How to Tell If a Student Used AI (And Why That's the Wrong Question)

There is a particular sinking feeling that comes with a piece of work that doesn't sound like the student who handed it in. The vocabulary is a size too big. The structure is suspiciously tidy. The voice you've heard in class for six months has gone somewhere else. Your first instinct is entirely reasonable: you want to know whether they wrote it, so you paste it into a detector. It's worth being straight about what happens next. Detectors don't return evidence — they return a confidence score, and a confidence score is not something you can put in front of a student, a parent, or a head of department. They're unreliable in both directions, missing text that was generated and flagging text that wasn't. And the false positives don't land randomly. They land on students writing in a second language, students who have been drilled into formal structures, students whose writing is careful and slightly unusual. The tool that promises to protect academic honesty will, some of the time, accuse the child who worked hardest.

So the honest position is that you cannot tell, not reliably, not from the text alone. That sounds like a defeat, and it isn't — because "did they use AI" was never the question you actually needed answered. Step back to what you wanted when you set the task. You wanted to know whether this student understands this material and can do this thing. AI use only matters because it obscures that. If you can answer the real question another way, the detection question quietly stops mattering.

The most reliable way to answer it is to move some of your evidence back into the room, where you can watch it happen. This doesn't mean banning homework or hand-writing everything under exam conditions. It means making sure that for every important thing you need to know about a student, at least one piece of evidence was produced in front of you: a five-minute written response at the start of the lesson, a paragraph drafted on paper, two questions answered out loud, an exit ticket that asks them to apply what their essay claimed to understand. Small, frequent, low-stakes. You're not building a surveillance system. You're making sure you're never dependent on a single unverifiable document to know where a child actually is.

The second move is to redesign the task so that using AI becomes visible rather than hidden — better still, so that it produces something worth reading. Ask for the process alongside the product: the plan, the first attempt, the thing they changed their mind about. Ask them to critique an AI-written answer instead of producing one — hand them a generated paragraph with a plausible error buried in it and ask what's wrong with it. Ask for a specific link to something that happened in your classroom: the text you read together, the demonstration that went wrong, the example from Tuesday. A model can write about the topic. It cannot write about your lesson.

When you do need to have the conversation, don't open with an accusation you can't support. Open with curiosity and let them do the talking. "Talk me through how you got to this paragraph." "Which part took you longest?" "Explain this sentence to me in your own words." A student who wrote it will answer easily and often at length. A student who didn't will usually tell you so themselves, gently, without a confrontation. That conversation is more accurate than any detector, and — this is the part that matters — it leaves the relationship intact whichever way it goes.

None of this is free. It means keeping small pieces of in-class evidence for each class, and noticing when written work and classroom work stop agreeing with each other. That noticing is exactly the sort of thing that erodes at the end of a long term, when you're carrying five classes and can no longer remember who demonstrated what in front of you in week three. It's also the sort of thing a tool with a memory of your teaching can hold — building the low-stakes checks out of the sequence you're already teaching, and keeping a record of what each class has actually shown you in the room, not only what was handed in. That continuity is what aime is built around. Whether it's aime or a notebook, the principle stands: what a student can do in front of you is worth more than any verdict about a document.

You are not going to win an arms race against a text generator, and you shouldn't spend your evenings trying. Assess what you can see. Ask the questions only someone who sat in your room could answer. Keep the conversation open rather than forensic. The students using AI to avoid thinking will be found out by the work itself, over weeks, in the gap between what they hand in and what they can do — and the ones who weren't will never have to prove their innocence to a confidence score.