Can you tell if a student used AI to write an essay?

Not reliably from the finished text alone, and not from a detector score. AI writing tools produce clean, human-sounding prose, and detectors both miss edited AI work and falsely flag genuine student writing, especially from multilingual students. What tells you is process evidence: how the document was written over time, whether the student can explain their choices, and whether the work matches what you have seen them do in class.

Are AI detectors accurate for schools?

Not accurately enough to accuse a student. Independent testing has repeatedly shown high false-positive rates, and several major vendors and universities have pulled back from AI detection because it wrongly flags honest work. A detector score is a rumor, not evidence. Basing an integrity case on one is how good students get hurt.

What is a better way to catch AI cheating?

Shift from detecting the output to seeing the process. In-class writing and oral defense reveal what a student can do unaided. Version history shows whether a document grew naturally or arrived fully formed. And governing AI on school devices, so tutoring is allowed while whole-assignment generation is blocked and logged, gives you real signal instead of a guess.

How to Tell If a Student Used AI, Without Trusting a Detector

You usually cannot tell whether a student used AI by reading the finished essay, and you certainly cannot tell from a detector score. Modern AI writes clean, human-sounding prose, and detectors both miss edited AI work and falsely accuse honest students. What actually tells you is the process: how the work was made, whether the student can explain it, and whether it matches what you have seen them do in front of you. Here is how to get that signal, and what to stop trusting.

Why the detector is a dead end

It is worth being blunt, because a lot of integrity cases still lean on this. AI detectors do not work well enough to accuse a child. Independent testing keeps finding two failures at once: they clear AI text that has been lightly edited, and they flag genuine student writing as machine-made, with multilingual students taking the worst of the false positives. Several universities and vendors have quietly stepped back from AI detection for exactly this reason.

A detector gives you a percentage. A percentage is not evidence. If you would not accept "the machine said so" as proof in any other part of your professional life, do not accept it here, where the cost of being wrong is a student's record and your relationship with them.

What the finished text can and cannot tell you

Reading the essay is not useless, it is just weak. Some soft signals are worth noticing:

The voice does not match the student's earlier, in-class writing.
The argument is fluent but oddly generic, correct without being specific to your prompt or your class discussion.
Citations are clean but a couple of them do not actually exist when you check.

These are reasons to look closer. They are not conclusions. Treat them as "start a conversation," never as "file a case."

The real signal is process, not product

Everything reliable comes from knowing how the work was made, not from staring at what it became. Three sources hold up.

1. Work you watched happen

The most trustworthy sample of a student's ability is the writing they do in front of you. A short in-class piece, a handwritten reflection, a few minutes of them talking through their argument out loud. None of it needs to be high stakes. It just needs to be unaided, so you have a true baseline of what this student sounds like when it is only them.

2. The history of the document

A real essay has a messy history: it grows, gets reordered, has sentences deleted and rewritten. Work that was generated and pasted tends to arrive nearly whole, in a couple of large chunks, with no visible struggle. Version history and revision timelines will not prove intent, but they show you whether a document was built or delivered.

3. Can they defend it?

The oldest tool still works. Ask a student to walk you through a choice they made: why this example, why this structure, what they cut and why. A student who did the thinking can talk about it easily. A student who did not will struggle to explain work that is supposedly theirs. A two-minute conversation tells you more than any scanner.

Change the environment, not just the assignment

All of the above is detection after the fact. The larger move is to change the environment so you get honest signal by default, on the school devices where much of the work happens.

Banning AI outright does not do this. It pushes the tool onto phones and home laptops where you have no view at all, and it takes the legitimate study aid away from the honest majority. What works better is governing the behavior: let AI stay reachable for tutoring, explanation, and feedback, and block the specific act of generating an entire assignment, with each blocked attempt landing in a log with context. Now the tool that helps a struggling student is available, the tool that writes their essay for them is not, and you have a real record instead of a hunch.

What to do Monday

You do not need a new platform to start. You need to stop outsourcing your judgment to a scanner and put process back at the center:

Collect one piece of unaided, in-class writing per student early in the term, as a baseline.
When something feels off, open a conversation and ask the student to explain their work, before you ever consider a case.
Design at least some assignments to be defended out loud, so the thinking is the graded artifact.
Where you manage the devices, govern AI so help is allowed and wholesale generation is blocked and visible.

The goal was never to win an arms race against a text generator. It was to know your students well enough to recognize their work, and to build a classroom where doing the thinking is the only path that pays off.