The question is not whether to ban or allow AI in schools. Banning is illusory, since students have access to it outside the walls and detection remains unreliable. Allowing it without a framework is irresponsible, because some uses undermine precisely what school is meant to build. The real question is more subtle: which uses build, and which short-circuit?
We tackled the question from a family angle in our article aimed at parents. Here is the school side, which raises different problems.
The principle that settles it
A single criterion, drawn from the work we examined on cognitive delegation, lets you classify almost every use.
A school exercise has two functions: producing a result, and transforming the student who produces it. The result interests no one. An essay on a novel brings nothing to the world; what matters is what happens in the head of the person writing it.
So: a use that produces the result while removing the transformation is harmful. A use that accompanies the transformation is useful. This criterion is more operational than any list of approved tools.
What works
The infinitely patient tutor. This is the use with the clearest benefit. A student who hasn't understood can ask again ten times, without fearing the class's judgment or the adult's weariness. For a shy student, one in difficulty, or one whose parents can't help with homework, this reduces a very real inequality.
The phrasing that changes everything, and which must be taught explicitly: asking to be quizzed rather than informed. "Ask me three questions to check I've understood" makes the student work instead of making them read.
Teacher-produced materials. Exercise variants, texts adapted to different levels, detailed answer keys, concrete scenarios to illustrate a concept. Here, AI saves considerable time on tasks that teach the teacher nothing, and gives back time for what the machine can't do.
Adaptation to special needs. Rephrasing a prompt for a dyslexic student, producing a simplified version, transcribing a lesson. This is the direct extension of what we described yesterday on accessibility.
Learning doubt. This is perhaps the most formative use, and the most neglected. Having an AI produce an answer containing an error, then asking the class to find it, teaches critical thinking better than any lecture. A student who discovers that a machine asserts a falsehood with perfect confidence retains a transferable lesson.
Having the assignment produced. Obvious, but it needs saying: the student gets a grade and loses the learning. The problem isn't cheating, it's that the contract is emptied of its meaning.
Automatic cheating detection. Tools meant to spot generated text produce false positives, and penalise in particular students who write in a highly structured way or in a language that isn't their own. Accusing an innocent student costs more than letting a cheater through.
Replacing the teacher. No serious data supports the idea that a tool replaces the pedagogical relationship. What we know about the role of the relationship in learning points in exactly the opposite direction.
The real task: changing the exercises
The most robust response is neither technical nor disciplinary. If an assignment can be entirely produced by a machine without it showing, the problem may also lie with the assignment itself.
Several approaches hold up naturally. Oral work, which checks real understanding and which no tool can do in the student's place. In-class work, which puts the production back under the teacher's eyes. Topics grounded in lived experience, which no corpus contains. And above all, the exercise of critiquing an AI production rather than producing one: it demands more mastery than the original exercise.
This last avenue is interesting because it turns the threat into a pedagogical tool. Correcting a generated essay requires knowing what a good essay is.
What to take away
School has been through the calculator, the online encyclopaedia, then the smartphone. Each time, outright banning failed and so did naive adoption. What worked was redefining what is assessed.
The difference this time is that the tool doesn't replace a technical gesture but the effort of thinking itself. That justifies more vigilance, not necessarily more prohibitions. And it brings back to the centre a question school has always asked: what do we actually want students to be able to do on their own?