Integrity policies written for invigilated halls fail at 1 a.m. in a shared internet café. This study tests an AFU disclosure-and-artefact policy: students may use AI tools if they declare them and submit process notes that a marker can audit. Viva samples showed that students who could not explain a method they had ‘written’ were caught more reliably than by detector scores alone. Detector scores were unstable on African English and French academic prose, producing false positives that damaged trust. The University of Edinburgh should ban undetectable-AI panic and require oral or process evidence on high-stakes work. Integrity is a teaching design problem for a continental digital campus, not a procurement problem for a scanner.