What should a testing program do when AI writes the candidate's answers?

The numbers say detection has already lost

The question is no longer whether candidates use AI. In 2026, 95 percent of UK undergraduates use AI, and 94 percent use it for assessed work, according to the Higher Education Policy Institute. Twelve percent paste AI text straight into their submissions. The program that still treats AI use as a rare event is policing a world that no longer exists.

The scale is the story. Turnitin found that 54 percent of Australian submissions between October 2025 and April 2026 showed some AI use, and 10 percent were over 80 percent AI. Around 30 percent of US STEM dissertations in 2024 contained AI text, against under 2 percent in 2023. The proven cases follow the same curve: about 7,000 in the UK in 2023-24, or 5.1 per 1,000 students, up from 1.6. Ireland went from 178 students in 2023/24 to 577 in 2025/26.

The enforcement gap is what matters. Proven cases are a fraction of actual use, because most AI-written work is never detected. In a University of Reading test, 94 percent of AI-written exam submissions went undetected. The program that relies on detection is relying on a tool that misses nineteen out of twenty.

Why the detector fails in both directions

The false negative is the bigger problem by volume, but the false positive is the bigger problem by harm. The University of Waterloo switched off Turnitin's AI detection in September 2025 after it flagged human text as 100 percent AI.

The bias is measurable. A Stanford study in 2023 found that seven detectors wrongly flagged 61 percent or more of essays by non-native English speakers. A detector that punishes the students who write most carefully in a second language is not a detection tool. It is a liability.

The asymmetry is the point. A false negative lets a cheater through. A false positive accuses an innocent student, and the accusation follows them. A program that treats a detector score as proof will eventually accuse the wrong person, and the appeal will cost more than the exam was worth.

The assessment has to change, not the policing

The Office for Students pushes paper, in-person, and oral exams, because those formats are hard to fake with a language model. A handwritten paper exam in a hall cannot be written by a chatbot, and an oral exam tests whether the candidate can defend the answer, not whether they can produce it.

The change does not have to be all-or-nothing. A program can keep the written exam and add a short oral defence of a random sample, or a supervised in-person component for the high-stakes paper. The point is to test the thing AI cannot do: think in front of a person, under pressure, with no model to ask.

What counts as evidence now

A detector flag is a signal, not evidence. It starts a review: the marker reads the work, compares it with the candidate's other work, and asks the candidate to explain it.

The review has to be human, and the human has to be the one who decides. Hybrid proctoring works the same way: AI flags plus a human reviewer beat either alone, because the machine is fast and the human is accountable.

Publish the rule. A program that tells candidates the detector is a signal, the review is human, and the oral defence is the backstop gets fewer appeals and better exams, because the candidates know what the standard is before they sit down.

Keep reading: what a Turnitin AI score actually proves, how AI flags and human reviewers split the work, and what a second device does to an exam cover the nearby cases in this series.