Are your students cheating with AI? The answer is harder than you think

I will start with the sentence that has to land before anything else:

AI detectors are unreliable, and no penalty should ever rest on one.

The technical reason is simple. These detectors measure how predictable a text is, meaning how close it sits to expected phrasing. And two things produce predictable text: a language model, and a student writing in a second language who is carefully following the structures they were taught.

In our schools, the second group is most students.

The wrong question and the right one#

The wrong question: how do I catch students using AI?

The right question: what am I trying to measure, and does my assessment actually measure it?

If your assessment is "write five hundred words on the causes of the First World War", it stopped measuring understanding two years ago. That is a problem with the assessment, not with the student.

Assessments that cannot be cheated#

A short oral. Three minutes with the student about what they submitted. Someone who wrote it can defend it; someone who did not, cannot. Faster and more accurate than any detector.

Process, not product. Ask for drafts, version history, or a photo of the first handwritten page. Process is harder to fake than an outcome.

Personal context. "Connect this to something that happened in your neighbourhood", or "use an example from the school trip". The model does not know their neighbourhood.

Judgement, not recall. Give them a generated text containing three errors and ask them to find them. This measures higher understanding and uses the tool rather than fighting it.

In-class performance. What happens in front of you needs no detector.

That one line#

Ask for a line at the end of every piece of work: what was used, and for which part.

"I used the assistant to reorder my paragraphs and to check grammar; the ideas are mine." That declaration changes the whole relationship. Use becomes something mentioned rather than hidden, and you become the person teaching the boundary rather than the person policing it.

What do you say to a student you suspect?#

Do not accuse. Ask.

"Talk me through your thinking in the third paragraph" reveals everything and leaves an innocent student a dignified exit. "This was written by AI" costs you the student, and possibly the class, if you are wrong.

What this means for you#

If you teach: change one assessment this term so it cannot be cheated. Better than chasing ten students.

If you lead a school: explicitly forbid basing penalties on detectors in your written policy. It protects your school legally and your students personally.

If you are a parent: if your child is accused on the basis of a detector, calmly ask for other evidence. A detector alone is not evidence.

In closing#

You cannot prevent use. You can absolutely design assessment that measures what you meant to measure. The first is a losing battle; the second was always your job.

Try this today: take the last assignment you set and ask whether a model could complete it in two minutes. If yes, the problem is the assignment.

Common questions#

Are AI cheating detectors reliable?

No, and no penalty should rest on one. They measure how predictable text is, and predictable text is produced both by language models and by students writing in a second language, who are most of our students.

How do I design assessment that cannot be cheated?

A short oral about the submitted work, asking for process rather than product, requiring personal local context, asking students to find errors in a generated text, or in-class performance.

What do I say to a student I suspect?

Do not accuse; ask. "Talk me through your thinking in paragraph three" reveals everything and leaves an innocent student a dignified exit, unlike a direct accusation that costs you the class if you are wrong.

Should I ban use entirely?

A ban is unenforceable. More effective is a declaration line on every piece of work stating what was used and where, which turns use into something mentioned rather than hidden.