UNAM, Mexico's largest university, will make roughly 58,000 admitted applicants retake its entrance exam in person after an AI-proctored remote rollout produced scores that bore little resemblance to five years of prior results. Nearly 160,000 people sat the 120-question test from late May through early June 2026, the first year UNAM ran the exam entirely remotely using a lockdown browser and AI webcam proctoring. Top scores jumped so sharply that the university's own commission concluded the results could not stand.
The numbers are the story. Between 2021 and 2025, 3.5% of test-takers scored 100 or higher. In 2026, that figure hit 16.3%. At the top end, 0.9% of applicants scored 110 or higher over the prior five years; this year, 5.5% did — a more than fivefold increase in a single cycle. A shift that large in a stable, high-stakes admissions test does not happen through better preparation.
UNAM contracted Respondus for its LockDown Browser and Territorium for AI-driven webcam monitoring. Respondus disables printing, copying, web navigation, instant messaging, and hundreds of other functions once an exam begins. Territorium's system watched each webcam feed for a person swap, phones, earphones, or the applicant leaving the frame, escalating alerts to one human supervisor per 150 applicants.
Key facts
- 01UNAM will require roughly 58,000 admitted students to retake its entrance exam in person after AI-proctored remote testing failed to prevent cheating.
- 02Scores of 100 or higher on the 120-question test jumped from a 2021–2025 average of 3.5% to 16.3% in 2026.
- 03Top-tier scores of 110 or higher climbed from 0.9% to 5.5%, a more than fivefold increase year-over-year.
- 04The remote exam used Respondus LockDown Browser and Territorium's AI webcam proctoring, with one human supervisor per 150 applicants.
- 05Nearly 2% of exams were canceled for conduct issues, but statistical modeling suggested almost half of test-takers cheated.
That stack canceled nearly 2% of total exams for conduct issues — a rounding error against the score inflation the commission ultimately documented. UNAM itself acknowledged "the probability that a significant number of applicants may have received help." The remote setup created attack surfaces the software was not built to close.
Cheating tips circulated openly before the test window opened. Advice included positioning a second monitor outside the webcam frame to access ChatGPT or other models, hiding earphones under hair, and hiring proxies to sit off-camera. Because the exam was multiple choice rather than essay-based, the usual forensic tells for AI-assisted answers — pasted blocks of prose, characteristic phrasing — did not apply. Answers only had to be selected, not written.
“that almost half of the students were cheating when taking the exam online.”— Raul Rojas, AI expert
AI expert Raul Rojas, speaking to NPR, said statistical modeling suggested almost half of the students were cheating when taking the exam online. That estimate, combined with the score distribution, drove the commission's recommendation.
The commission — formally the Comisión Técnica de Personas Expertas para la Revisión del Proceso de Selección de Ingreso a Licenciatura para el Ciclo Escolar 2026-2027/1 — recommended a "control exam" administered in person. It applies not just to 2026 admits but to everyone who secured a place based on minimum successful scores in their program since 2021. Placement at UNAM now depends on the new test. The rector apologized to honest applicants who will have to prepare and sit for the exam again despite doing nothing wrong.
“necessary to give certainty and guarantee equity in access”— UNAM rector, Rector, Universidad Nacional Autónoma de México
The timing is tight. Classes were scheduled to begin August 10, 2026, giving UNAM roughly a week from the commission's announcement to run the retake, unless the school delays the semester. Details of the control exam had not been published as the recommendation landed.
Some caveats are worth stating. UNAM has not published the specific vectors it believes drove the score inflation, so the split between AI-assisted cheating, leaked question banks, proxy test-takers, and traditional cheat sheets is unknown. Multiple choice makes root-cause analysis harder, not easier. The 2% cancellation rate suggests the proctoring stack caught some behavior but was overwhelmed by the volume and creativity of workarounds.
The broader lesson for AI proctoring vendors is unforgiving. Respondus and Territorium built their systems for a world where the primary cheating threats were browser tabs, phones on the desk, and unauthorized humans in the room. A cheap second monitor angled away from the webcam, running ChatGPT on a separate device, defeats that threat model entirely. Every remote-testing product currently in market — for professional certifications, university admissions, and corporate compliance — faces the same gap. The economics of running a legitimate exam remotely, at scale, with cheap generative models sitting one glance away, are getting harder to defend. UNAM just paid the tuition for the rest of the industry.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
