Generative artificial intelligence is forcing universities to reconsider what a degree should prove. As students use tools such as ChatGPT in coursework, institutions are weighing stricter, supervised examinations against assessment models that treat AI as part of the professional environment graduates will enter.
The shift reflects a basic tension. Employers increasingly expect graduates to understand and use AI responsibly, yet universities must still establish that students can reason, write and solve problems themselves. A policy built only around detecting prohibited use may protect the appearance of academic integrity without showing how learning actually took place.
From policing outputs to examining judgment
A 2026 survey cited by the Higher Education Policy Institute and Kortext found that 94% of UK undergraduates used generative AI for assessed work, up from 88% the previous year. The finding does not establish that students are outsourcing their assignments: AI use can include brainstorming, editing, translation or research support. But it does show that blanket assumptions about work being completed without AI are increasingly unrealistic.
HEPI argues that universities should focus on “verifiable judgment”: the ability to assess evidence, reason through uncertainty and take responsibility for conclusions. Its recommendations include oral examinations, iterative projects, supervised problem-solving and portfolios that show development over time.
That approach could give lecturers stronger evidence than a single essay submitted at the end of a course. Drafts, reflections and brief conversations can reveal whether a student understands an argument and can defend choices made during the work. They may also reward the learning process rather than merely the polish of a final product.
Why some universities are bringing back closed exams
Other institutions are moving in the opposite direction for at least part of the curriculum. The University of Bath has announced a framework distinguishing between open assessments, where generative AI may be allowed or required, and closed assessments, which are time-limited, invigilated and held in person from the 2026–27 academic year.
The case for closed examinations is straightforward: some knowledge and skills must be demonstrated without external assistance. Medical, legal, technical and safety-critical fields may require students to recall information, interpret evidence quickly or make decisions under controlled conditions. In-person assessment can also reassure students that grades are based on comparable rules.
Yet a return to traditional exams has costs. Timed tests can privilege recall over collaboration, creativity and sustained problem-solving. They can also disadvantage students with disabilities, caring responsibilities or severe exam anxiety unless institutions provide meaningful accommodations. More invigilation may address one form of misconduct while doing little to prepare students for workplaces where AI tools are widely available.
Employers want AI literacy—but not blind reliance
Advice on the 2025 graduate labour market from St John’s University highlighted working knowledge of AI tools as a skill employers increasingly seek. That demand changes the question universities face: graduates may need to show not only that they can produce an answer, but that they can verify an AI-assisted answer, identify errors and explain the limits of the technology.
For students, inconsistent rules remain a practical problem. One assignment may permit AI for idea generation, another may ban it entirely, while a third may require disclosure without defining what counts as acceptable assistance. Universities that fail to state expectations clearly risk turning ordinary study practices into disciplinary hazards.
Several higher-education organisations are therefore calling for assessment instructions to identify explicitly when AI is prohibited, permitted or integral. The University of Manchester, for example, approved a policy assigning every summative assessment one of four AI-use categories from September 2026.
A redesign, not a technological arms race
AI-detection software is unlikely to settle the dispute. Detection can produce false accusations, and a student who submits original work may have limited ability to challenge an opaque automated judgment. Designing tasks that require personal reflection, staged evidence and live explanation may be more reliable than trying to determine whether a finished paragraph “sounds” machine-generated.
The transition will still require resources. Oral defences and detailed feedback take staff time; portfolio assessment demands coherent digital systems; and unequal access to paid AI tools could widen existing gaps. Universities will need to provide training and, where AI is required, ensure that students are not disadvantaged by cost or disability.
The emerging compromise is neither unrestricted automation nor a complete return to paper exams. It is a clearer division of purpose: assessments should state when students must work unaided, when AI is a legitimate professional tool, and how they must demonstrate oversight. That could make degrees more credible while preparing graduates for a labour market in which knowing how to use AI is only the beginning.
Sources: Higher Education Policy Institute; Times Higher Education; St John’s University; University of Manchester.