Contents7
The debate about AI in schools is presented as a choice between two camps: ban it to protect learning, or teach it because it is already everywhere. An experiment published in June 2025 changes the terms of the question. It shows that the same AI model can harm learning or preserve it, depending on how the tool is designed.
Our view, which we set out below: the useful debate is not “for or against”, but “which tool, designed how, for which task”.
What the Turkish experiment shows
In June 2025, the journal PNAS published the study “Generative AI without guardrails can harm learning: Evidence from high school mathematics”, led mainly by researchers from the University of Pennsylvania. It is a field experiment: nearly a thousand high school students in Turkey did maths exercises with or without access to GPT-4, in two forms.
- “GPT Base” imitated the standard ChatGPT interface.
- “GPT Tutor” used instructions designed to protect learning: guiding the student with hints without giving the answer, relying on the solution to each exercise and the common mistakes supplied by the teachers.
The results, as the paper’s abstract presents them:
on marks during access, for students who had GPT Base
on marks during access, for students who had GPT Tutor
on marks once access was withdrawn, for students who had used GPT Base, compared with those who never had it
With GPT Tutor, this negative effect is largely mitigated: these students do about as well as those who never had the tool, without doing better than them.
The authors explain that, without guardrails, students use AI as a crutch during the exercises, then do less well on their own. A correction was published in August 2025; it concerns only one author’s affiliation and does not affect the results.
Watch out
The results do not automatically carry over to every subject or every age
The experiment concerns maths, in a single high school in Turkey, over four 90-minute sessions (about 15% of the semester’s curriculum).
Position 1: ban it, so that students learn
The first camp starts from a simple principle: we learn by doing. A student who asks a machine for the answer does not make the effort that makes knowledge stick. The Turkish experiment gives this camp a strong argument: the tool improves immediate performance, but not skill, and can even degrade it. Banning AI in class and for homework protects the time for effort.
This camp is right about the mechanism. Its limit is practical: students have AI in their phones, at home as in the playground. A ban in class says nothing about what they do in the evening.
Position 2: teach it, since it is already here
The second camp holds that school must prepare students for the real world, and that this world includes AI. It is better to learn to use it, to check what it produces and to recognise its mistakes than to act as if it did not exist. Students who do not learn to use it at school will use it anyway, but without a method.
This camp is right about how AI is really used. Its limit appears in the Turkish experiment: giving access without a framework is exactly the “GPT Base” condition, the one that left lower marks once the tool was withdrawn.
What UNESCO and the French Ministry of Education set out
Institutions have chosen a framework rather than a ban. In September 2023, UNESCO published its first global guidance on generative AI in education and research, which call for teachers to be trained in the use of AI tools.
is the minimum age this guidance sets for using AI tools in class
of the more than 450 schools and universities surveyed had a formal policy or guidance on the use of generative AI
In France, the Ministry of Education published a framework for the use of AI in education in June 2025. As summarised by the Académie de Rennes (a regional education authority), it allows students to use generative AI tools in class from the 4e (age 13 to 14), within an explicit framework. It treats as fraud the use of AI for an assignment without the teacher’s explicit authorisation and without the student making the work their own. And it judges unreliable the tools that claim to detect a text produced by an AI.
These texts side with both camps at once: AI comes into school, but within a framework, and what is assessed remains the student’s own work.
What we think
We design apps, and this study speaks to us first of all about design. The same technology gave two very different results depending on a single variable: the instructions that frame the tool. GPT Tutor was not a different model; it was the same model, asked to guide rather than to answer, with the solutions prepared by the teachers.
Our position: banning AI in schools is an understandable but fragile response; rolling it out without a framework is a mistake.
Key takeaway
The reasonable path lies between the two
- Distinguish the moments. While a skill is being learned, AI must not do the task in the student’s place. Once the skill is acquired, it can help to go further.
- Choose tools designed for learning, which ask questions, give hints and refuse to hand over the raw answer.
- Assess without the tool, at least some of the time, to find out what the student can really do alone.
What would change our mind: studies showing that open access, without guardrails, has no lasting negative effect in other subjects and other contexts. Or, conversely, that even tools designed as tutors degrade learning in the long run.
What this changes in practice
For a parent, the question to ask is not “is my child using AI?” (the answer is probably yes) but “are they using it to understand, or to hand in the homework?”. Asking them to explain their answer without the tool is a good test. The same attention applies outside homework, to the AI companions that teenagers chat with.
For a company, the Turkish experiment has a direct equivalent: training newcomers. A junior who does all their first tasks with AI can produce quickly without learning the job. The day the tool gets it wrong, they will not see it. We develop this point in our analysis of AI and intelligence. The EU’s AI Act, incidentally, asks companies that use these tools to take “measures to support the development of AI literacy of their staff”, without requiring them to guarantee a specific level: that is the wording of Article 4, as amended in July 2026.
For anyone designing an educational tool, the lesson is clear: the assistant’s default behaviour is a pedagogical decision. Letting AI answer everything is a choice, not neutrality, and it is settled when the tool is designed, not afterwards. The web and mobile app development section gathers what we learn from building our own. And if you are preparing a tool of this kind, our app development offer puts a first version in the hands of real users within a few weeks: enough to observe early on whether they use it to understand or to copy.
Frequently asked questions
Does the study prove that AI makes students worse?
No, not in general. It shows that access without guardrails during learning lowers results once the tool is withdrawn, in maths, in one Turkish high school, over four sessions. It says nothing about other subjects or about supervised uses, and the tutor version largely avoided this effect. It is a serious signal against open access during learning, not proof that every use of AI does harm.
What exactly is an “AI tutor”?
It is an assistant that has been given instructions to guide the student instead of answering in their place: asking questions, giving hints, correcting a step, without handing over the complete solution. In the experiment published in PNAS, the tutor was built on the same model as the base version, with the solution to each exercise and the common mistakes supplied by the teachers. That framing is what limited the negative effect.
Should AI be banned for homework?
A ban is hard to enforce, and the Ministry of Education’s framework judges unreliable the tools meant to detect a text written by an AI. It is often more effective to plan part of the assessments without the tool and to ask the student to explain their reasoning. In France, this framework already treats the use of AI for an assignment without the teacher’s explicit authorisation as fraud.
Sources
- “Generative AI without guardrails can harm learning: Evidence from high school mathematics” (new tab)doi.org
- Its first global guidance on generative AI in education and research (new tab)unesco.org
- The Académie de Rennes (new tab)pedagogie.ac-rennes.fr
- Article 4, as amended in July 2026 (new tab)ai-act-service-desk.ec.europa.eu
“Generative AI without guardrails can harm learning: Evidence from high school mathematics”
Read on doi.org(new tab)Its first global guidance on generative AI in education and research
Read on unesco.org(new tab)Jim · founder of Brain Plus AI
Also in web and mobile apps
Stepping back8 min
Does AI make us dumber? What the studies say
Not all the studies doing the rounds on this subject are worth the same. We rank four of them by strength, from a field experiment to a preprint on 54 people, and say what emerges.
Stepping back8 min
Are AI companions dangerous for teenagers?
An American survey indicates that most teenagers have already talked to an AI companion, but its figures come from a campaigning non-profit and call for caution. The design question, though, is a legitimate one, and Europe has already answered it in part.
Stepping back7 min
AI and disinformation: lying machine or remedy?
One study shows that AI writes more convincing false information than humans do, another that it reduces conspiracy beliefs, but its journal has expressed a reservation. What we can take from them, and what the law requires of those who publish.