A head who bans, a teacher who looks away, and a student ahead of both

Banning does not work and permitting everything does not work, and the evidence satisfies nobody. What do we actually know about AI in education?

AA Abdelilah Arahal
5 min read Updated 20 September 2026

Almost every school holds three positions at once: an administration that officially bans it, a teacher who knows students use it and looks away, and a student two years ahead of both.

That situation is worse than any clear policy, whatever the policy says.

Why bans fail

Not because banning is ethically wrong, but because it is unenforceable, for three reasons.

Detection is unreliable. Tools that detect generated text produce false positives at worrying rates, and err more often with people writing in a second language. Accusing a student on that basis is an injustice waiting to happen.

Use is invisible. A student who uses a model to understand a concept and then writes it themselves leaves no trace, and should not be prevented anyway.

The gap widens. A ban stops the conscientious student and not the others, so it punishes conscientiousness.

What to measure instead of the product

Instead of measuring Measure
The finished essay The drafts and the changes between them
The correct answer A spoken explanation of it
The completed research The student's critique of two conflicting sources
The solution The question the student asked to reach it
Individual work alone A short defence in front of the class

The second row is the strongest and the cheapest: ask the student to explain their answer aloud. No technology, no policy, no detection tool. Whoever understood can explain, whoever copied cannot, and the difference appears in two minutes.

What we know and what we do not

I will be direct here, because this field is full of confident claims.

Reasonably supported: that personalised explanation helps, that use as a revision and example-generation aid assists learning, and that a teacher using it to prepare materials saves genuine time.

Not yet known: the long term effect on critical thinking, and it cannot be known, because not enough time has passed. Anyone giving you a definitive figure about a ten year effect is selling something.

What worries me personally: not cheating. The removal of difficulty. Learning happens in the moment you are stuck, and a tool that removes that moment instantly may remove the learning with it. That is not a research claim, it is a concern grounded in how learning works.

What a school can actually do

A one page policy, not fifteen. Three clear levels: forbidden in these tasks, permitted with disclosure in those, required in these.

Teach disclosure rather than concealment. Ask students to write at the end of their work: what they used, in which part, and what they changed themselves. That builds a professional habit they will need later.

Train teachers first. A teacher who has not used the tool cannot design an assignment that resists it. That is the largest gap I see.

Prompt
I teach [subject] to students at [level]. Here is an assignment I currently set: [describe it] Redesign it to measure understanding rather than the final product: 1. Suggest three alternative formats that are hard to complete by pasting a model's output. 2. For each, state exactly what it measures and how I can mark it quickly. 3. Suggest one short spoken question that reveals whether the student understands what they submitted. 4. Write a sentence I can put on the assignment explaining to students what is and is not permitted. Bear in mind that my marking time is limited.

In closing

A student who uses the model to understand learns more than before. A student who uses it to avoid understanding learns less than before. The same tool in both cases.

The whole difference is in the design of the task, and that sits with the teacher rather than with any policy.

Start today with one assignment: add a two minute spoken question to it. That small change reveals more than any detection tool you could buy.

And if your school is writing its policy now, that is what we design together in the train the trainer course.

Common questions

Can AI-generated text be detected?
Not reliably enough to justify a disciplinary decision. Detection tools produce false positives at worrying rates and err more often with second language writers.
What is the alternative to banning?
Changing what is measured. Ask for drafts, a spoken explanation, and a critique of two conflicting sources rather than the final product alone. A short spoken question beats any detection tool and costs nothing.
Does AI harm thinking skills?
We do not definitively know, and not enough time has passed for an answer. The reasonable concern is the removal of difficulty, because learning largely happens when you are stuck.
What should a school do first?
Train teachers before writing policy. A teacher who has not used the tool cannot design an assignment that resists it, and that is the largest practical gap in schools today.
How do I teach responsible use?
Require disclosure rather than concealment: students write at the end what they used, in which part, and what they changed themselves. That is a professional habit they will need later.
Share

No comments yet

Leave a comment

Never published. Used only if I reply to you directly.

Comments are read before they appear.

Related reading

12 min read

The best framework for building AI agents in 2026

Seventeen agent frameworks, ranked by what survives production rather than by GitHub stars. Some of the most popular names sit in the bottom tier, and one of them is a security incident waiting to happen.

Start with a short call

Fifteen minutes to understand what your team does and what you want to change. If training is not the right answer, I will say so.