Deep Dive Nº 08 · Integrity

The detector that can't tell you from a bot

94% of UK undergraduates now use generative AI on assessed work. The software built to catch them misreads more than six in ten essays by non-native English speakers as machine-written. Universities are starting to quietly drop it — here's what that actually changes for you.

For three years, "we'll run it through a detector" has been the quiet threat behind most AI guidance in UK higher education. That threat is losing its teeth. Not because universities had a change of heart, but because the tools themselves turned out not to work — and the students most likely to be wrongly caught by them were rarely the ones actually cheating.

Waterloo dropped Turnitin's AI-detection feature in September 2025. Curtin followed in January 2026. UCLA and UC San Diego had already stepped away in 2024. UK institutions have since started restricting or abandoning the same tools — not as a pedagogy debate won on principle, but as a response to evidence that had become too specific to ignore.

The gapEveryone's using it, the policy hasn't caught up

The starting numbers, from HEPI's Student Generative AI Survey 2026 (1,054 full-time UK undergraduates, with Kortext): almost every student is already doing the thing the policy argues about.

94%
of UK undergraduates use generative AI to help with assessed work
48%
feel their teaching staff are actually helping them build AI skills

As two senior University of Westminster academics, Doug Specht and Gunter Saunders, put it in Times Higher Education: "the policy process ran in one direction; practice ran in another." Rules were still debating whether using ChatGPT counts as plagiarism while the overwhelming majority of students had already folded it into how they work.

Why detection brokeThe tools don't measure what they claim to

The evidence against AI detectors isn't a single disputed study — it's now a settled, repeated finding across independent research. Stanford researchers, publishing in the journal Patterns, tested seven widely used detectors and found over 61% of essays by non-native English speakers misclassified as AI-generated, while the same tools performed near-perfectly on native-English writers. A separate 805-sample evaluation put average detector accuracy at just 39.5% on unmodified AI text — falling to 17.4% once simple evasion techniques were applied. An independent review by Weber-Wulff and colleagues, the most comprehensive run to date, concluded that none of the tools tested met the bar required for a high-stakes decision about a student's future.

"The detection-and-punish era did not end because universities won the argument about pedagogy. It ended because the tools failed, the ombudsman started ruling against institutions, and the students most likely to be wrongly flagged turned out to be the ones paying the bills."

That last point has a paper trail. Of cases published by the Office of the Independent Adjudicator in July 2025, three out of four were upheld or partly upheld against the university. And the students a detector is statistically most likely to flag — non-native English speakers — are drawn heavily from the same group universities most need financially: international students make up 24% of all UK higher education students and 51% of postgraduates, at a time when 43% of English universities are forecasting deficits. A tool that disproportionately accuses your most commercially important students of cheating was never going to survive contact with a complaints process.

If you're ever flagged

A detector score is not evidence on its own — treat "it came back 80% AI" as the start of a conversation, not a verdict. Ask explicitly to see the full basis for the accusation before any misconduct meeting, not just the percentage.

Don't reach for a "humaniser" or paraphrasing tool to get a lower score if you're worried. The same research that discredits detectors shows their accuracy gets worse, not better, once text is run through evasion tools — which just adds noise to an already unreliable process, on both sides.

What replaces itProcess, not a percentage

The redesign researchers and institutions converge on isn't "no rules" — it's moving the evidence from a single probabilistic score to something you actually produced over time: staged submissions, drafts, oral defences of your own work, and AI-literacy built into the task itself rather than policed after it's handed in. Specht and Saunders frame it plainly: "this is a reframe, not a capitulation" — the question moves from whether you produced text unaided to whether you can evaluate, critique and responsibly integrate AI-assisted work into your own thinking.

None of this makes AI use in coursework risk-free, and none of it means detectors have vanished everywhere. It means the single number a piece of software spits out is no longer something a university — or you — should treat as the whole story.

✦ End of issue ✦

Get the next one in your inbox

One AI tool, one prompt, three quick wins, every week. Free, and under 3 minutes to read.

Subscribe Free →