On November 19, 2026, over 550,000 students sat for South Korea’s Suneung exam, armed with government-mandated stickers to cover their clothing buttons. The Prime Minister announced the measure just two days prior, aiming to block camera-equipped smart glasses and button cams used for AI-assisted cheating. However, the implementation was a classic case of patching a layer-7 problem with a layer-1 solution, resulting in a nationwide disruption that caught the unprepared rather than the sophisticated.

The Failure of the Emergency Alert

The government’s primary detection method was a mandatory, unopt-out emergency alert sent to every phone in the country between 8:25 and 8:30 AM. The logic was that hidden phones would ring, revealing the source of the AI connection. This failed immediately for two reasons: dedicated cheating hardware often lacks speakers or screens, and savvy users simply switch their devices off before the exam starts. The alert, lasting about 10 seconds, served as a false positive generator, disrupting the entire nation while missing the targeted demographic of pre-planned cheaters.

Infrastructure Gaps in Detection

The article argues that the real vulnerability lies in the transmission and reception layers, not the capture layer. While stickers covered cameras, the government admitted to not deploying radio detectors or live carrier checks. A device with its own cellular module or a pocket router does not ring during an alert but still transmits data. The author suggests that handheld radio detectors and real-time analysis of data sessions originating from exam buildings would have been far more effective infrastructure-level controls than visual inspections of clothing.

The Solved Benchmark Problem

Beyond the logistical failures, the core issue is that the Suneung has become a solved benchmark. A YouTuber’s AI glasses reportedly scored 96 out of 100 on a practice math test in just 18 minutes. The exam rewards speed and pattern recognition on known problem types, which is exactly what large language models excel at. By focusing on blocking the tool rather than changing the metric, the government is fighting a losing battle against hardware that continues to shrink and become indistinguishable from everyday accessories.

A Proposal for Open-Model Exams

The author proposes a sustainable fix: an open-model section of the exam. Instead of banning AI, every student would be issued the same sealed device running a fixed model snapshot, with no public internet access. The task would be a large, open-ended problem with planted errors, graded on the student’s reasoning logs and their ability to verify the AI’s output. This shifts the ranking from 'who can hide a camera' to 'who can effectively direct and audit an AI,' a skill relevant to modern engineering workflows.

Key Takeaways

  • The nationwide emergency alert failed because dedicated cheating hardware often lacks audio components or is simply switched off.
  • Stickers on buttons are a superficial patch; radio detectors and live carrier checks are the actual technical controls needed.
  • The Suneung is now a 'solved benchmark' for AI, with models scoring 96/100 in under 20 minutes on practice tests.
  • A proposed 'open-model' exam section would standardize AI access, grading students on reasoning and verification rather than recall.
  • The government’s plan to retroactively pull AI service logs lacks a defined legal basis and clear scope, raising privacy concerns.

The Bottom Line

Stop patching the user interface of cheating and start upgrading the exam’s architecture. If the test measures what LLMs do best, the infrastructure will always be one step behind the hardware.