Picture a Year 10 history essay that is, frankly, too good. The argument flows, the vocabulary is polished and the pupil has never written like that before. The teacher runs it through an artificial intelligence (AI) detector, which says it is human. She still does not believe it, and she still does not know what to do. A new report on AI cheating in schools suggests that this quiet stand-off, repeated in classrooms everywhere, is now doing more harm than the cheating itself.

In short: researchers at Harvard’s Center for Digital Thriving surveyed more than 1,000 United States (US) teachers and principals and interviewed 31 young people. When teachers described a dilemma with AI, 74% of their examples were about suspected cheating. Few schools had a specific AI policy, detectors cannot settle the question, and pupils describe their own fear of being falsely accused. The researchers call this two-way suspicion. Their answer is to talk to pupils, assignment by assignment, about where the line sits.

What the Center for Digital Thriving found

The report, An AI Policy Isn’t a Playbook, was published in September by the centre, which is based at Project Zero within the Harvard Graduate School of Education. Its authors are Beck Tench, Emily Weinstein and Carrie James, with Allison Starks and Sara Konrath. Harvard’s Usable Knowledge summary followed on 6 October.

The survey drew nationally representative samples of 719 teachers and 299 principals in US public schools from the RAND Corporation’s American Educator Panels, fielded in April and May 2025. Interviews with pupils aged 15 to 19 ran from April to June this year. The numbers are American and about eighteen months old, but I think they travel well to British classrooms.

A quarter of teachers (25%) and 29% of principals said generative AI had caused them a dilemma or a difficult decision. The figure climbed sharply with the age of the pupils: 11% of elementary teachers, 29% in middle school and 44% in high school. When the researchers coded the stories behind those answers, 74% of teachers’ examples and 69% of principals’ were about cheating.

Policy was thin. Only 18% of teachers and 24% of principals said their school had specific policies or guidelines on generative AI, and that figure even includes those who said the policy needed significant improvement.

Why AI cheating in schools is turning into a trust problem

The most striking pages describe what happens to a classroom when nobody can prove anything. One teacher wrote: “When a student digs in and says they didn’t use it, there’s no actual proof.” The pupils described the same burden from the other side. One high-achieving 15-year-old told the researchers: “We can’t make ourselves sound too smart anymore, we cannot give our full effort anymore, because then we’re at the risk of being detected for AI.”

Read that again. A pupil is deliberately writing less well so that a teacher will believe the work is theirs. Another said: “It sucks because I never was going to cheat, but now it feels like I have to.”

Suspicion runs upwards too. Pupils notice when teachers use AI to plan lessons or mark, and some read the “original work only” rule as hypocrisy. The researchers name the whole pattern two-way suspicion: both sides suspect the other of using AI, and trust erodes whether or not anyone actually did. The quote that stayed with me came from a 16-year-old: “I don’t usually go to my teachers. I kind of just go to Chat.”

Detectors will not rescue us

Software will not settle this. When an international team of academic integrity researchers tested 14 detection tools, including Turnitin, they concluded the tools were “neither accurate nor reliable“. Lightly edited or machine-paraphrased AI text slipped through about half the time. Worse for many of our classes, human writing that had been put through machine translation was more likely to be flagged, which leaves pupils who draft in another language and translate their own work more exposed to false accusations. The authors were blunt: detector reports cannot be the only basis for accusing a student.

That matches the Joint Council for Qualifications (JCQ) position on coursework, which we covered in our guide to the JCQ AI guidance for 2026: detection software must not be relied on alone. A detector can raise a question. It cannot answer one.

When a pupil starts writing worse on purpose so that you will believe them, the suspicion has already cost more than the shortcut.

From referee to coach

The report’s central image comes from Ted Lasso, the television comedy about an American coach running an English football club. Teachers, the authors argue, have been handed a whistle and asked to referee a game nobody taught them, when what pupils need is a coach. Rather than another rulebook, they offer five free classroom routines, which they call plays. Three stand out for a busy department.

Align on the Line is the one I would try first. Pick a real task, say a Year 9 analytical essay on Macbeth. Ask the class to list every way someone could use AI to help with it. The word “could” matters, because nobody has to confess anything. Each pupil places those uses on a five-point scale, from 1 (totally fine) through 3 (not really sure) to 5 (crosses a line). In the report’s own example, about an essay on The Great Gatsby, spellchecking lands at 1 and “write me the whole essay” lands at 5. The interesting arguments happen in the middle. Is asking a chatbot for the main themes a fair starting point, or the very thing the essay is meant to test? The class then records where the line falls for this assignment.

Making AI Pressures Visible borrows Ethan Mollick’s phrase “secret cyborg” for someone who uses AI but hides it. Pupils list moments when others feel pressure to keep AI use quiet. Talking about other people lets them name a pressure without owning it, and the researchers say the room opens up when the teacher shares one first.

Name the Game gives everyone vocabulary. The centre’s AI to Z glossary includes “AI-dar” (the perceived ability to spot AI use) and “why botherism” (the feeling that if AI can do a task there is little point learning it). The authors suggest running it with colleagues first. A staff briefing that starts with an honest chat about our own AI-dar would beat another slide on plagiarism.

One caveat the authors state openly: these plays come from design work with teachers and young people, and they have not been tested in controlled studies. Treat them as well-made conversation tools, not proven interventions.

What I would do this half term

None of this needs a new policy document, just a few habits.

• Before the next extended piece of writing, run a 20-minute Align on the Line session and put the agreed line on the task sheet.
• When a piece of work worries you, start with a conversation about how it was made rather than an accusation. Ask the pupil to talk you through a paragraph or show an earlier draft.
• Build more of the thinking into lessons, with plans, drafts and short spoken explanations, so you see the work grow. That also helps with cognitive offloading, where pupils hand over the very thinking a task is meant to build.
• Be open about your own AI use. If you used a chatbot to draft a quiz, say so and explain why that is different from a pupil using one to write the essay the quiz is preparing them for.
• Never use a detector score as evidence on its own, and be especially careful with pupils who translate their own writing from another language.

The report’s closing mindshifts are simple: move from assuming to asking, from a focus on the technology to a focus on the young person, and from “us versus them” to “us and them”. That last one is the heart of it. We have written before about psychological safety in learning: pupils ask better questions and take more risks when they feel safe enough to be honest with us. A classroom where every polished paragraph is a potential charge sheet is not that place.

Cheating with AI is real, and some pupils will hand over whole assignments. The bigger risk is that our response teaches the honest majority to hide, to write down to our expectations and to take their questions to a chatbot instead of to us.

Where does your school draw the line on AI for homework and coursework, and have you tried agreeing it with your classes? Tell me in the comments.