AI detectors are unreliable enough that a flag is not proof of anything. They measure how predictable your writing is, not whether a machine wrote it — so careful, formal or second-language writing gets flagged constantly. A Stanford study found roughly 61% of essays by non-native English speakers were wrongly flagged, and HEPI's position in 2026 is that a detector score should never be sufficient evidence of misconduct on its own.
You spent three weeks on it. You read the actual books. You wrote every sentence yourself, at 1am, fuelled by the kind of coffee that tastes like a decision you'll regret. And then a little coloured bar tells your lecturer it's 43% “AI-generated.”
If your stomach just dropped reading that — you're not being dramatic, and you're not alone.
The detectors are genuinely, measurably unreliable
This isn't cope. It's the current evidence.
The Stanford finding above is the one most people quote, and it deserves the attention: not a rounding error, but the majority of essays from one group of students, wrongly flagged. The Higher Education Policy Institute — about as establishment as UK higher-ed commentary gets — spent the summer of 2026 publishing on exactly this, and their line is blunt:
A detector score should never be sufficient evidence of misconduct on its own.HEPI, August 2026
Some universities abroad have already switched their detectors off as too unreliable to act on. Most UK universities haven't — yet. Which means, for now, real students are still getting flagged for work they wrote themselves. In 2026 the Office of the Independent Adjudicator upheld complaints from students — including an autistic student and several non-native English speakers — who were accused on the strength of a detector score and later cleared.
So if you got flagged: it is entirely possible the tool is simply wrong. That happens constantly.
It's worth being precise: these tools err both ways. A high score isn't proof of cheating, and a clean score isn't a clearance. A measure that fails in both directions isn't measuring the thing its name claims to.
Why does your own writing “look like AI”?
Here's the uncomfortable bit. Detectors don't detect AI. There's no watermark in a paragraph, no fingerprint left behind. What they detect are patterns — and some of those patterns are just… good, clean, careful writing. The exact things you were told to do.
You're more likely to trip a detector if you:
- Write in short, even, tidy sentences — you were told to be concise.
- Use formal, hedged academic phrasing — “this essay will argue that”, “it can be suggested that”.
- Structure everything neatly with clear topic sentences — you were told to signpost.
- Are writing in your second, third or fourth language and lean on safe, correct, textbook constructions rather than risking an idiom.
- Are neurodivergent and write with consistent structure, repeated signposting and even sentence rhythm — all rewarded by a rubric, all penalised by a detector.
None of that is cheating. It's a student doing what the marking rubric asked. But it reads as low surprise to a detector, and low surprise is what these tools score as machine-like. The better you follow the rules, sometimes, the more you look like a robot to a robot.
If you want the mechanics rather than the summary, we've written up perplexity and burstiness — the two statistical measures nearly every detector is built on.
The “everyone's doing it” problem — and where the line actually is
Let's be honest about 2026. HEPI reports 94% of UK undergraduates now use generative AI for assessed work in some form. Pasting AI text straight in has gone from 3% of students in 2024 to 12% in 2026. AI is in the room whether your department likes it or not.
Which creates a genuinely confusing situation: if nearly everyone uses it somehow, where's the line between “fine” and “misconduct”? Most students who feel guilty aren't cheating — they're just not sure. Roughly, the line most UK departments draw:
- Usually fine: brainstorming ideas, getting a topic explained, checking your grammar and spelling, planning a structure — then writing the essay yourself.
- Usually not fine: generating paragraphs and submitting them as your own words, having AI write your argument, or paraphrasing AI output to disguise where it came from.
Could you explain and defend every argument in your essay, out loud, without the AI in front of you? If yes, it's your work. If the ideas and the words are the model's and you tidied them, that's the danger zone.
The exact rule is set by your course, and it varies — sometimes between modules on the same campus. If you're unsure, the two-minute email to your module leader is the cheapest insurance you'll ever buy. And if you only used AI to plan or proofread, you have nothing to hide; you just need to be able to say so with a clear conscience. Our answers on planning an essay with AI and using AI for grammar go through the usual boundaries in detail.
What to actually do — before you submit, not after
You can't control your university's detector. You can stop walking into a meeting blind. Three moves:
- Check your own work privately, first. See what a scan says about your essay while you can still do something about it — not in a misconduct meeting where the evidence is already on the table.
- Understand why something looks AI-ish and rewrite it in your own voice. Not to “beat” a detector — to write more like a human who actually thinks, which is what markers reward anyway.
- Keep your receipts. Drafts, notes, version history, library loans. If you're ever asked “how did you write this?”, boring proof of a boring process is your best friend.
Write in one continuously-edited document from plan to final draft, rather than assembling a clean copy at the end. The version history of that one document is the most persuasive thing you can produce if anyone ever asks — and it costs you nothing to create.
If you've already been accused
A flag is an allegation, not a verdict. Don't fire off an anxious 2am email that over-explains — that's the most common way honest students make things harder for themselves. Instead: ask in writing which regulation you're alleged to have breached and on what evidence, request the detector report itself rather than just the headline number, freeze your drafts and version history before you touch the file again, and get your students' union academic advice service involved early. They do this constantly and it's free.
The full step-by-step is here: what happens if I'm wrongly accused of using AI? — and our longer guide on showing the work is yours covers the process evidence in more depth.
Where SafeGrade fits (the honest version)
SafeGrade is a UK student tool that scans your essay before you submit — for AI-looking patterns, grammar, referencing, and a rough grade band — and tells you what's raising flags so you can fix it in your own words. Two things make it different from the free web checkers:
- Your essay stays yours. It's analysed privately on our own infrastructure — never stored, never used to train anything, never handed to a third party. Given that the whole anxiety here is about your work being seen and judged, that matters.
- It's built for UK students — Harvard and APA referencing, UK grade bands, and Sage, an advisor that knows your university and course.
And because we know you're skint (it's 2026, everyone is), the first scan is free — no card, no catch. If checking your own essay before your lecturer does is worth £3.99 a month to you, Pro is there. If it isn't, the free scan is still yours.
We're not here to help anyone hide anything. We're here so that the person who did the work doesn't get done for it.
One free scan. No card, no catch. See what your essay looks like while you can still change it — and keep the receipts either way.
Run one free scan →