UK Universities Are Abandoning AI Detectors, Not Fixing Them
HEPI research and a seven-university survey show UK institutions retreating from AI detection tools toward oral exams and staged coursework, after evidence the tools misfire most on international students.
The tools are being switched off, not upgraded
For two years, UK universities responded to generative AI by tightening detection. Now several are doing the opposite. According to reporting compiled by Resultsense and a July 2026 Higher Education Policy Institute (HEPI) analysis, a number of institutions have restricted or dropped AI-detection tools such as Turnitin's AI indicator, GPTZero and Copyleaks, citing accuracy concerns rather than switching to a better-calibrated alternative.
This is not a story about one flawed product being replaced by a sounder one. It is a retreat from the entire premise that a percentage score can settle whether a student cheated.
Who the detectors actually flag
The HEPI paper, titled "Catching the wrong students: AI detection, international students and the fairness crisis in UK universities", makes the equity problem explicit. It cites Stanford research finding that seven widely used AI detectors misclassified more than 61 percent of essays by non-native English speakers as machine-generated, while performing close to perfectly on writing from native speakers. Formulaic sentence structure and simpler vocabulary, common features of second-language academic writing, appear to read as "AI-like" to these systems.
HEPI's authors point out that this is not a marginal quirk. International students make up 24 percent of UK higher education enrolment and 51 percent of postgraduate students, at a moment when 43 percent of English universities are forecasting budget deficits. Institutions are, in effect, financially dependent on the cohort their own detection software flags most often for misconduct.
A technical weakness in a piece of software becomes a question of procedural fairness and natural justice the moment it is used to open a disciplinary case.
What students report
HEPI's wider student survey work found that 75 percent of students who use AI say they are stressed about the possibility of a false accusation, rising to 81 percent among international students. That anxiety exists independently of whether a student has done anything wrong. Several qualitative responses in the underlying research describe students becoming self-conscious about writing "too well," softening or roughening their own prose to avoid triggering a detector.
Meanwhile, the scale of actual undisclosed use is not small. A survey led by researchers at Edinburgh Napier University, covering more than 6,600 students across seven UK universities, found that 32 percent admitted some unpermitted AI use in assessed work. Detection tools are not stopping this; they are producing false accusations against some students while missing a third of the cohort who used AI without permission.
The alternative on offer: redesign, not surveillance
HEPI's recommendation, echoed by researchers at MIT's Sloan School of Management cited in the same coverage, is to stop treating detection scores as the primary evidence base in misconduct proceedings and instead redesign what students are asked to produce. In practice this means staged submissions with visible drafts, process tracking through version history, and oral defences of written work, an approach that makes the origin of the ideas verifiable without needing to interrogate the prose itself.
This is a harder sell for large modules with high student-staff ratios than a same-day percentage score. But it is consistent with where JCQ guidance already points UK schools, and with the reasoning behind the US ruling this year in which a court found in favour of a student wrongly accused after Turnitin returned a "100 percent AI-generated" score on work that a separate detector scored at zero. Detection outputs are probabilistic; treating them as proof was always the weak point.
Why this matters below university level
Sixth-form and college staff preparing students for UK higher education should expect the assessment culture their leavers walk into to look different from the one current Year 13s were warned about. If universities are moving away from a single AI-detection score and towards staged, process-visible, orally defended work, that has implications for how post-16 providers build coursework habits now, particularly for students used to producing a single final draft the night before a deadline.
It also raises the stakes for UK schools' own detection practices. If HEPI and Edinburgh Napier's findings hold at scale, then any GCSE or A-level centre still treating a Turnitin-style score as standalone evidence of malpractice is relying on a method that better-resourced institutions are actively abandoning.
What to do
- Treat any AI-detection score, in a school or university, as a prompt for a conversation rather than a verdict. Corroborate with drafts, version history or a short viva-style discussion.
- If your institution still leans on detection software for high-stakes decisions, ask what its false-positive rate is for second-language writers specifically, not just its overall accuracy claim.
- Where possible, build staged submission and brief oral check-ins into coursework now, so students arriving at university are already used to showing their working rather than only their final text.
What to watch
- Whether other UK universities follow the pattern of restricting detection tools, and whether the Office for Students or JCQ issues formal guidance discouraging detector scores as standalone evidence.
- Whether assessment redesign, oral defences and process tracking prove workable at scale in large modules, or whether cost and staff time pull institutions back towards detection despite its flaws.
- Whether international student recruitment numbers and misconduct-case data are published together by any UK university, which would make the equity argument in HEPI's report testable rather than anecdotal.