The group of students using an adaptive AI tutor with pedagogical guardrails achieves scores 127% higher during assisted exercises; the group using standard ChatGPT progresses by 48%. These two figures, from the Bastani et al. study published in PNAS in 2025 and cited by the OECD in its Digital Education Outlook 2026 report, sum up better than a thousand-page report the problem posed by artificial intelligence in the classroom. Because that same student who used the standard chatbot, when placed before an exam without access to the tool, scores 17% lower than one who never opened a chatbot.
This is not a technological question. It is a pedagogical question as old as schooling itself: what does it mean to learn, and who is responsible for it?
The Essentials
- According to the Bastani et al. study (PNAS, 2025), cited by OECD Digital Education Outlook 2026, the group using an adaptive AI tutor with pedagogical guardrails progresses 127% on assisted exercises, while the group using standard GPT-4 (GPT Base) progresses 48% — but the latter regresses by 17% on an unassisted exam.
- The study conducted in Turkey distinguishes two radically different uses: the standard chatbot that provides the answer, and the AI tutor that forces the student to reason before correcting.
- The GPT Tutor group does not register significant gains on the exam, but its results are similar to the control group — meaning the AI tutor avoided the regression observed among standard chatbot users. The pedagogy of the tool, not the tool itself, determines the outcome.
- The divide is less between students with or without access to AI than between students with or without sufficient pedagogical guidance to use it as a thinking tool rather than as a crutch.
A Leveraging Effect That Backfires
The reference study concerns Turkish high school students learning mathematics, divided into three groups: those who work without assistance, those using standard GPT-4 (GPT Base group), those using an AI tutor designed to guide without directly providing the answer (GPT Tutor group). The results, taken up and put into perspective by the OECD in its Digital Education Outlook 2026 report, are clear.
During the training phase, the standard GPT-4 group outpaces the others. AI answers, corrects, proposes variations — exercises are completed faster, with an apparently much higher success rate. A teacher observing the classroom would see students who are focused, efficient, visibly making progress.
At the exam, which takes place in the same 90-minute session as the practice phase, immediately after it, the machine is no longer there. And the standard GPT-4 group drops below the level of the control group — the one that had no tool at all. The leveraging effect backfires. What the AI seemed to be building, it had actually short-circuited.
The mechanism is not mysterious. Learning to solve a problem means traversing the discomfort of not knowing. It means searching, fumbling, failing a first time, understanding why. This is what learning neuroscience calls “desirable difficulty”: the cognitive effort that makes knowledge lasting. When AI eliminates this effort by providing the answer before the student has had time to search, it also eliminates learning.
The AI Tutor Does the Opposite — and So Do the Results
The third group in the study tells a different story. The adaptive AI tutor used in this group is designed according to explicit pedagogical principles: it does not give the answer, it asks questions. It evaluates the student’s level in real time and adapts the difficulty. It encourages, prompts, gives progressive clues. When the student makes a mistake, it invites them to understand why rather than to start over.
On the exam, this group does not register significant gains compared to the control group: its results are similar to theirs, which means the AI tutor avoided the regression observed among standard chatbot users, without producing measurable progress. The difference between the two AI groups is not a difference in technological power — they use similar generation models. It is a difference in pedagogical design.
The OECD draws from this a conclusion that the institution formulates with its usual caution, but which amounts to this: AI in education is neither good nor bad. It amplifies the pedagogy within which it operates. Poorly used, it accelerates dependency. Well used, it can do what no teacher in an overcrowded classroom can do alone: adapt to the exact pace of each student.
This is not a new promise. Intelligent tutors have existed since the 1980s, and some meta-analyses attribute to them effects comparable to those of individual human tutoring — roughly a gain on the order of one standard deviation, which places a median student at the level of the top 20%. What is new is that generative AI allows these tools to be built without the prohibitive costs of previous systems, and deployed on a scale that was previously out of reach.
The Divide Is Not What We Think
Public debate about AI in schools revolves around a simple divide: those with access to tools and those without. This reading is partially true — digital access inequalities persist, even if they are narrowing in most OECD countries.
But the Turkish data points to a deeper divide, and one more difficult to bridge with simple equipment. What differentiates students who benefit from AI from those who suffer from it is not access to the tool. It is the capacity to use it as an instrument of thought rather than as a substitute for thought.
This capacity is not innate. It is taught. It requires sufficient pedagogical guidance for the student to understand what they are seeking, what they do not yet know, and how AI can help them bridge that gap without crossing it on their behalf. This is precisely the kind of guidance that is unequally distributed according to schools, educational tracks, and social backgrounds.
The irony is cruel: students from privileged backgrounds, who already tend to better mobilize educational resources, are also those most likely to be guided toward a reflexive use of AI. Others, left to themselves facing a chatbot without pedagogical framework, use it as a shortcut — not from laziness, but because no one explained to them why difficulty is useful. The gap in social mobility replays itself right there.
We find a similar dynamic in the working world. AI demands that beginners adopt the posture of a senior: knowing how to evaluate the quality of an answer, identifying what one does not know, formulating a precise question. These metacognitive skills are those that school is supposed to build — and they are exactly what poorly used AI in the classroom risks preventing from forming.
