A statistic that gave away the scale of the problem
The Universidad Nacional Autonoma de Mexico ran its first fully online undergraduate entrance exam in May and June 2026, with roughly 158,727 applicants sitting the test across three separate weekends. When results came back on July 17, one number stood out immediately: 16.3 percent of exams scored perfect, compared with a 3.5 percent average across the five prior admissions cycles from 2021 through 2025. A jump of that magnitude on a high-stakes exam does not happen through improved teaching or a lucky applicant pool.
UNAM's monitoring systems, combined with the score anomaly, pointed investigators toward phone use during the exam, unauthorized outside assistance, and suspected identity substitution, meaning someone other than the registered applicant may have taken all or part of the test for them. The university has kept the specific AI tools involved undisclosed so far, while the broader pattern matches what proctoring researchers elsewhere have flagged repeatedly: remote assessment without in-person controls is precisely where generative AI does the most damage to exam integrity at scale, and it does that damage quietly until someone checks the aggregate numbers.
The response was fast and unusually decisive
UNAM moved quickly once the anomaly was confirmed. A technical committee formed on July 27 to evaluate the scope of the problem, and by August 2 the university had announced its remedy: roughly 3,175 exams, about 2 percent of the total, were nullified outright, and a broader group faces mandatory in-person retesting. That broader group includes applicants who were already preliminarily admitted based on the compromised exam, plus rejected candidates whose original scores matched or exceeded the lowest passing threshold recorded across the previous five admissions cycles.
Rector Leonardo Lomeli addressed the fallout directly: 'I offer an apology to the applicants who were preliminarily admitted, but this measure is necessary to guarantee transparency, provide certainty, and ensure access to the university is equitable.' That is a rare instance of an institution publicly absorbing the reputational cost of its own exam design failure rather than quietly patching the process and hoping nobody outside the university noticed the scale of what happened. Most institutions facing a comparable finding would be tempted to handle it internally and adjust future exams silently, which makes UNAM's public accounting an unusually direct test case for how transparent an admissions office should be when its own process is the point of failure.
Retesting nearly 60,000 people is not a small operational lift
The retesting population is enormous by any standard: reports put the number of applicants who must sit the exam again at close to 60,000, once accepted and borderline-rejected candidates are combined. Running an in-person exam at that scale, on short notice, for an institution that only just discovered it could not trust its own online format, is a logistics problem most universities never have to solve in a single admissions cycle. UNAM filed a formal complaint with authorities alongside the retest announcement, signaling the university considers this a matter for investigation beyond its own internal remedy.
Student reaction has been predictably split. Some affected applicants are demanding the entire exam be scrapped and repeated in full, arguing that partial remedies still leave uncertainty about who cheated and who did not. Others want their original results honored, arguing they should not be punished for a system failure that was UNAM's to prevent, not theirs to answer for individually. Campus protests followed the announcement, and the university has promised further updates as the technical committee's investigation continues into the fall term.
Why anomaly detection beat the proctoring software
The detail most worth studying here is the detection method itself, more than the scandal's size. UNAM caught this by noticing that the aggregate score distribution no longer looked statistically plausible against five years of prior data, rather than through proctoring software flagging individual sessions in real time. That population-level statistical approach turned out to be a fundamentally more effective detection strategy than the session-by-session monitoring most remote exam platforms are built around today.
For any institution running high-stakes assessment remotely, whether that is a university entrance exam, a professional certification, or an internal corporate assessment, the UNAM case argues for building population-level statistical monitoring alongside individual proctoring, not instead of it. A single AI-assisted test-taker is hard to catch in the moment. A sudden five-fold jump in perfect scores across 158,000 exams is not, and it took UNAM roughly three weeks from result release to public announcement once someone looked at the distribution instead of just the individual sessions.
What this means beyond Mexico
UNAM is one of the largest universities in the world, and its shift toward fully online entrance exams was itself a modernization bet, one that just produced the largest known academic integrity failure tied to that format to date. Every institution considering a similar move, expanding a high-stakes assessment to a remote or online format to cut costs or widen access, now has a concrete data point on what can go wrong and how visible the failure becomes once it does.
The lesson here is to treat statistical baseline monitoring as a mandatory control before scaling any high-stakes exam online, built in from the start rather than bolted on once results already look suspicious. UNAM had five years of prior score distributions to compare against, which is exactly the kind of historical baseline every institution running recurring assessments should be building now, before their own version of this story becomes public.
