Sampled vivas are pivotal in combating AI cheating

Last week’s revelation that the proportion of UK students using generative AI tools for assessments has jumped from 53 to 88 per cent in one year was yet another wake-up call for those of us worried about how to protect standards in assessment.

Higher education is often seen as a credentialing system: students invest in degrees that open doors to career opportunities. Yet if assessment defines a degree’s value, what happens when technology challenges our traditional modes of assessment? Generative AI forces universities to rethink not just how they assess learning but how they ensure meaningful student engagement.

For centuries, oral exams were standard, requiring students to defend their knowledge through debate. Today, viva voce assessments remain central to PhD examinations, testing critical thinking and ownership of ideas – skills equally valued in the workplace. Yet as undergraduate numbers grew in the 19th century, written exams were adopted as the only scalable solution for assessment, allowing remote evaluation and offering a verifiable record for moderation and quality assurance.

The integrity of that model, however, is increasingly questionable. Contract cheating was already allowing students to bypass learning and surrender both their authorial voice and intellectual engagement with their studies. For students who feel compelled to engage in such practices, generative AI has drastically lowered the barriers to entry. More fundamentally, the possibility of automating written work calls into question the value of written assessments in university credentialing.

To be fair, universities have responded swiftly since ChatGPT’s release in late 2022. Many introduced AI policies addressing assessment design. In the UK, for instance, UCL’s three-category framework (introduced in 2023) classifies assessments into those where AI is prohibited, permitted in an assistive role, and integral to the task. But while these policies clarify acceptable use, they do little to ensure students actively engage with learning.

Unlike traditional “verbatim” plagiarism, generative AI produces bespoke text, making detection difficult. Some tools claim to identify AI-written text, but their accuracy is questionable. AI’s tendency to hallucinate sources and make mistakes is often touted as a telltale sign, but humans do the same and it is becoming rarer as AI models improve.

Reverting to traditional in-person written exams ensures integrity, but relying on them exclusively is short-sighted. We should not throw away the inclusivity that has resulted from the Covid-era diversification of assessment, embracing coursework and take-home, open-book and online exams. Besides, employers won’t penalise AI use; graduates who integrate it effectively will have an advantage in the workplace. Universities must therefore prepare students to engage with AI critically and responsibly.

All of which brings us back to the viva. Many UK universities retain the option for viva voce assessments, but primarily for investigating suspected academic misconduct in written assessments. Yet oral assessments offer more than deterrence of misconduct, shifting the focus of assessment from policing AI use to assessing comprehension, ensuring students engage with ideas.

Discussion-based learning is already embedded in higher education through small-group tutorials and seminars. Nowhere is this more evident than the Oxbridge system, where students articulate their understanding and address gaps with tutors well before final assessments. However, these models are resource-intensive and difficult to scale.

Some institutions are reintroducing vivas into undergraduate assessment to uphold integrity. The University of South Australia, for instance, has reportedly replaced written exams with structured oral assessments in some subjects, significantly reducing misconduct. But for most universities, adopting oral assessment would be near impossible without overburdening staff – especially in an era when student-staff ratios are only growing as funding pressures force redundancies.

A practical solution would be to adopt sampled viva moderation of written assessments. Those students selected – perhaps 10 per cent of the total – would be informed shortly beforehand, ensuring they prepared by engaging meaningfully with the material.

Building on existing moderation processes, whereby a subset of student work is reviewed for marking consistency, oral sampling would be presented to staff and students not primarily as an AI detection tool but an opportunity for students to demonstrate comprehension and the ability to explain their own ideas. It would encourage them to take ownership of their work by shifting assessment from a one-way submission to a meaningful dialogue.

However, if a viva revealed a significant gap between a student’s written work and their demonstrated knowledge, institutional responses would be merited. Further discussion would be needed on how to handle such cases within existing policies while maintaining fairness, academic integrity and student support. That would include establishing clear guidelines for extenuating circumstances and reasonable adjustments (flexible scheduling, for instance, or alternative formats, such as recorded responses). Those eligible should include students with disabilities, language barriers or anxiety-related conditions.

Generative AI has transformative potential in education. Early findings from a World Bank programme in Nigeria suggest that an AI tutor could help students achieve the equivalent of two years of learning gains in a six-week intervention. In another large-scale trial, students working with AI-supported tutors were more likely to master topics. In both cases, thoughtful, educator-driven interventions are essential.

Rather than merely reacting to new technology, universities must set the standards for its responsible use. We must ask ourselves whether we are assessing text production or genuine understanding. Sampled viva voce moderation can help ensure that it remains the latter. 

Duncan Brumby and Anna Cox are professors of human-computer interaction at UCL. Advait Sarkar is an affiliated lecturer at the University of Cambridge, and an honorary lecturer at UCL. Sandy Gould is an academic in the School of Computer Science and Informatics at Cardiff University.