England's University Regulator Says AI Has Made Fair Coursework Assessment "Almost Impossible" and Puts Fines on the Table

England's higher education regulator has formally concluded what many lecturers already suspected: AI has broken the standard model of coursework assessment. The Office for Students (OfS), publishing jointly with Advance HE on November 11, says generative AI has made the fair evaluation of unsupervised assignments "almost impossible" and describes unchecked AI use as a generational risk to the credibility of English degrees. Its prescription is unambiguous — universities should dramatically increase secure, AI-free assessment such as in-person exams, live presentations and oral defenses.

The evidence base is substantial. Researchers conducted work from June to September 2026 across at least 127 institutions, gathering input from 244 higher education staff and 91 students or student representatives through field events, 64 written survey responses and 22 interviews. Financial Times reporting cited survey figures indicating more than 90 percent of students use AI to help with coursework.

The findings are unusually blunt for a regulatory document. Students reported that peers complete assignments in minutes with large language models and score about the same as students who worked independently — and that some cannot recall the basic content of what they submitted. Staff described batches of submissions with near-identical writing styles and identical reference lists, with remotely submitted dissertations particularly exposed. One staff member told researchers that in some fields they had "almost lost the ability to evaluate students effectively."

The report's most consequential conclusion is that detection is a dead end. Human markers cannot reliably identify AI-generated text, and detection tools do not reliably detect AI use either, so policy statements and after-the-fact misconduct procedures are no substitute for well-structured assessment. OfS strategy and delivery director Josh Fleming said detection software is not reliable enough to be used in disciplinary action, and that "a better approach is to design assessment so that cheating is not even an option." Test results, where used, should count as at most one piece of evidence — suspicion alone should never lower a grade.

The guidance steers universities toward "secure assessment": in-person exams, supervised presentations and the revival of viva voce oral defenses, with some institutions asking students to critique or edit AI-generated content or document their research process as assessable work. The report also flags the cost reality — one-to-one and live assessment is expensive — and warns that without proper adjustments it can disadvantage disabled students. Universities are still expected to teach AI literacy, since students see it as essential for employment.

Enforcement now has teeth. Under the quality conditions that apply to registration, institutions that cannot demonstrate assessments which "genuinely measure student knowledge and ability" face specific registration conditions, monetary penalties, or suspension of student-funding or OfS grant eligibility. The Department for Education backed the message, calling on universities to take the issue seriously and on the regulator to enforce existing standards.

England is now the clearest test case for what post-AI assessment looks like at system scale: stop trying to catch the cheating, and redesign the exam so the cheat does not work. The open question is whether institutions can absorb the staffing costs of live assessment — and whether other regulators, watching from a distance, follow the same path of putting funding on the line.