Mexico's Largest University Moved Its Entrance Exam Online and the Marks Came Back Too Good

The National Autonomous University of Mexico ran the entrance examinations for its various programmes remotely for the first time this year. The results came back markedly higher than in previous years, high enough that the university treated the distribution itself as evidence, and candidates are being made to take the exam again on suspicion that artificial intelligence assistance was widely used.

The reasoning is statistical rather than forensic, and it is sound. A national admissions exam sat by very large numbers of candidates produces a distribution that barely moves from one year to the next. Cohorts do not become collectively better at algebra over twelve months. When the curve shifts across an entire intake, the honest explanation is almost never that the students changed, and the university does not have to catch anyone individually to know the conditions were different.

Remote invigilation is where the design failed. Proctoring software watches the machine the exam is on, the webcam feed and the room behind it. It has no view of a second device sitting outside the frame, which is all that is needed when the assistance is a chat window rather than a person. The controls were built for an era when cheating meant a hidden note or a knock at the door.

The cost lands on the candidates who did nothing. There is no method that separates a genuine high score from an assisted one after the fact, so the only remedy available to the university is collective, and every applicant sits the paper again regardless of how they earned the first result. Admissions in Mexico are a hinge point for a young person's prospects, and a second sitting under new conditions is not a neutral inconvenience.

Institutions everywhere are reaching the same conclusion at roughly the same time. High stakes assessment that can be taken anywhere cannot also be assumed to be unaided, which leaves supervised rooms, oral defence, or assessment designs where the model's help is assumed and the marking accounts for it. The convenience of examining at distance was real. It was also priced on a security assumption that no longer holds.