Why a blanket ban does not do what it looks like it does
A ban does not remove use. It removes disclosure. Students still use it, alone, badly, with nobody to tell them when it has confidently misled them — and the ones who suffer most are the ones who trusted it.
A free-for-all is not the answer either. Some tasks are the exercise. Handing them over is like paying someone to do your push-ups, and the students who most need the struggle are the ones most tempted to skip it.
What you need is a rule that changes by task, because the honest question is never "did you use AI". It is "did you learn the thing this task existed to teach".
Put the level on every task
Write three levels on the wall at the start of term and mark one on every piece of work you set.
Level 0 — none. This task is the exercise. Doing it is the point.
Level 1 — assistant. You may use it to explain something you do not understand, to check spelling, or to get unstuck. Not to produce anything you hand in.
Level 2 — collaborator. You may use it to draft. You must show what it produced, what you changed, and why.
This does three things at once. It removes ambiguity, so nobody can claim they did not know. It makes the rule enforceable, because "Level 0" is a specific instruction rather than a mood. And it teaches the real judgement — that the right amount depends entirely on what is being learned.
Make disclosure normal by doing it first
Require one line: what tool, what for, what you changed.
Then model it. Tell the class when you used AI to generate their practice questions. Tell them what you had to fix. That single act does more for the honesty norm in your room than any policy document, because it demonstrates that using it is not shameful and hiding it is what is odd.
Teach verification, not rules
The skill that matters is catching it being wrong. Rules do not teach that. One exercise does.
Pick a question in your subject where you know the answer precisely and the model tends to slip — a multi-step calculation, a date, a local historical detail, a species that only lives in your region. Get the answer in advance, on paper. Read it to the class as though it were authoritative. Let them accept it. Then show them the truth.
The point is not that it was wrong. The point is that it was wrong in exactly the same voice it uses when it is right. A student who has personally been taken in once, in a safe room, treats the tool differently for the rest of their life. Do this in the first fortnight.
The argument they will make, and the answer
"I will have AI at work anyway, so why learn this."
They are partly right, and saying otherwise loses them. The answer is not denial. It is this: the people who get value from these tools are the ones who can tell when the output is wrong, and that judgement only comes from having learned the thing yourself. You cannot supervise work you do not understand. The person who never learned to write cannot tell a good draft from a fluent bad one, and that person is not more employable, they are more replaceable.
The access problem, said plainly
Some students have a paid account and a good phone. Some have a shared handset and patchy data. Some have neither.
Any policy that assumes access builds a two-tier classroom, where the advantage compounds quietly. So: no task should require a student to have their own AI access to complete it. If a task benefits from AI, either provide the output on paper for everyone, or make it optional and ungraded. This constraint is not a compromise. It keeps the assessment about what students know.
Before you move on