When students use AI: what to change about the homework

Your students have access to AI, they know how to use it, and some of their homework is already produced by it. Three responses are available and two of them do not work.

What does not work

Detection. AI-detection tools produce false positives on ordinary careful writing and false negatives on lightly edited generated text. Accusing a student who did their own work is a serious harm, and it takes exactly one such accusation to lose a family permanently. Do not build a policy on a tool with that error rate.

A ban. Unenforceable on work done at home, and it moves the whole thing out of view, including away from the students who would have asked you how to use it well.

What does work: change the task

• Ask for the working, and mark it. Generated answers are answers. A task that awards most of its marks for the steps is one AI helps with far less.
• Set homework on the student's own material: their own mock paper, their own error from last week, a problem they wrote. It cannot be generated because it is not the same task for two people.
• Verify in the room. Two minutes at the start of the session: one question about the homework, answered aloud, by a named student. Homework someone else did becomes visible immediately, and no accusation is needed, because the gap speaks for itself.
• Weight what happens in front of you. If the grade rests mainly on supervised work, homework returns to being practice, which is what it was for. Much of the anxiety here comes from grading things nobody supervised.

Teach the honest use

The students who will do best are not the ones who never use AI; they are the ones who use it to check, to get unstuck, and to have something explained a second way, and who then do the thinking themselves. That distinction is teachable and almost nobody is teaching it. Show them what "ask it why my answer is wrong" does, and what "ask it for the answer" costs them in the exam, where it will not be there.

Say it once, at the start

State your rule in the first session, in one sentence, before it is a problem: what AI may be used for, what it may not, and what you will do. A rule announced in advance is a rule; a rule announced after an incident is a punishment, and it will be argued with.

Sources

• UNESCO — Education
• OECD — Education and Skills