What do I do when a student hands in work an AI wrote?

Start with a conversation about the work, not with a detector score. Ask the student to talk you through two specific choices they made — why this example, why this order — and listen to whether the answers exist. That five-minute conversation is more accurate than any software, and unlike software it is defensible to a parent, a principal, or the student themselves.

Why not to lead with a detector

AI-detection tools produce false positives, and they do not produce them evenly. Writing by students with English as an additional language, by autistic students with formal prose styles, and by anyone who writes in plain short sentences is flagged at higher rates. A confident accusation built on a percentage score is a bad position to be in when a parent asks how the number was produced, and you cannot answer.

Detectors are not useless. They are a reason to look more closely, in the same way an odd smell is a reason to check the fridge. They are not evidence.

A sequence that holds up

One: look for the tells in the work itself. Sources that do not exist. A register that jumps between paragraphs. Correct general statements with nothing specific to your class, your text, or your discussion last Tuesday. Flawless mechanics from a student whose mechanics are not flawless. A conclusion that restates without adding.

Two: ask about the process, not the product. "Where did you find this?" "What did you cut?" "Which part gave you the most trouble?" A student who wrote it answers instantly and often with irritation. A student who did not goes vague.

Three: name what you see, not what you conclude. "These three sources do not appear to exist, and this paragraph does not sound like your other work. Help me understand." That sentence is true, specific, and survives being repeated back to a parent.

Four: separate the act from the child. Most students who do this are not cheats by disposition. They are behind, frightened, or they genuinely did not think a summary counted. The consequence should be redoing the work under conditions you can see, not a label.

Then fix the assignment

If a task can be completed well by a language model that has never met your class, the task was measuring the wrong thing. This is uncomfortable, but it is the most useful thing in this whole article.

Things a model cannot do: connect the reading to something that actually happened in your room, respond to a peer's specific argument from the seminar, use a data set the class generated, show the draft history, or defend the piece out loud for three minutes. Adding any one of these changes the economics of cheating more than any detector.

Ask for the middle, not just the end. An outline, an annotated source list, a rough draft with the changes visible. Process artifacts are tedious to fake and they are where the teaching happens anyway.

Make the oral defense routine, not punitive. If every student explains their work for two minutes, the explanation is not an accusation. It is just how the class works.

Let them use it, and require the declaration. "Which tools did you use, for what, and what did you change?" A student who has to describe their use has to think about it, and a student who lies about it has crossed a much brighter line than a vague honour code.

Say the rule before you need it

Most of these incidents come from genuine ambiguity. "Do not use AI" is not a rule a fourteen-year-old can apply — their spellchecker uses AI, their search results are AI summaries, their phone keyboard predicts the next word. Write the rule at the level of the step: planning, yes; drafting, no; proofreading, yes with a declaration. Put it on the assignment sheet, not in a policy document nobody opens.

The part that is actually your job

A student who hands in generated work has told you something: they do not yet see the difference between producing text and knowing something. That is a teachable gap, and it is the gap that AI literacy is for. The incident is unpleasant and it is also the best teaching opportunity you will get all term.