Schools should respond to AI cheating by fixing the task before chasing the student: say how much AI each piece of work allows, and assess the thinking that matters under supervision. A detector score can’t prove anything, so treat it as a reason to talk, and follow JCQ’s rules for coursework, where unacknowledged AI use is malpractice.
Picture a pile of Year 10 history homework that’s landed on your desk. A student who usually manages half a page has handed in two polished pages. The spelling’s American. One reference doesn’t seem to exist. A colleague suggests running it through a free online checker. Before anyone does, have a proper look at the task. If a machine could’ve done it start to finish at the kitchen table, the task is part of the problem.
What counts as AI cheating in school
AI cheating is using AI in a way the task didn’t allow, or using it and not saying so. So you’ll need the rule before the work’s set, or students are guessing. It’s one of the first decisions in a school’s wider plan for AI.
For qualifications, you don’t have to draw the line yourself. The Joint Council for Qualifications defines AI misuse as using AI without properly acknowledging it and handing in work that isn’t the student’s own (JCQ AI Use in Assessments, 2025). Copying or paraphrasing whole AI answers counts. Using AI for parts of an assessment counts too, when the work no longer reflects the student’s own.
Your homework needs the same clarity. If the sheet’s silent on it, two very different students are both “using AI”. One checked their spelling. The other had it write everything, and you’ve no fair way to tell them apart.
What JCQ and Ofqual expect for coursework
It’s coursework, non-exam assessment and internal assessment where the rules bite. Most exams are closely supervised with no internet, so JCQ says they should be unaffected by AI tools.
In coursework, they’re firm. Teachers “must not accept work which is not the student’s own”.
Students who’ve used AI must acknowledge it. They name the AI source and the date, and keep a copy of what they asked and what came back. Misuse is malpractice, and that’s serious: sanctions can include disqualification.
Ofqual’s turned the heat up since. Its March 2026 pack for senior leaders says the Chief Regulator has written to exam boards about “serious concerns about AI-generated coursework”. In the same pack, Ofqual says “a student using AI to produce coursework and passing it off as their own is cheating” (Ofqual AI coursework pack, 2026). The pack asks leaders to check which qualifications include coursework, brief staff on JCQ’s policy and back teachers who raise concerns.
Recorded numbers aren’t big yet. In summer 2025, 100 GCSE, AS and A level malpractice cases involved plagiarism through AI misuse, up from 85 a year earlier (Ofqual malpractice statistics, 2025). That’s three in four plagiarism cases, but only 2% of student malpractice. They’re also cases that ended in a penalty, so they can’t show what went unnoticed.
Outside qualifications, it’s your call. The Department for Education says schools “may wish to review homework policies, and other types of unsupervised study” (DfE generative AI guidance, 2025).
This is general information, not legal advice, checked on 9 October 2026: Ofqual and the DfE cover England, JCQ’s rules cover its member boards’ qualifications across the UK, and your exams officer and data protection officer can check the detail for your school.
How teachers check for AI
The best way teachers check for AI is by knowing the student’s work and talking to them about it. Software can add to the picture, but it can’t settle the question.
Detectors haven’t been reliable from the start. OpenAI’s own detector, launched in January 2023, caught 26% of AI writing and wrongly flagged human writing 9% of the time. It was withdrawn that July “due to its low rate of accuracy” (OpenAI classifier note, 2023). That year, a Stanford team ran seven detectors over 91 essays by non-native English speakers. On average, they wrongly called 61% of them AI-generated. Essays by US eighth graders were judged almost perfectly (Liang and colleagues in Patterns, 2023). So I’d think hard before trusting a score on a pupil with English as an additional language.
In a 2024 study, detectors spotted unaltered AI text only 39.5% of the time on average. Simple editing tricks dropped that to about 22%. The authors’ verdict was that detectors “cannot currently be recommended for determining academic integrity violations” (Perkins and colleagues in IJETHE, 2024). At the University of Reading, researchers slipped answers written entirely by AI into real take-home exams. Markers missed 94% of them (Scarfe and colleagues in PLOS ONE, 2024). So human eyes aren’t a reliable check either.
JCQ’s 2025 guidance says detection tools vary in accuracy, and treats them as one source of evidence. It says teachers “will know their students best and so are best placed to assess the authenticity of work”. Its warning signs are ones you’d spot anyway. They include writing that’s unlike the student’s classwork, references you can’t find and American spelling.
And don’t paste a pupil’s work into a free online checker. The DfE’s leadership materials say doing that without permission “is likely to breach intellectual property rights” (DfE AI leadership transcripts, 2026). A free tool may also learn from what’s uploaded.
Then talk to them. Ofqual’s pack asks teachers to get students “to explain their thinking and describe their process”.