What Countries Making Progress Are Doing
The OECD documents several national initiatives attempting to address this problem with concrete means.
In South Korea, the Ministry of Education launched in 2024 a program to deploy adaptive AI tutors in public schools, with specific teacher training on the pedagogical use of data produced by these systems. The idea is not to replace the teacher with the machine, but to give them a precise dashboard on individual gaps for each student — and to let them decide where to intervene when AI has identified a blockage. Initial feedback, still preliminary, suggests improved pedagogical differentiation in pilot classrooms.
In Finland, research teams from the University of Helsinki are working on protocols for “post-AI debriefing”: after a work session with an AI tool, students are invited to explain aloud what they understood, what they would have done differently without the tool, and what they still do not know. This simple practice transforms passive interaction with the chatbot into an exercise in active metacognition.
In the United Kingdom, Khanmigo, the AI tutor developed by Khan Academy based on GPT-4, is being tested in several hundred schools. The tool is explicitly designed to never give a direct answer in mathematics: it asks questions, proposes steps, and invites the student to validate their own reasoning. Current evaluations do not yet allow conclusions about long-term effects, but intermediate data indicate a reduction in giving up on difficult problems — a marker often more predictive of school success than raw scores.
What Technology Alone Cannot Decide
It would be tempting to conclude that the problem is solved as long as we have the right tool — the adaptive tutor rather than the generalist chatbot. This would miss the essential point.
The best-designed AI tutor in the world produces limited effects if the teacher overseeing it has not understood how to integrate it into a coherent pedagogical progression. It produces zero effects if students have no reason to engage with difficulty rather than with shortcuts. It produces perverse effects if the school deploys it without training its teachers and without modifying its assessment practices.
This is where the responsibility of education systems is complete. Not in the choice of tool, but in teacher training, the design of pedagogical sequences, and the definition of what we seek to evaluate. A system that continues to evaluate speed of execution rather than reasoning capacity will produce students who optimize for speed — and AI is unbeatable at that.
The question of school inequality asserts itself here with particular acuity. Schools that have the pedagogical resources to train their teachers in reflexive use of AI are rarely those who would need it most. The digital divide that mattered yesterday was a divide in equipment; the one taking shape today is a divide in adult pedagogical skills. It is a problem of public investment in teacher continuing education — not a problem of silicon.
Some signs point in the right direction. The OECD has recommended since 2024 that AI policies in education must obligatorily integrate a teacher training component at least as well funded as the equipment component. The European Commission has inscribed in its digital plan for education the objective of training 90% of teachers in “critical use of digital tools” by 2030 — an ambitious objective given current levels, but whose inclusion in public policy at least marks recognition of the problem.
Educational science research, long marginalized in debates about school technology, is regaining a central place. Teams like those of John Hattie in Melbourne or Robert Coe in the United Kingdom have for years documented high-impact pedagogical practices, independent of the technology question. Their work converges with OECD data: what makes students progress is precise feedback, exposure to difficulty, and explicit explanation of thinking strategies. AI can serve these objectives or bypass them. The choice belongs to those who design schools, not to those who design AI.
The Exam as Revealer
There is something symbolic about the fact that the rupture manifests itself at the exam. The exam without assistance is the moment when the student is alone with what they have actually learned — not what they seemed to understand thanks to machine assistance. It is the chemical indicator that distinguishes learning from assisted performance.
Some researchers advocate rethinking the exam itself: if AI is everywhere in professional life, why continue to evaluate skills “without AI”? The argument has some coherence. But it assumes that AI will always be available, always reliable, always adapted to the problem to be solved — and that the capacity to reason without it is worthless. This bet is risky, and the Turkish study data suggest it is all the more so because students who learned to reason alone are also those who best use AI when they have access to it.
The real question is not “should AI be allowed on exams?” It is: what type of exam actually measures what a student can do with and without assistance, and how does school prepare them for these two modes of thought? A few education systems are beginning to experiment with hybrid evaluations, with assisted and autonomous phases, to precisely measure the gap between the two. The distance between assisted performance and autonomous performance becomes in itself a pedagogical indicator.
The results of the Turkish study do not say that AI harms students. They say that AI poorly used widens the gap between those who know how to learn and those who do not yet know. This is a distinction that deserves to guide policy choices in the coming years: not toward banning tools, not toward their blind deployment, but toward investment in the pedagogy that determines what we do with them.
The true school divide of the next decade may not be between connected and disconnected students, but between students who learned to think with a machine and students who the machine learned to think for.
Sources
- OECD Digital Education Outlook 2026 — https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_062a7394-en.html
- Primary study: 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.” PNAS, 122(26) — https://www.pnas.org/doi/10.1073/pnas.2422633122
- SSRN preprint – Bastani et al. (2024) — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4895486
- Kulik & Fletcher (2016), meta-analysis on intelligent tutors — https://journals.sagepub.com/doi/10.3102/0034654315581420
- Wikipedia – Intelligent Tutoring System (history) — https://en.wikipedia.org/wiki/Intelligent_tutoring_system
- Khan Academy — Khanmigo, AI tutor based on GPT-4: https://www.khanacademy.org/khanmigo
- European Commission, Digital Education Action Plan 2021-2027: https://education.ec.europa.eu/focus-topics/digital-education/action-plan