How Universities Are Reinventing Exams in the AI Era

Universities across North America, Europe and Australia are quietly dismantling the take-home essay. Between January and May 2026, at least 47 institutions — including the University of Sydney, KU Leuven and the University of California system — have announced revised assessment policies that restrict unsupervised written submissions, restore in-person exams or introduce mandatory oral defenses. The trigger is no longer ChatGPT alone, but a generation of multimodal AI agents that can draft, edit and cite at near-doctoral level. Faculties are scrambling to redefine what an exam measures when the answer is a prompt away.

The shift matters because assessment is the backbone of accreditation. If a degree no longer certifies that a student can think independently, employers, regulators and ranking agencies start asking uncomfortable questions. The current wave of reforms is less a pedagogical upgrade than a defensive move to protect the credibility of the diploma itself — and it is forcing universities to decide, often in a single semester, what they actually want graduates to know.

The context: how AI broke the take-home essay

The take-home essay, the dominant assessment format in the humanities and social sciences since the 1970s, was designed for a world where research took days and writing took hours. According to a 2026 report by the Higher Education Policy Institute, 88% of UK undergraduates say generative AI can now produce a passable first draft of a 2,000-word essay in under five minutes.

That gap between the time a task assumes and the time it actually requires has hollowed out the format. Professors at the University of Cambridge, in a working paper circulated in March 2026, described unsupervised essays as «assessing access to compute, not cognition». The complaint is echoed in faculty senates from Toronto to Melbourne.

Detection has not closed the gap. After a series of high-profile false-positive cases — most notably a 2025 lawsuit filed by a Texas A&M graduate student wrongly flagged by an in-house detector — institutions have grown wary of relying on probabilistic tools. The result is a pivot away from policing and toward redesign.

What the new exam looks like

The emerging template combines supervised in-person components with oral or applied tasks. A March 2026 EDUCAUSE survey of 412 institutions in 31 countries reported that 64% of respondents had reintroduced or expanded handwritten exams during the 2025–2026 academic year, while 41% had added oral defenses to courses where none existed before.

At KU Leuven, first-year law students now sit a 30-minute viva for every written assignment over 1,500 words. The defense, conducted by two examiners without AI access, probes the student’s reasoning rather than the text itself. Comparable formats — sometimes branded «process portfolios» or «authentic assessment» — are spreading through engineering and medical schools in continental Europe.

In the United States, the University of California issued system-wide guidance in February 2026 encouraging «AI-integrated» assessment: students may use generative tools, but must submit a prompt log, an annotated draft history and a reflective commentary. The model treats AI as a calculator rather than a ghostwriter, a framing also adopted by the University of Helsinki and the Australian National University.

The data on student behaviour

Surveys from 2025 and 2026 converge on a clear pattern: students use AI heavily, disclose it rarely and resent ambiguous rules. A January 2026 Pew Research Center poll of 4,200 US college students found that 71% had used a generative tool for graded work in the previous year, while only 14% had been explicitly told by their instructor whether such use was permitted.

The disclosure gap is the policy crux. Without clear rules, students default to silence, and faculty default to suspicion. The Russell Group’s 2026 guidance attempts to break the deadlock by requiring instructors to publish, for every assessment, a traffic-light statement: red (no AI), amber (AI permitted for specific tasks) or green (AI integral to the assessment).

Independent studies suggest the framing changes behaviour. A randomised trial at the University of Edinburgh, published in Assessment & Evaluation in Higher Education in May 2026, found that students given an explicit amber-zone rubric were 38% more likely to declare AI use than those given a generic honour code, with no measurable drop in grade quality.

«We are not going back to the typewriter. The question is whether we can design tasks where the student’s contribution remains legible, even when the draft passes through a model.»

The push-back from inside the faculty

Not every department is convinced. A faculty survey at the University of Toronto, released in April 2026, found that 46% of humanities instructors believed oral defenses were unworkable at scale, citing class sizes above 200 and adjunct staff with no contractual time for vivas. The economics of the reform, critics argue, fall on the least protected labour.

There is also a pedagogical dispute. Proponents of the «AI-integrated» model, including several computer-science departments at ETH Zürich and Carnegie Mellon, argue that banning generative tools trains students for a workplace that no longer exists. Opponents, often based in literature and philosophy, counter that fluency in unaided writing is itself a skill worth certifying — and that outsourcing first drafts to a model erodes the cognitive habits the degree is meant to cultivate.

The split mirrors a broader debate about AI literacy in higher education, one that StudyVerso has tracked through reporting on tools and study workflows, including a recent comparison of Claude 4.7 vs ChatGPT-6 for studying in 2026. The pattern across institutions is the same: technology adoption is racing ahead of pedagogical consensus.

A snapshot of the new formats

What it means for students and the sector

For students, the immediate effect is a heavier in-person workload and a steeper learning curve on disclosure norms. For institutions, the bet is on whether the cost of redesign — estimated at €1,800 per course by a 2026 League of European Research Universities working paper — can be absorbed without raising tuition or shrinking enrolment.

Beyond the universities themselves, the ripple effects reach into the EdTech market. Demand for AI-aware assessment platforms has surged, while a wave of consumer study apps — from established players like Quizlet and Notion AI to newcomers such as Modo Cheto — are repositioning around supervised practice and oral-defense rehearsal rather than essay generation. Investors who funded the homework-AI boom of 2023 are now asking how those tools fit into a classroom that has stopped grading homework.

Regulators are watching. The European Commission’s Joint Research Centre is preparing a 2027 report on assessment integrity, and the US Department of Education has signalled that accreditation reviews will begin examining AI-assessment policies as part of routine evaluations. Even publishers of standardised tests, including ETS and Pearson, have announced redesigns of their writing components for the 2027 testing cycle. The pressure to reinvent exams is no longer confined to seminar rooms.

The unresolved question is whether the reforms produce graduates who can think with AI, or merely graduates who can perform without it for two hours in a sealed lecture hall. Universities are picking a side, often quietly, often before the evidence is in. The next cohort will be the first to take the test.