Ask how they started and what they cut. Get them to explain one paragraph in their own words. Write down what they said that day, because you won’t remember it word for word. If you’re still concerned, follow your malpractice procedure. For qualification coursework, if you’re not confident the work’s theirs, you can’t accept it, whatever a detector says.
I call the habit of pausing to check information before you act on it Information Hygiene. It’s easy to point it at the student with the chatbot answer. Point it at the adult holding a detector score, too.
Why redesigning the task works better than catching students
Cheating didn’t arrive with AI. A 2026 survey of 4,354 students at six US high schools found 72% reported some form of cheating. That’s in line with historical baselines, the researchers say. Their reading is that AI hasn’t changed how common cheating is in high school (Chen and colleagues in ETR&D, 2026).
What’s changed is how cheap the old take-home task became. Ofqual’s pack says that “for students facing deadlines, the temptation is real”. If a machine can do the homework in a minute and it’s due tomorrow, the task has made copying easy. ‘Find five facts about the Tudors and write 300 words’ is a set of steps a machine can follow. The page that comes back tells you little about the student. DfE leadership materials suggest homework that focuses on “tasks that can’t be easily completed by AI”. I drew the same line when I looked at which parts of teaching AI can take on: AI can draft the routine pieces, and the judgement stays with the teacher.
Give every task a clear AI level
Every task should say how much AI is allowed, so nobody’s guessing. One public option is the AI Assessment Scale, created by Mike Perkins, Leon Furze, Jasper Roe and Jason MacVaugh. Its 2025 version has five levels (Perkins, Roe and Furze in JALT, 2025).
No AI means a controlled setting with AI kept out. AI Planning lets students use AI at the start, for topic ideas, an outline and first research. With AI Collaboration, it helps with ideas, drafts and feedback. Full AI means AI use is expected, and AI Exploration builds the task around using AI creatively to solve problems.
The authors say the levels aren’t a ranking. No AI isn’t the ‘proper’ version, and Full AI isn’t cheating with permission.
Their sharpest warning is about schools labelling take-home essays ‘No AI’. There’s no realistic way to enforce that, they say, and it leaves the assessment “both invalid and insecure”. So if a task really can’t allow AI, do it in the room.
Put the level at the top of every task sheet, with a line on how they’ll acknowledge AI use. Then write the levels into your school’s AI policy, so every department uses the same words. Tasks at AI Collaboration or Full AI are where you can teach students to question what AI tells them.
A redesign table for your next department meeting
Run your department’s most common tasks through this. The levels are the AI Assessment Scale’s.
| Task as it’s often set | What AI can do | A redesign | Level to state |
|---|---|---|---|
| 500-word take-home essay | Write a fluent draft in seconds | Plan at home, write the essay in class | AI Planning at home, No AI in class |
| Research project or coursework | Summarise sources, draft sections, invent references | Drafts seen at set check-ins, a two-minute talk-through, every reference checked | AI Planning or AI Collaboration, acknowledged; for qualification coursework, JCQ’s rules decide |
| Science write-up | Write a method and conclusion | Students write up their own group’s results in class, from the raw data | No AI |
| Presentation | Write the script and slides | AI helps with slides; assess the questions afterwards, live | AI Collaboration for slides, No AI for questions |
Pick the two tasks you set most. Change one before the next homework goes out, and put its level at the top of the sheet.
How Colin helps
I’d start with your task sheets before your software budget. You won’t need a licence for most of what I’d change.
My first question about cheating is a behaviour one. What makes copying the easy option, and what would make the honest route easier? That’s the lens I bring as the founder of The Human Behaviour Institute and a Professional Affiliate of The British Psychological Society. With a leadership team, I’d work through your most-set tasks and agree the level each one needs.
If you’d like to talk about a keynote or advisory work for your school, get in touch.
Related questions
What is the 30% rule for AI?
There isn’t one in JCQ’s rules for UK qualifications. Its guidance sets no acceptable percentage of AI. Any AI use in assessed work has to be acknowledged, and the work has to be the student’s own. Detectors can’t reliably measure a percentage anyway.
Can AI writing be detected if it’s rewritten?
Less reliably, and it wasn’t reliable to begin with. JCQ says detection tools give lower scores once students have amended AI text. Simple edits cut average accuracy to about 22% in the 2024 study above, which is why I’d talk to the student instead.
Should schools ban AI?
A blanket ban can’t be enforced outside supervised conditions. That’s the scale authors’ point about ‘No AI’ labels on take-home work. Set a level for each task instead, and keep No AI for work done in the room. I think bans driven by worry cost schools more than they save, as my piece on fear of AI in classrooms explains.