Tuesday, September 22, 2026

How to Tell If a Student Used AI on an Assignment (Without Accusing Anyone Unfairly)

If you suspect a student used AI on an assignment, the honest answer is that you probably cannot prove it — and that detection software will not give you proof either. What you can do is notice the right signals, ask better questions, and design the next assignment so the problem largely disappears. This is a guide to doing that without accusing anyone unfairly.

It is worth saying at the outset why this is difficult. A false accusation damages a student badly and is very hard to walk back. A missed case feels like letting someone get away with it. Most teachers are caught between those two fears, and the advice available tends to be either "use a detector" or "redesign everything you teach." Neither is much help on a Tuesday evening with thirty essays in front of you.

Why AI detectors should not be your evidence

Start here, because a lot of teachers are relying on something that will not hold up.

AI detectors work on probability. They look at how predictable the word choices are and return a percentage. That percentage is not a measurement of whether AI was used — it is a measurement of how unsurprising the writing is. Those are different things, and the gap between them is where students get hurt.

Two groups get flagged disproportionately. Students writing in a second language often produce simpler, more predictable sentence patterns, which reads to a detector exactly like generated text. So do students who write carefully and formally — the ones following the structure you taught them. A detector can penalize a student for doing what you asked.

There is also no way to interrogate the result. If a student asks "which part did it flag, and why," you have no answer beyond a number. That is not a position you want to be in during a meeting with a parent.

Use a detector as a prompt to look more closely if you like. Do not use it as the reason for a conversation, and never quote a percentage to a student or a parent as though it were evidence.

What actually suggests AI use

The reliable signals are not stylistic. They are gaps between the work and the student.

A jump you cannot account for. A student who has written in short, disconnected paragraphs all semester hands in something with smooth transitions and a controlled argument. Growth is real and worth celebrating — but growth usually shows up as a better version of the same voice, not a different voice entirely.

Fluent writing about things that were never taught. The essay is well constructed and references ideas that were not in your lessons, not in your lessons, not in the assigned reading, and not in any other material you shared. Independent research is possible. It is also worth asking about.

Confident, specific, wrong detail. This is the strongest signal there is. A generated paragraph will state a date, a statistic, or a quotation with total assurance and get it wrong in a way a student who did the reading would not. A student who skimmed the material makes vague errors. Generated text makes precise ones.

Sources that do not exist. Check two citations. If a title, author, or page reference cannot be found, that is close to conclusive — students who fabricate sources normally invent plausible ones, not perfectly formatted ones.

Nothing personal in it. You asked them to connect the topic to their own experience, and the response is generic. Generated text is poor at the specific, slightly odd details of a real life.

Perfect structure, no position. Five well-balanced paragraphs that never commit to an argument. Students tend to have opinions and express them clumsily. Evenness is not always a good sign.

One signal on its own means very little. Three together is worth a conversation.

What to say to the student

This is the part that determines whether the situation improves or becomes a formal complaint. Two principles: ask about the work, not about their character, and give them a route that is not a confession.

Do it privately, sitting down, without the essay held up as an exhibit. Open with something like:

"I've read your essay and I want to understand your thinking on it. Talk me through how you got to your main argument — where did that idea come from?"

A student who wrote it will answer easily. They will remember the bit they found hard, the paragraph they rewrote, the source they nearly used. A student who did not will give you a summary of what the essay says rather than an account of writing it. That difference is unmistakable and you do not need to be an expert to hear it.

If the answers are thin, go more specific rather than more confrontational:

"In paragraph three you mention [detail]. Tell me more about that — where did you find it?"

Then offer the exit:

"I'll be straight with you — parts of this don't sound like your writing, and I'd rather ask than assume. If you used AI for some of it, tell me now and we'll sort out what happens next together. If you didn't, that's fine, and I'd like you to walk me through your process so I understand."

That phrasing does specific work. It states your concern honestly rather than laying a trap. It makes admitting easier than lying. And it leaves the student's dignity intact if you have got it wrong, which sometimes you will.

Avoid "I know you used AI," "the detector says," and anything said in front of the class. The first two you cannot support, and the third turns a teaching moment into a humiliation the student will remember longer than the lesson.

When you are not sure — and you often won't be

Sometimes you finish the conversation still uncertain. That is uncomfortable, but it is a normal outcome, and there is a reasonable way to handle it.

Grade the work in front of you. Then set a short, related task done in class, on paper, in twenty minutes — explain one of your arguments in more depth, or apply the same idea to a new example. Set it for everyone, not just the student you are unsure about.

If the in-class writing matches the essay, you had your answer and no harm was done. If it is dramatically weaker, you now have something concrete to work with, produced under your supervision, rather than a suspicion. This is more useful than any detector and it does not single anyone out.

Designing assignments so it matters less

The long-term answer is not better detection. It is asking for work that generated text is bad at.

Require material only your class has. A specific discussion from Tuesday, the data your students collected, the way your school handles something. AI cannot invent what it has never seen.

Ask for process, not just product. Hand in the outline and one rejected draft alongside the final piece. Faking a plausible bad first draft is more effort than doing the work.

Build in personal connection. "Relate this to something you have seen or experienced" is hard to generate convincingly and easy to probe in conversation.

Make the thinking visible in class. Ten minutes of planning on paper before they go home means you have seen the argument form. It also helps the students who genuinely do not know how to start.

Ask them to critique, not summarize. "What is the weakest part of this argument and why" produces far more distinctive answers than "summarize this argument."

None of this makes AI use impossible. It makes it more effort than doing the work, which is the realistic goal.

Have the conversation before it happens

Most students using AI are not trying to cheat in the way we assume. They are behind, or stuck, or panicking at eleven at night with nothing on the page. The tool solves an immediate problem and the ethics get worked out later, if at all.

That means a clear, early, specific conversation prevents more than any policy does. Tell them what is fine and what is not — and be concrete, because "don't use AI" is not a workable rule for a generation whose search results are AI summaries. Using it to explain a concept you did not understand is different from having it write your paragraph. Say which is which.

Tell them why it matters, in terms they care about. Not academic integrity in the abstract, but: the writing is how you learn to think, and if something else does it, you arrive at the exam having practiced nothing.

And tell them what to do instead when they are stuck at eleven at night. Most of them genuinely do not know, and that gap is where the problem lives.

The short version

Detectors are not evidence and should not be quoted to students or parents.

The real signals are gaps between the work and the student, particularly confident wrong detail and sources that do not exist.

Ask about the process, not the person, and give an honest route to admitting.

When unsure, set a short supervised task for the whole class rather than pursuing one student.

And design assignments that ask for what only your students can provide — it solves more than any detection ever will.

Building assignments that test understanding rather than recall?

Try the Worksheet Generator

Related reading

No comments:

Post a Comment