AI detectors misflag international students, HEPI warns

TL;DR:

• Stanford researchers found seven widely used AI detectors misclassified over 61% of essays by non-native English speakers as machine-generated, while performing near-perfectly on native writers.
• Of four Office of the Independent Adjudicator cases published in July 2025, three involved international or second-language students and three went against the university.
• Waterloo, Curtin, UCLA and UC San Diego have dropped Turnitin’s AI detection; no UK university has followed.

Writing in a HEPI blog, Strathclyde PhD candidate Brendal Aformeziem sets out a straightforward mismatch: detection tools flag the linguistic patterns of second-language writing — shorter sentences, narrower vocabulary, less idiomatic phrasing — as evidence of machine generation. The tool, as the piece puts it, cannot distinguish artificial intelligence from a student working in a language that is not their own.

The reliability picture is poor even setting bias aside. A large-scale evaluation across 805 samples found average accuracy of 39.5% on unmodified AI text, falling to 17.4% once students applied simple evasion techniques, with some tools misclassifying half of all human-written work. Turnitin’s own chief product officer has acknowledged that scores are probabilistic and should not be the sole basis for misconduct proceedings.

What the ombudsman found

The OIA cases show the consequences. In one, a student who explained they had used Google to find synonyms had that recorded as an admission of using AI to paraphrase; the adjudicator found the university had never considered whether Turnitin might perform less reliably for non-native speakers. In another, a flagged student was not shown the evidence before their viva and had their use of Grammarly dismissed without explanation. A third involved an autistic student accused, penalised and eventually cleared — for the second time, having previously been flagged and found to have written the work themselves.

The financial angle

International students are 24% of UK higher education enrolment and 51% of postgraduates, at a point when 43% of English universities forecast deficits. Institutions are financially dependent on the cohort their detection systems flag most often.

Looking forward

The policy ask is specific: suspend detection as primary evidence pending independent validation, have the QAA and OIA issue joint guidance that scores alone cannot ground disciplinary action, and fund assessment redesign — staged submissions, oral defences, embedded AI literacy. For UK universities, the exposure is now reputational and legal as well as educational, and the overseas institutions that acted did so on evidence that has only strengthened since.