Two years ago, I handed back a set of essays on The Great Gatsby and paused on one. The prose was clean. Too clean. Words like "emblematic" and "juxtaposition" appeared in sentences that sounded nothing like the student who had asked me three weeks earlier what a thesis statement was. I pulled him aside after class, told him I had concerns about the essay, and watched his face go through confusion, then hurt, then a kind of flat shutdown. He said he wrote it himself. I said I wasn't sure about that. He stopped participating in my class for the rest of the semester. He wasn't wrong to.
That confrontation was one of the worst moments of my teaching career, not because I was necessarily wrong to be suspicious, but because I handled it like a prosecutor instead of a teacher. I had no proof. I had instinct and a vocabulary list. And I damaged a relationship that took months to partially repair. If you are trying to figure out how to tell if a student used AI, I want to start by telling you that story, because the "how to tell" question matters a lot less than what you do with your suspicion once you have it.
Why AI Detection Tools Are Not the Answer
The first thing most teachers reach for is a detection tool. Turnitin has one. GPTZero has one. There are a dozen others. I tried several of them. I regret it.
The fundamental problem with AI cheating detection tools is that they are probabilistic, not definitive. They look at patterns in text and assign a likelihood score. That sounds reasonable until you realize what kinds of writing produce those same patterns: writing from students who learned English as a second language, writing from students who were taught to write in a very structured, formal style, and writing from students who are just genuinely good at clear, precise prose. Turnitin's own research has acknowledged false positive rates that should make any responsible educator pause before using a score as evidence. Flagging a student as a cheater based on a software guess is not fair. It is not even close to fair.
A detection score is not evidence. It is a reason to have a conversation, and only if you handle that conversation carefully.
The tools also create a kind of arms race dynamic that benefits no one. Students learn to prompt AI differently to evade detection. Teachers start treating every clean essay as suspicious. The whole thing corrodes trust in both directions.
What Actually Signals AI Use
After a couple of years of watching this closely, I think the real signs of student AI use are almost never in the text itself. They are in the mismatch between the text and everything else you know about the student.
Here is what I mean. If a student has been turning in choppy, short-sentence paragraphs all semester and suddenly submits something with elegant subordinate clauses and a confident argumentative voice, that gap is worth noticing. Not because clean writing is suspicious, but because that specific student's writing has never looked like that before. Similarly, if an essay references "the broader societal implications" of a theme but never once mentions the specific passage we spent forty minutes on in class last Tuesday, something is off. AI has no memory of your classroom. It cannot reference the moment a student in the back row said something that changed how everyone thought about the text.
Process signs matter too. A student who submits a polished final draft but has no prewriting, no rough notes, no evidence of revision, is showing you something. Not proof, but a pattern. The absence of struggle, in a subject where struggle is normal, is its own kind of signal.
Build Process Into the Assignment, Not Just the Rubric
The most durable solution I have found to the whole detect AI writing students problem is to stop making it possible to skip the process. Require a rough draft submitted at least four days before the final. Require a reflection paragraph where the student explains one decision they made in revision. Do a five-minute verbal check-in where you ask the student to walk you through one paragraph out loud, not to grill them, just as a normal part of conference feedback.
That verbal check-in is the most important tool I have. It is not an interrogation. It is just teaching. "Tell me what you were going for in this opening paragraph" is a question I would ask any student, regardless of suspicion. When a student can answer it fluently and specifically, I learn something. When a student cannot talk about their own writing at all, I also learn something. Either way, I have done my job as a writing teacher.
Verbal check-ins also do something else: they signal to students that writing is a thinking process you take seriously, not just a deliverable you collect and grade. That framing alone changes how some students approach the work.
How to Have the Conversation When You Do Suspect AI
Sometimes, even with all the process scaffolding in place, you will still have a gut feeling that something is off. Here is what I do now, and what I wish I had done with that student two years ago.
I open with curiosity instead of accusation. "Walk me through how you wrote this paragraph" gets you so much further than "Did you use ChatGPT?" The first question invites the student into a conversation about their own thinking. The second question backs them into a corner where the only rational response is to get defensive, whether they cheated or not.
"Walk me through how you wrote this" is not a trap. It is a teaching question. The answer tells you everything you need to know.
If a student used AI and knows they weren't supposed to, this question will usually make that apparent without you ever having to accuse them directly. If they wrote it themselves, you will hear them talk about their own thinking, and you will probably learn something about them as a writer. You cannot lose with that question.
Make Honesty Less Risky Than Hiding
The deeper issue underneath all of this is that students hide AI use because they are scared. They are scared of a zero, scared of a phone call home, scared of being labeled a cheater. So they lie, even when they might have been honest if the stakes felt lower.
One thing that has genuinely helped in my classroom is being explicit about where AI is allowed and where it is not, and making clear that telling me the truth carries no punishment when honesty happens early. If a student comes to me before submission and says "I used AI to help me outline this, is that okay for this assignment?" that is a student practicing exactly the kind of judgment I want to build in them. I am not going to punish that. I am going to have a real conversation about it.
When students know there are sanctioned uses of AI in your class, they are far more likely to disclose unsanctioned ones too, because the blanket shame is gone. They know you are not categorically opposed to the tool. They know you care about the learning, not the performance of never touching it.
A Different Kind of Classroom, With a Different Kind of Visibility
I want to be honest: even with all of this, the detection game for work done outside your classroom walls is genuinely hard. That is part of why I have been interested in platforms like Authentiya, which are built specifically for in-classroom writing work. When students write inside Authentiya, teachers can see every AI interaction the student had during the writing process, not as surveillance, but as a transparent record. There is no guessing. There is no confrontation based on instinct. The information is just there, built into the workflow.
That kind of transparency changes the dynamic entirely. The conversation shifts from "did you use AI?" to "here's what I see you used AI for, let's talk about whether it helped you learn." That is a much better conversation to be having with a teenager than the one I had with that student over his Gatsby essay.
If I could go back, I would not lead with suspicion. I would lead with process, with questions, and with a classroom culture where a student could tell me the truth without dreading what came next. That is not a perfect system. But it is a much more human one.