How to Handle AI and Cheating Honestly

Start here, because this is the part that harms students

AI detectors are unreliable. They produce false positives, and those false positives fall hardest on students writing in a second language, students who write in an unusual register, and students whose writing is unusually plain or unusually formal. A detector score is not evidence. It should never be the basis of an accusation.

That paragraph is the most important thing in this course, so it goes first.

A detector does not know whether a machine wrote something. It produces a statistical guess about how predictable the text is. Writing that is careful, formulaic, taught to a structure, or produced by someone composing in their second language can all look predictable. So can writing by a student who has been drilled in a paragraph frame, which is to say, writing produced by good teaching.

The consequence is not abstract. A student wrongly accused of cheating loses trust in the teacher, in the school and often in their own work. Some of them never fully get it back. And the students most likely to be wrongly accused are frequently the ones with the least standing to argue.

What counts as evidence instead

Process, not product. If you can see the work being made, the question mostly answers itself. Version history in a shared document. Planning notes. A draft handed in at the halfway point. A photograph of a mind map. Rough work. None of this is surveillance, it is just asking to see the working, which teachers have done forever.

In-class writing you have seen. A short piece written in front of you at the start of a unit gives you a baseline of what this student sounds like. Not to catch them out, but so that a sudden change is a reason to ask rather than a reason to assume.

The student's own account. A student who wrote something can talk about it. They can tell you why they cut the second paragraph, what they nearly argued instead, where they got stuck. A student who did not write it usually cannot, and that conversation is more informative and much fairer than any score.

Help me redesign an assessment so it is harder to outsource
and still tests what I care about.
Subject and year group: [ ]
Current task: [describe it]
What I actually want to assess: [one or two sentences]
Constraints: [time available, whether devices are allowed, class size]
Suggest three redesigns. For each, tell me:

what changes about the task
why that makes it harder to hand to a tool
what it costs me in preparation or marking time
what it might make harder for students who struggle with writing
under pressure, so I can decide if the trade is worth it
Do not suggest anything that relies on detection software.

The conversation, if you decide to have one

If you have real concern based on process evidence, the conversation matters enormously. Go in curious, not accusing. You are trying to find out what happened, and there are many explanations, including help from a parent, a tutor, an older sibling, a website, panic the night before, or nothing at all.

Open with the work, not the accusation. Ask them to talk you through it. Ask what they nearly wrote instead. Ask which bit they found hardest. If a student wrote it, this is a pleasant conversation about their thinking. If they did not, the gap usually appears without you ever having to make a claim you cannot support.

Never open with a score. Never say a system flagged this. You are then defending a number you do not understand against a student who cannot disprove it, which is a bad position for everyone in the room.

Design your way out where you can

Policing scales badly and design scales well. Things that help, in rough order of how little effort they cost.

Ask for the process alongside the product. A plan, a draft, a reflection on what changed.

Anchor tasks to what happened in your room. This lesson's discussion, this trip, this experiment that went wrong, this local example. A tool that was not there cannot supply it convincingly.

Assess some things live. A short verbal defence, a two minute explanation, a question in the next lesson.

Ask for disclosure and mean it. If students can say what they used without being punished for saying so, you find out far more than any detector would tell you. That is lesson five.

Here is a piece of student work. Do not assess authorship.
Give me:

Five questions about the ideas in it that anyone who understands
this material could answer, whoever typed it
Two questions about specific choices made in this particular text
One question that would show me whether the student can apply
the same idea to a new example
Keep the questions conversational. They will be asked out loud
and they should not sound like a trap.
Work:
[paste it, with the name removed]

Write a short explanation for [year group] students about where
help ends and cheating begins with AI, for our subject.
Cover: using it to understand something, using it to check your own
work, using it to write something you hand in as yours.
Rules: under 200 words. Speak to them as capable people, not suspects.
No threats. Give two concrete examples of acceptable use and two of
unacceptable use, both drawn from ordinary schoolwork.