Detection-based approaches to AI cheating have one structural weakness: they require the tool to outpace the model generating the work, and that race is already lost. The lecturer who announced the viva requirement was not describing a policy innovation — they were describing the only remaining leverage point an educator has, which is the moment a student must speak . A student who submitted AI-generated prose and cannot name the source they supposedly consulted has been caught not by a detector but by their own silence.