The AI Homework Problem: What Actually Works Instead of Detection

Every staffroom has had this conversation. A pupil hands in work that is suddenly three grades better and oddly free of personality. Someone suggests a detector. Someone else has heard the detectors are unreliable. The conversation ends with everyone slightly more tired and nothing decided.

Here is a position worth adopting: stop trying to detect, and start designing tasks where using AI dishonestly costs more effort than doing the work. That is achievable within your existing curriculum, this term.

Why detection is the wrong tool

Detectors report a statistical impression of how predictable text is. Three consequences follow, and all three are documented in the research literature and in the vendors' own caveats.

False positives cluster on the pupils least able to defend themselves. Writing by pupils using English as an additional language is systematically more likely to be flagged, because learner writing uses more common words and simpler structures - the same features that look machine-like.

The false negative rate is high and trivially exploitable. Asking for a rewrite in a specified voice defeats most detectors. Any pupil who wants to evade detection can, in one extra step.

The score is not evidence. You cannot show a family a probability and call it proof. A pastoral leader will ask what you actually observed, and if the answer is only a percentage, the case collapses - as it should.

One more point that gets missed: even a perfectly accurate detector would only tell you that a pupil avoided doing the thinking. It would not tell you what they can do. That is the information you actually needed.

What to do when you suspect misuse

Have a routine, so you are not improvising in front of an upset pupil.

1. Ask, do not accuse. Talk me through how you got to this paragraph. A pupil who did the work answers easily. A pupil who did not usually tells you within a minute, often with relief.
2. Ask about the thinking, not the words. Why did you choose this example over the one we used in class? Ownership of a decision cannot be faked by someone who never made it.
3. Compare with in-class writing. Your own record of what this pupil produces unaided is the only baseline worth having, and it is stronger evidence than any tool.
4. Separate the integrity issue from the learning issue. Whatever the outcome, the pupil still has not practised the skill. Set the learning task regardless.
5. Follow your policy, in writing, once. Escalate through the route your school has agreed. Never leave a serious accusation as an undocumented corridor conversation.

Seven redesigns that make misuse pointless

None of these require new technology and none add marking load once established.

1. Assess the process, not just the product. Require a plan, a draft and a reflection. Mark the movement between them. A pupil who submits a polished final piece with no trail has not met the brief.

2. Anchor the task in something local and recent. Our fieldwork data. Tuesday's demonstration. The argument the class had about the source. Nothing outside the room knows about these.

3. Require the pupil's own voice on the record. Two minutes explaining their argument to you, or a ninety-second audio note submitted with the work. This is the single most effective change I have made; it takes half a lesson for a class of thirty using a carousel.

4. Write in class, extend at home. Do the thinking under your eye; use homework for the parts where help is legitimate - reading, gathering, formatting, revising.

5. Make the task require a judgement, not a summary. Which of these three interpretations is weakest and why is far harder to outsource than explain the causes of. It is also better assessment.

6. Permit AI use explicitly, with disclosure. Add a required line: what I used, what I asked it, what I changed. Undisclosed use becomes the offence, which is enforceable. Disclosed use becomes a teachable artefact, and you learn a great deal about who is using what.

7. Grade drafts, not only finals. A final piece unrecognisable from a graded draft is a visible conversation starter with no accusation attached.

A worked redesign

Original: Write 800 words on whether social media harms teenagers. Due Monday. Outsourceable in thirty seconds, and the output would be decent.

Redesigned:

• In class: pupils survey their own year group and produce a small dataset.
• Homework: read two provided sources, annotate one, note where each disagrees with the class data.
• In class: write the argument, 40 minutes, sources and annotations allowed, devices closed.
• Homework: revise for clarity. AI permitted for revision, with the disclosure line required.
• Assessed: the argument, the annotations, and one paragraph on what the class data could not tell them.

Same curriculum content. Same word count. Marking load essentially unchanged. There is now no version of this task that AI can complete for a pupil, because the required material exists only in your classroom.

What to say to your class

Say it once, plainly, in week one, and then be consistent.

I am not going to run your work through a detector, because detectors are unreliable and I am not prepared to accuse anyone on a guess. I will ask you to explain your thinking, and that will tell me what I need to know. Where AI is allowed, I will say so and you will tell me what you used. Where it is not, the reason is that you need the practice - and the exam will not care what a tool can do.

Pupils respond to that far better than to surveillance, because it is honest about the reasoning. The classes where I have said this have had fewer integrity problems, not more.

FAQ

My school has already bought a detector. Now what? Use it, if required, as one signal that prompts a conversation - never as the basis of a sanction on its own. Insist that any escalation records what was observed in discussion, not the percentage.

Is not banning AI simpler? Simpler to state, impossible to enforce, and it pushes use underground where you cannot teach into it. Permitted-with-disclosure gives you visibility.

Does the oral check scale to 150 pupils? For major pieces only, and as a sample: five pupils per class chosen at random, announced in advance. The possibility of being asked changes behaviour across the whole cohort.

What about coursework with external moderation? Follow your awarding body's rules exactly; they are more prescriptive than school policy and they change. Check the current specification each year rather than relying on last year's guidance.