How can I approach a student if I suspect that they have used AI in ways that aren’t allowed (but I don’t have proof)?

You are completely correct about AI detection software not being reliable (read more about that here), and I would also be very careful with the tell-tale signs that you think you see and that you interpret as AI – your gut is as unreliable as the detection softwares!* 

I would start by considering what it is exactly that you hope to gain from confronting the student, and what you stand to lose by doing it.  

It is possible that the student will admit to using AI in ways that you explicitly did not allow (but check your course policy first – is it really so clear that you can say that you are not allowing a specific use, like using AI to polish language or support editing?). When they admit to it, what happens then? Will you ask them to redo the work, or fail them on the task, or notify an oversight committee? And can you be 100% sure that they did not just admit to it to get out of an uncomfortable confrontation, and now they are carrying consequences for something they did not, in fact, do? 

On the flip side of the coin, what happens if they do not admit to it? If they did not do it, you have now created a situation where they feel wrongly accused, probably wonder if you will always be looking out to catch them cheating even though they never did, so your relationship with that student has suffered. If other students hear about you wrongly accusing someone, that will also influence your relationship with them. And will you actually believe them if they say that they did not do it, or will you also be looking differently at them after this? It is of course also possible that your hunch is right, they did use AI, but they still don’t admit to it. Since you don’t have proof either way, where does it leave both of you now? 

So in a nutshell, I think that you really don’t have anything to gain from confronting the student.  

What you could, however, do, is address the topic more generally with the whole class. Not that you suspect someone is going against your policy, but going over your policy again, ask students what questions they have regarding that policy, if they know what that means in practice. Or, depending on what specifically looked like AI to you, addressing that you have seen a lot of this stye of writing, but that another style would be more appropriate; that you would appreciate if they edit their drafts in certain ways to avoid repetition; ...; meaning addressing specifically what it is that you want to be different in the texts they produce next time, rather than making it about whether AI wrote it or your students did.

*Three special cases that I can think of:  

1) You detect plagiarism. That is academic misconduct and needs to be addressed as such, no matter whether it was the student themselves or AI and the student did not notice.

2) You find made-up references or clear misrepresentation of facts. You still don’t know completely for sure that they come from AI, but that does not matter. Those are obvious problems with the content of the essay, and you can address them as such, for example by asking the student to send you the (non-existent) article, or explain how they found the wrong “facts”. 

3) The student forgot to edit the AI output to the extent that you find things like “Certainly! Here is the response to the essay question using scientific language”. In that case, you might just ask the student to explain why they added that sentence in the context.

My experience is with suspecting AI generated 'peer reviews'. Here there tend to be stylistic markers (repetition, lack of connection to content, strict conformity to standard form, formulaic phrasing, contradictory 'assessments', non sequiturs, and more) that accumulate to give an inescapable sense of inhumanity when the review is read through carefully sentence by sentence.

You cannot make an accusation on the basis of 'gut instinct' or 'impression'. If you want to raise the issue, you have to collect data and present evidence that can be discussed. This means going through the essay very carefully and marking every place that, when taken together, makes you twitch and how it does so, and put together the argument of why this (cumulatively) makes you strongly suspect generative AI, and why this is evidence of a use that is not allowed. The basis of a just system is 'innocent until proven guilty'; the onus is on the accuser to prove guilt, not on the accused to prove their innocence.

Prior knowledge of the students' performance in other tasks should never form part of this 'evidence'; this is simply prejudice. A final grade of B can be obtained by a combination of a C on two exams and an A on the third, or an A on two exams and a D on the third. A student's marks on the various assignments and exams for a module can easily vary by over 30%, without generative AI. The point is that 'unusually' good or bad work may have more to do with the topic, preparation or the conditions immediately surrounding the work, than any rule-breaking.

Then there's the question of what to do. Some possibilities include:

• Discuss the analysed essay and the evidence you have that makes you suspect generative AI use beyond what is considered acceptable with the student individually. The fact that you have spent the time going through their work should help mitigate some of the trust issues: you've shown you cared. So then you have two issues to address: what to do with the current assignment, and what can the students do in future, both of which the student should have a say in. Beyond directly addressing the strong suspicion of 'unacceptable' use of generative AI, a meeting might also address: what is the 'penalty' this time? what learning can they take from the meeting? can you help the student develop an 'action plan' for their future use of generative AI?
• Raise the issue more generally in class without identifying any particular student (and ideally, also drawing examples, especially of generative AI use, from several essays (if necessary 'create' new ones)). This might take the form of a discussion of one or more of: stylistics, of logic, of structure, of the 'shape' of an essay, or other disciplinary relevant topics.
• Think about what you're teaching students about essay writing (and make changes). Are you teaching — and allowing — them to think broadly, to consider — and discard — multiple perspectives, only bringing in those that substantially expand their essay (in whatever direction), or are they being 'taught', encouraged, or otherwise learning that an essay should hold itself to a single point, never developing tangents, and that 'depth' can be achieved through introducing irrelevant or arbitrary aspects? (This may include unnecessary theorising, see Kitching, 2008.)

Russell et al (2026; preprint) compare generative AI and human-written fiction, finding that "AI stories over-explain themes and favor tidy, single-track plots while human stories frame protagonist' choices as more morally ambiguous and have increased temporal complexity" and that "that AI-generated stories cluster in a shared region of narrative space, while human-authored stories exhibit greater diversity." They also discuss differences between different generative AI models, but the overall inhumanity remains.

• Kitching, G. (2008). The trouble with theory: the educational costs of postmodernism. Pennsylvania state university press.
• Russell, J., Rajendhram, R., Pham, C. M., Iyyer, M. and Wieting, J. (2026). StoryScope: Investigating idiosyncrasies in AI fiction. ArXiv preprint https://arxiv.org/abs/2604.03136v4