Historically, 3.5% of applicants to Mexico’s National Autonomous University scored 100 or higher out of 120 on the entrance exam. In 2026, 16.3% did. That roughly fivefold spike didn’t happen because Mexican teenagers suddenly got smarter over the summer. UNAM ran its undergraduate admissions exam fully online for the first time, trusted AI proctoring to keep roughly 160,000 candidates honest, and watched the score distribution warp beyond recognition. Now about 58,000 of those applicants must sit an in-person control exam that will override their original results entirely. The semester starts August 10.
Twelve Filters, None of Them Enough
UNAM’s two-layer security setup looked serious on paper — and fell apart in practice.
- Respondus LockDown Browser blocked other apps, copy functions, and web browsing during the test
- Territorium’s AI webcam monitoring flagged extra people on screen, visible phones, earphones, and candidates leaving frame — with each human supervisor watching roughly 150 applicants simultaneously
- What slipped through anyway: external monitors running ChatGPT off-camera, hidden earphones under hair, full identity substitution, and commercially sold real-time AI answer services
- The catch rate: UNAM annulled about 3,000 exams — roughly 2% — yet AI expert Raúl Rojas estimated statistically that nearly half of all 158,000 candidates cheated in some form
Scores at 110-plus jumped from a historical 0.9% to 5.5%. These anomalies clustered in the upper deciles, pointing toward organized, targeted cheating rather than any general improvement in applicant quality. UNAM’s own technical expert commission reviewed the data and called it unambiguous.
“Almost half of the students were cheating when taking the exam online.” — AI expert Raúl Rojas, according to The New York Times
UNAM paused admissions, convened its Comisión Técnica de Personas Expertas, and filed criminal complaints against companies selling fraudulent exam assistance. The chosen remedy is brutally analog: pencils, rooms, and proctors with actual eyes. Results from this in-person control exam will completely override the online scores.
58,000 People Who Mostly Did Nothing Wrong
The retake punishes honest applicants for a failure of institutional design.
The control exam catches not just admitted candidates but also rejected applicants whose scores met or exceeded historical cut-off thresholds — people potentially squeezed out because cheaters inflated the curve above them. It’s the academic equivalent of a sold-out concert canceled after bots grabbed half the seats. The honest fans absorb the inconvenience. Rector Leonardo Lomelí publicly apologized, calling the retake “necessary to give certainty and guarantee equity in access” to UNAM.
Ars Technica’s August 2026 coverage already frames this as a landmark AI-era cheating scandal — proof that generative AI tools can beat institutional monitoring at scale while that same monitoring cannot retroactively identify exactly who cheated. The lesson isn’t that remote exams are finished. It’s that deploying AI proctoring without robust statistical oversight is institutional wishful thinking. Calling it a security solution doesn’t make it one.
Specific dates and locations for the control exam remain undisclosed. Across Mexico, 58,000 people are studying again — some for the second time, and a few hoping no one looks too closely at the math.
