AI Proctored 160,000 Exams. Half the Students Cheated Anyway — Here's What Developers Should Learn

In July 2026, Mexico's National Autonomous University (UNAM) administered its college entrance exam to roughly 160,000 applicants under AI webcam surveillance — and the technology failed so spectacularly that 58,000 students now must retake the test in person. Statistical analysis of the results suggests that approximately half the original test-takers cheated, while the AI proctoring system flagged roughly 2% of them. The fraud wasn't exposed by the software; it was exposed by a score distribution that had no plausible innocent explanation.

The Claim

UNAM deployed a combination of AI-powered webcam monitoring and LockDown Browser — software that restricts what a test-taker can access on their computer during an exam. The setup is representative of what the industry currently sells at scale. Vendors claim that combining computer vision, behavioral analysis, audio monitoring, and browser lockdown achieves suspicious-behavior detection rates of 90–95%. One industry guide puts the overall reduction in cheating from AI-supervised versus fully unsupervised remote exams at 96%. The AI proctoring market is projected to reach $9.17 billion by 2033, growing at 18.7% annually, and roughly 70% of educational institutions have adopted some form of the technology. By those numbers, UNAM had done everything a reasonable administrator would have done.

The university also had human supervisors available to respond to AI-generated alerts, though at a ratio of approximately one person per 150 applicants. The system was monitoring cameras, flagging eye movements, tracking gaze direction, detecting secondary devices — the standard toolkit. UNAM presumably chose this approach because in-person exams at this scale, distributed across a large and geographically dispersed applicant pool, carry their own serious logistics problems. Remote AI proctoring was the practical answer.

What We See

What the AI caught was approximately 2% of students cheating. What a statistical comparison against historical data caught was approximately 50%. The share of applicants scoring in the highest performance tier jumped from a five-year historical baseline of 3.5% to 16.3% in this sitting. That is not noise or year-to-year variance; it is a signal visible to anyone running a distribution comparison against prior cohorts. The fraud was ultimately detected not by the proctoring software but by the same kind of baseline anomaly analysis used in financial fraud review — the AI was incidental.

Our read is that this is the more important lesson, and one that proctoring vendors are structurally incentivized to obscure. The same industry guide that claims a 96% reduction in cheating also describes detection rates of 90–95% for suspicious behavior — but neither number appears to be measured against the actual prevalence of cheating under AI supervision. They are measured against fully unsupervised assessments, where essentially nothing catches anything. Compared to a verified real-world baseline like the one UNAM's score distribution just provided, those figures collapse.

The circumvention methods required no technical sophistication. Multiple reports indicate students used off-camera monitors running ChatGPT while appearing to look at their screen, hid earpieces to receive answers from confederates outside the room, and in some cases used proxy test-takers — someone else sitting the exam entirely. None of these attacks required defeating the AI. They simply worked around whatever the webcam could see. This is consistent with the broader 2026 picture: voice-activated AI tools accessed via earpiece, secondary devices kept below camera frame, and AI browser extensions running alongside the monitored window are all documented circumvention methods in current use, and none of them require a technical exploit.

The Brown University case, reported separately this summer by Inside Higher Ed, reinforces the pattern from a different angle. Professor Roberto Serrano's advanced mathematical economics course had a take-home midterm where the class average reached 96% — with nearly half the class scoring a perfect 100. When Serrano submitted the exam questions to ChatGPT, the outputs matched student submissions closely enough to make the source clear. When the final moved to an in-person format, the class average fell to 48%, and roughly a third of students either skipped the exam or dropped the course entirely. That is the controlled comparison the UNAM data cannot quite provide: the same students, same material, different monitoring environment. The result is hard to interpret any other way.

Where It Falls Short

The industry's detection claims are not fabricated — AI monitoring is measurably better than nothing, and it likely deters some portion of would-be cheaters who lack the nerve or logistics to work around a camera. But the specific claim of "95% detection accuracy" deserves scrutiny before anyone builds high-stakes policy around it. UNAM provides a rare real-world data point at scale, and it suggests the operational detection gap is enormous. A system that catches 2% of fraud while the score distribution signals 50% is not performing anywhere near 95% effectiveness by any definition that matters.

The human oversight ratio compounds the problem structurally. One supervisor per 150 students receiving AI-generated alerts in real time is not meaningful oversight — it is queue management at a ratio that makes response physically impossible. The AI produces signal; there is simply no staffing infrastructure to act on it. This is a system design failure that sits underneath the AI capability question, and the two are easy to conflate when a vendor's pitch bundles them together as a single product.

What UNAM's outcome actually demonstrates is that statistical anomaly detection — comparing a cohort's results against a verified multi-year historical baseline — was the only layer that worked at scale. That is not an argument against AI proctoring as a partial deterrent or a cost-reduction measure compared to fully manual approaches. It is an argument that any high-stakes remote assessment system treating the AI monitoring layer as the primary fraud detector, rather than one signal input into a broader statistical framework, is relying on a premise the evidence does not support.

Sources

arstechnica.com Cheat AI Explained — How Students Use AI to Cheat and How to Detect It (2026) Brown Professor Suspects Most of His Class Used AI to Cheat AI Proctoring Guide: Reduce Cheating by 95% in 2026