South Korea Tax Exam AI Cheating: Top Scorer Confession and National Reforms

On Sunday, November 15, 2026, the man who topped South Korea’s tax accountant qualification exam told millions of viewers he did not earn it honestly. Han Won-seok, 34, appeared on attorney Lee Don-ho’s YouTube channel and said he used a miniature camera, generative AI, and a smartphone to read answers while two proctors watched the room.

The confession closes months of suspicion after a 397.5 / 400 first-round score and a perfect mark in introductory tax law, a subject where more than 16,000 test takers averaged roughly 30 points. It also lands four days before the national CSAT college entrance exam, when every parent in the country is already nervous about fairness.

For builders, the story is not about tax law. It is about what happens when multimodal AI meets paper tests that still assume pencils, silence, and honor codes. explainx.ai has tracked the same tension in England’s coursework grading shift, OpenAI’s education guidance program, and why exam scores beat homework metrics in AI-and-learning research.

TL;DR

How the cheating method allegedly worked

According to Han’s on-camera account, summarized by Yonhap News and The Korea Herald, the loop was deliberately low-tech on the human side and high-tech on the answer side:

1. Capture — a miniature camera photographed exam book pages under two assigned proctors.
2. Solve — images went to a generative AI service (model unnamed publicly).
3. Read — Han checked output on a smartphone during timed multiple-choice and written sections.

He said he used the same pattern in the March 2026 CPA first-round exam, which he failed, and declined to detail gear brands so others would not copy him.

The HRD Service said Han sat center row, second row in round one and front row, center in round two, with two proctors in the room each time, according to Herald reporting. That layout matters: this was not a remote proctoring failure. It was in-person oversight losing to hidden hardware plus cloud inference.

A magnifier inspecting a ballot slot, evoking high-stakes exam integrity checks after AI cheating

Why the scores raised alarms

Tax accountant exams in Korea are famously selective. Korea Herald reporting cites HRD statistics: 722 passes out of 6,606 second-round takers in 2026, a 10.92% pass rate.

Han’s first-round 397.5 included 159 of 160 multiple-choice items correct. Three of four subjects were perfect; public finance was 97.5. The outlier was introductory tax law: perfect for Han while the cohort mean was 30.18 and 74.21% scored below the 40-point failure floor for that subject, per the same Herald piece.

Round two written work averaged 82.87 / 100 against a 49.75 passing cutoff, and Han ranked first on the written sections, according to Herald reporting. Fellow candidates reportedly questioned the gap long before Sunday’s video.

What authorities said on November 15

The Human Resources Development Service of Korea (HRD Service), which runs the tax accountant exam, issued a same-day statement carried by Yonhap. In English summary, it:

• Apologized to honest test takers and the public.
• Said it will file a criminal complaint against the confessing candidate on obstruction of official duties charges and coordinate with ministries on voiding his qualification.
• Promised a full review of all national qualification exams for irregular score patterns.
• Described active measures: more “AI inspectors” (dedicated supervisors for AI-device cheating), possible mass glasses checks, and evaluation of radio-frequency detectors to spot covert transmitters.

The Ministry of Economy and Finance is separately pursuing suspension through the Tax Accountant Qualification Deliberation Committee under the National Tax Service, per Herald and SBS English reports.

Han told viewers he submitted a written confession to Seoul police and would hand over remaining devices. Police assigned the case to the Seoul Metropolitan Agency, according to SBS.

CSAT timing and national nerves

The Straits Times and Korean outlets tie the scandal to Thursday, November 19, 2026, when roughly half a million students sit the CSAT (Suneung). Korea has not alleged CSAT cheating in this case, but the tax exam proof that AI plus a phone beats proctors arrives at the worst possible PR moment.

HRD’s pledge to audit every qualification track is effectively a promise that CSAT-style controls will be compared against what failed in April and July. Expect louder debate over:

• Faraday-style RF screening versus privacy complaints
• Smart glasses and micro-camera bans with physical inspection
• Whether AI answer apps should be blocked at the network level in exam halls (legally and technically controversial)

explainx.ai’s read: Korea is treating this as infrastructure failure, not one bad actor. That matches how LAUSD framed district-wide AI bans when integrity tools could not keep pace with student workflows.

