosky, J. & Kornell, N. — Self-Regulated Learning. Beliefs, Techniques, and Illusions. Annual Review of Psychology 64, 417-444 (2013).
• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. & Mariman, R. — Generative AI without guardrails can harm learning. Evidence from high school mathematics. PNAS 122(26), e2422633122 (2025). The journal later published a correction (DOI 10.1073/pnas.2518204122).
• Kestin, G., Miller, K., Klales, A., Milbourne, T. & Ponti, G. — AI tutoring outperforms in-class active learning. An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports 15 (2025). DOI 10.1038/s41598-025-97652-6.
• Kosmyna, N. et al. — Your Brain on ChatGPT. Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872 (v2, December 2025). Preprint, not peer-reviewed.
• Wang, J. & Fan, W. — The effect of ChatGPT on students' learning performance, learning perception, and higher-order thinking. Insights from a meta-analysis. Humanities and Social Sciences Communications 12 (2025). Retracted on 22 April 2026.
• Wu, X., Zhu, P., Zhang, J., Yin, M. & Wang, Y. — ChatGPT's impact on student learning outcomes. A meta-analysis of 35 experimental studies. Humanities and Social Sciences Communications 13 (2026).
• Brown, P. C., Roediger, H. L. & McDaniel, M. A. — Make It Stick. The Science of Successful Learning (Belknap Press, 2014).

Further reading

• Sweller, J., van Merriënboer, J. J. G. & Paas, F. — Cognitive Architecture and Instructional Design. 20 Years Later. Educational Psychology Review 31(2), 261-292 (2019). A synthesis of cognitive load theory; it helps explain why removing effort doesn't always help learning.

## You might also like - What happens to your brain when you use ChatGPT, according to MIT? - Why rereading no longer works now that you use AI - Suggestibility in front of the machine - The erosion of one's own judgement - Augmented cognitive laziness

## Elsewhere - Scientific Reports — AI tutoring outperforms in-class active learning (Kestin et al., 2025) - Code and data for the Bastani et al. experiment on GitHub - Plagiarism Today — Study Claiming AI Helps Students Learn Retracted - arXiv — Your Brain on ChatGPT (Kosmyna et al.)