Why universities are bringing back the oral exam — and how classrooms can scale it

Something old is happening in higher education. At Cornell, biomedical engineering students now sit an “oral defence” with their instructors. The University of Pennsylvania describes a large-scale shift toward in-person assessment and trains faculty in oral examination techniques. NYU reports more presentations, more mandatory office hours, more professors asking students to talk through their work. At UC San Diego, an engineering professor has spent three years researching how to run oral exams at scale, and other universities now invite her to train their staff. In the UK, university law schools have reinstated invigilated exams after finding themselves buried in essays that plainly weren't written by the students submitting them.

The viva voce — the “living voice” exam — predates written examination by centuries. In Norway and Denmark, it never went away. Elsewhere, it survived mostly in doctoral defences. Now it's being dusted off across entire undergraduate programs, and the reasons why matter to every school, not just universities.

Why the shift is happening

Polished writing no longer proves understanding. In a recent Inside Higher Ed survey, 85% of college students said they had used AI in their courses, and a quarter admitted using it to complete assignments. The pattern professors describe is consistent: written submissions have never looked better, and students have rarely been less able to discuss them. A flawless document now tells you a document was produced. It doesn't tell you who did the thinking.

Detection failed. The first institutional response to ChatGPT was software that promised to spot AI writing. It hasn't worked — detection tools remain unreliable, producing both false alarms that accuse honest students and false negatives that wave through generated work. Institutions that leaned on detection ended up with piles of misconduct allegations that were difficult to prove and corrosive to pursue. Assessment design is replacing detection: rather than trying to catch generated writing after the fact, educators are choosing formats where the student's own understanding is the thing being observed.

The deeper reason: it was never really about cheating.The most striking thing in the coverage of this shift is how educators themselves frame it. One University of Pennsylvania professor put it plainly: the concern isn't policing misconduct — it's that students are losing skills and cognitive capacity when tools do their thinking for them.

That concern now has hard evidence behind it. A field experiment with nearly 1,000 high school students, published in theProceedings of the National Academy of Sciences, gave students GPT-4 access during maths practice. Performance during practice jumped 48%. But when the AI was taken away for the exam, those students scored 17% worse than classmates who never had AI at all. The researchers found most students had simply asked for answers and copied them — and, notably, the students didn't perceive that their learning had suffered. Early neuroscience work points the same direction: an MIT Media Lab study (still a preprint, so treat it as preliminary) found essay-writers using ChatGPT showed the weakest brain connectivity of any group, and most couldn't quote from essays they had submitted minutes earlier.

The pattern across all of it: when a tool does the generative work, the learning doesn't happen. And the student is usually the last to know.

Why explanation works as assessment

Asking a student to explain a topic does two jobs at once.

It verifies. As one Cornell professor told reporters, “You won't be able to AI your way through an oral exam.” A student explaining a concept in their own words, in their own voice, is demonstrating understanding directly — no detector required, because the assessment is the detection.

It teaches. Decades of learning research show that constructing an explanation is one of the most effective things a learner can do. Students who prepare to teach material learn it better than students who study for a test; students who generate explanations outperform students who reread and summarise. (We've written up that evidence separately — see why teaching a topic back strengthens learning.) The oral exam isn't just harder to fake. It's better pedagogy.

The problem: vivas don't scale

Here's where the university solution hits a wall in schools. A professor can examine a seminar group of twelve. A secondary teacher with five classes of thirty cannot run 150 individual vivas a term — the timetable doesn't exist. Live oral assessment also disadvantages students who need time to compose themselves, and it leaves no artefact to moderate, revisit, or share with parents.

The format is right. The logistics are impossible. That gap is exactly what a recorded teach-back closes.

The oral exam, made scalable

A Fishbole teach-back is an oral exam that fits a real school week. The teacher sets the topic. Each student researches it, structures their slides, and records themselves explaining it — capped at five minutes, in their own voice, with no AI writing tools, no avatars, no generated scripts — then shares it with their teacher via a short link. The teacher reviews on their own schedule, at five minutes maximum per student, seeing exactly what each student understands and where the misconceptions live.

Everything universities want from the viva survives the translation: the student must explain, understanding is visible, and there is nothing to run through a detector. And the things that made vivas impractical fall away: no scheduling, no panel, a rewatchable record, and a format that lets anxious students re-record until they're happy — submitting privately to the teacher rather than performing live in front of a room.

Universities rediscovered the oldest assessment there is because it's the one AI can't sit. Fishbole exists so that every classroom can use it — thirty students at a time, five minutes each.

Sources and further reading: the Wharton/PNAS field experiment on generative AI and learningand its plain-English summary; the MIT Media Lab Your Brain on ChatGPT preprint; Times Higher Education on the return of in-person assessment; and the research behind teach-back learning in our evidence write-up.