What this means for AI policy and product teams

1. “Generative AI” is a cheating modality, not a single app

Han did not cite ChatGPT, Gemini, or a local Korean model. Investigators may recover app logs later. Policy teams should assume any multimodal cloud model that accepts photos and returns structured answers fits the threat model.

Research detectors like Google DeepMind’s DetectorScope explicitly warn against using classifier scores for individual academic integrity decisions. This scandal is the live counterexample: the cheat was image → answer, not essay → detector.

2. Proctoring that worked in 2015 loses in 2026

Two human proctors per room did not stop a center-seat candidate with a covert camera and phone. Remote proctoring vendors often market gaze tracking; this case needed transmitter detection and device-free zones that Korean officials are now considering.

Product managers building copilots for finance and tax should note the reputational blast radius. The confessing candidate was preparing to enter a licensed profession with fiduciary duties. Clients will ask whether any credential earned under traditional testing still signals skill.

3. Assessment redesign beats whack-a-mole

OfS guidance in England, covered in our November post, pushes secure, supervised tasks and skepticism toward AI detectors alone. Korea’s nationwide score audit is the emergency version of the same idea: if the metric can be gamed with a phone, the metric must change.

A regulatory stamp fitting a document slot, echoing national exam rule changes after AI cheating

What people are asking

Could he really photograph every question in time? Herald math notes 80 minutes for 80 questions in some sessions, about a minute per item including bubbling. Han’s method depends on selective photography of hard items (especially tax law) rather than every line. That matches why his anomaly clustered in the hardest subject.

Will AI inspectors actually help? HRD’s term refers to human supervisors trained to spot AI-enabled cheating, not autonomous proctor bots. They are scaling headcount, not deploying a model.

Is Korea alone? No. OECD and classroom studies explainx.ai covered in PISA-related fact checks show integrity stress everywhere. Korea’s difference is speed and scale of response because licensing exams gate elite jobs.

What about smart glasses? Han did not claim glasses-based capture, but Yonhap explicitly lists mass inspection of eyeglasses as an option under review. Meta Ray-Ban-style devices are not widespread in Korean exam halls today; the policy debate is preemptive.

Will other passers be re-tested? HRD’s full review language implies statistical sweeps, not automatic re-examination of all 722 passers. Expect spot audits and appeals if outliers cluster.

Should companies trust Korean credentials? Too early. One confirmed cheat does not indict thousands of honest passers. Watch for committee rulings, device forensics, and whether HRD publishes audit methodology after the CSAT week.

Practical checklist for educators and cert bodies

If you run high-stakes tests in 2026, steal the operational items Korea is now forced to implement:

1. Treat phones and watches as hostile — storage outside the room, not pockets.
2. Add RF anomaly detection where law allows; log false-positive rates publicly.
3. Run statistical integrity screens after each administration; perfect tails in hard subjects should auto-trigger review before certificates print.
4. Pair closed-book segments with performance tasks that require proctored live work, the direction OfS describes for UK higher ed.
5. Publish sanctions quickly so the market sees consequences, not rumors.

None of that blocks students from using AI to learn; it separates practice from verified competence, the same boundary OpenAI’s ThinkYoung EMEA program tries to teach teachers.

Timeline: from exam day to confession

The Tax and Finance Newspaper preparation story Han gave before the confession is now read as part of the fraud narrative in Korean coverage, not as independent verification of skill.

Related reading

• England OfS: AI coursework and secure assessment
• OpenAI closes the AI guidance gap in education
• Homework scores vs exam scores under AI use
• DetectorScope and why detectors are not integrity courts
• Los Angeles district AI ban and academic integrity
• OECD AI writing claims vs PISA science scores

Figures on pass rates, averages, and seating come from Korean news reports citing HRD Service statistics as of November 15, 2026. Han’s statements are attributed to his lawyer’s YouTube broadcast and English-language outlets quoting it.