On November 15, 2026, the top scorer of Korea’s 63rd Certified Tax Accountant (CTA) exam turned himself in to police, confessing to a high-tech cheating scheme. The candidate, referred to as "A," achieved a near-perfect 159 out of 160 on the multiple-choice first stage and topped the essay-based second stage with 663 out of 800 points. His method involved a hidden micro-camera photographing the exam papers, which were then processed by a generative AI outside the testing room. The AI-generated answers were relayed to A’s smartphone, allowing him to transcribe responses that were technically not his own.

The Infrastructure of Suspicion

The scandal highlights a critical failure in current proctoring infrastructure. A sat in the front row center, directly under the eyes of four committee members, yet no suspicious behavior was noted. The Human Resources Development Service of Korea (HRD Korea), which administers the exam, found no anomalies in the electronic scoring data. The breach was identified not by the exam proctors, but by online communities that flagged the statistical improbability of his scores. This suggests that current monitoring tools, such as radio-wave detectors and CCTV, are reactive rather than proactive, focusing on specific hardware like smart glasses while missing simpler, more effective setups like a micro-camera and a phone.

AI Has Been Passing Professional Exams for Years

This incident isn't an anomaly; it's the inevitable result of AI capabilities meeting human incentives. Research from 2023 already showed ChatGPT scoring between 52.4% and 75.0% on the United States Medical Licensing Examination (USMLE), near the passing threshold. Similarly, GPT-4 averaged 85.1% on multiple-choice questions from four accounting certification exams, including the CPA. While these were sample-based studies, not actual exam sittings, they demonstrated that generative models could handle the cognitive load of professional licensing. The shift from multiple-choice to essay formats did not solve the problem; in fact, generative AI is arguably better at producing coherent long-form text than selecting single correct options, making the second stage equally vulnerable to outsourcing.

The Cost of Unverifiable Competence

The fallout extends beyond one candidate. Following the confession, other high scorers in the LEET (law school admission test) and MEET (medical graduate school test) faced baseless suspicion, illustrating how AI cheating degrades the value of every high score. When an exam cannot distinguish between human intelligence and AI output, the credential itself loses its signaling power. The author, Kwangseob Ahn, argues that the solution isn't just better surveillance, but a structural change to assessment: implementing oral defenses. Like the oral interview required for Korea’s Professional Engineer (gisulsa) exam, follow-up questioning forces candidates to explain their reasoning in real-time, a task AI cannot seamlessly fake without the candidate understanding the underlying logic.

Key Takeaways

  • Reactive Surveillance Fails: Current anti-cheating measures like radio-wave detectors are outdated and miss simple camera-phone setups.
  • AI Competence is Proven: Studies from 2023 confirm that LLMs can already pass or nearly pass major professional licensing exams.
  • Essay Formats Are Vulnerable: Generative AI excels at long-form writing, making essay-based exams susceptible to the same outsourcing.
  • Oral Defense is the Fix: Real-time questioning is the only reliable way to verify that a candidate possesses the knowledge they claimed on paper.

The Bottom Line

We can no longer trust written output alone as proof of competence; if we want credentials to mean something, we must return to the oral defense.