How to talk to a student you suspect used AI on an assignment

You've got a submission on your desk that reads too cleanly. Perfectly structured, confident vocabulary, no visible signs of struggle. Your instinct says something's off. What you do next matters more than most guidance acknowledges - both for the student and for your own position.

The first mistake is framing the conversation as an investigation. Starting with 'Did you use ChatGPT for this?' gives the student two options: confess or deny. Neither is particularly useful. A denial locks both of you into an adversarial position before you have any real evidence. A confession still doesn't tell you how much AI was involved, or whether it was deliberate misuse or naive experimentation.

There's also a practical risk. Formal academic misconduct proceedings require clear evidence. If you raise a concern without it and the accusation falls apart, the institutional and relational fallout can be significant - especially if the student is a non-native English speaker whose fluency you may have misread.

Before any conversation, make sure you have something concrete to go on. A submission that 'feels' AI-generated is not sufficient grounds for a formal concern, even if your instinct turns out to be right.

Useful signals split into two categories: output signals (things about the text itself) and process signals (things about how it was created). Understanding the difference shapes how you use what you have.

If the submission came through a tool that tracks writing behaviour, you may have access to whether the work was largely typed or pasted, how long the session lasted, whether there were natural pauses and revisions. Tools like Learnaway record the timing and type of events - a burst of typing, a 600-character paste, a gap of 20 minutes - without recording the actual text. A teacher can see that 85% of the essay arrived in a single paste event after two minutes of activity, without ever reading what the student wrote.

Lead with curiosity, not suspicion. The most useful questions are neutral enough that you'd ask them of any student about any piece of work. That framing matters: you're not interrogating, you're checking in.

Good conversation-opener questions:

A student who wrote the work will have a story. They'll remember the hard parts, the choices they made, the thing they looked up. A student who didn't write it often can't reconstruct the thinking behind it - not because they're a bad liar, but because there's no process to recall.

If you have specific process data, raise it without accusation: 'I noticed a large block of text appeared early in the session - can you tell me more about your process at that point?' This gives the student a chance to explain. Maybe they drafted in a separate app and pasted in, which is completely legitimate. The question opens a door; it doesn't close one.

Students who wrote their own work - even with some AI assistance - can usually describe their reasoning in specific terms. 'I wasn't sure about the second paragraph, so I moved the structure around.' 'I looked up what cognitive dissonance actually meant because I wanted to use it correctly.' That kind of memory is hard to fabricate.

What you're listening for is not a perfect process. Everyone cuts corners. You're listening for evidence that the student was present in the work: that they made choices, hit obstacles, and thought about the content rather than just submitting the output of a prompt.

Most of these conversations end one of three ways: the student demonstrates clear engagement and you move on; they acknowledge some AI use that, on reflection, doesn't cross your policy line; or there's a real gap between the submission and what they can explain.

Only the third case warrants escalation to formal procedures, and even then, document what you have first. A timestamped process log from a tool is far stronger evidence than a prose score from a detector. 'The submission contained a single paste of 850 characters at minute three of a six-minute session' is a statement of fact. 'The detector gave it a 91% AI score' is a probabilistic claim with a documented false-positive problem.

The instinct to act fast on AI suspicion is understandable, but most of the harm in these situations comes from acting before there's enough to go on. A well-placed curiosity question almost always tells you more - and leaves you in a stronger position - than jumping straight to accusation.