100 Points on an Exam Where 16,856 People Averaged 30

A 100 on an Exam Where 16,856 People Averaged 30

Korea has a national professional license called the Certified Tax Accountant (CTA). It’s the credential for people who file taxes, handle tax adjustments, and represent clients in tax disputes on their behalf. The exam has two stages. This year’s 63rd first-stage exam drew 16,856 applicants, of whom 18.09% passed. The second stage is run with a government-set minimum quota of 700 passers, and last year’s second-stage pass rate was 10.48%. It’s common for people to spend several years preparing for this exam.

On this year’s first-stage exam, in a subject called “Introduction to Tax Law,” the 16,856 test-takers averaged 30.18 points. One person scored a perfect 100. He got 159 out of 160 questions right on the first-stage exam, and on the essay-based second stage, he topped the field with 663 out of a possible 800 points. Let’s call this person “A.”

On November 15, 2026, A confessed to cheating and turned himself in to the police. According to his confession, a micro camera hidden on his body photographed the exam papers, a generative AI produced the answers outside the testing room, and A checked those answers on his smartphone.

This exam may sound unfamiliar to readers outside Korea, but any society that confers credentials through testing can’t escape the same question. How can an exam tell the difference between an answer solved by AI and one solved by a human?

Front Row, Center Seat, Right in Front of the Proctor

Examinee A’s seat for the second-stage exam was front row center, directly in front of the proctor. None of the four committee members supervising that session reported seeing anything suspicious. The Human Resources Development Service of Korea¹, which administers the exam, found nothing unusual in the first-stage scoring data, which is graded electronically. The suspicion didn’t start within the exam process at all — it started in an online community that found the score itself strange.

The same method didn’t always work, either. A had also admitted to using electronic devices and generative AI on the 61st Certified Public Accountant first-stage exam — and failed that one. A failed attempt at cheating just looks like an ordinary failure; a successful one leaves nothing behind but a high score. Either way, the answer sheet shows no trace.

In his confession video, A said his motive was fear — fear of not knowing how much longer his exam-prep years, which had already begun in his 30s, would drag on. The bigger the exam riding on passage, the more test-takers will feel that same temptation.

The government’s response has been to roll out radio-wave detectors, expand CCTV in test centers, and station staff trained to spot smart glasses at Suneung, Korea’s national college entrance exam, test sites. And yet the warning the Service built into its training materials for tax-accountant second-stage exam proctors was about smart glasses — and the suspicion first raised online was also about AI smart glasses. What was actually used was a micro-camera and a smartphone. Supervision targets whatever device has already been exposed; cheating simply moves on to the next one. Once the radio-wave detectors arrive, some other method will take their place.

The Next Day, Other High Scorers Started Getting Suspected

After the confession came out, a post started circulating in Korean online communities that compiled this year’s unusual results into seven lines. It listed the LEET² (the law school admission test), the MEET (the medical and dental graduate school admission test), the certified tax accountant exam, the customs broker exam, the certified appraiser exam, and the Grade 5 civil service exam, all in one lineup. Of these, I could only confirm the perfect LEET score through news reports; I couldn’t verify whether the rest of the claims were true.

There’s no evidence that the people on that list cheated using AI. The person who got a perfect LEET score gave a detailed media interview about how they studied, and there’s been no reporting tying them to any wrongdoing. And yet there they were, on the same list.

This kind of suspicion predates this particular scandal. When the LEET produced its first-ever perfect scorer this September, 19 years after the exam was introduced, the first reaction in test-taker communities was reportedly: “Isn’t this rigged?” The tax accountant scandal just gave that vague suspicion one concrete case to point to.

This suspicion is itself a cost that cheating leaves behind. If an exam can’t tell apart a high score made by AI from one made by a person, then even someone who earned a high score honestly has no way to prove their innocence. The higher the score, the deeper the suspicion. Exam scores exist to vouch to society that “this person knows this much” — and that guarantee is now shaking from the very top.

AI Was Already Near the Passing Line on These Exams Back in 2023

The fact that AI can tackle professional licensing exams isn’t a sudden development. A study published in February 2023 in the journal PLOS Digital Health reported that ChatGPT scored between 52.4-75.0% on publicly available questions from the United States Medical Licensing Examination (USMLE). That exam’s passing threshold sits around 60%. A paper released the same year by US and German researchers found that GPT-4, given sample questions and tools like a calculator, averaged 85.1% on multiple-choice questions across four accounting certification exams, including the Certified Public Accountant (CPA) exam. OpenAI’s claim that GPT-4 scored in the top 10% on the bar exam was later re-examined in a 2024 paper that called it an overstatement.

These weren’t perfect scores, and the numbers themselves were contested. None of this happened in an actual exam room, either. Still, it doesn’t change the fact that a model from three years ago was already hovering near the passing line on exams that take humans years of preparation to clear.

Multiple choice or essay — same vulnerability, same reason

There’s a reason multiple-choice exams are so common. Grading is fast, machines can read the answers, and there’s only one correct answer per question, so disputes are rare. But that same advantage works just as well for someone cheating. All the AI outside the exam room has to pass to the test-taker inside is a single number per question. Copying it down takes almost no time, and the answer sheet shows no trace of how the answer was reached. Looking at the sheet alone, you can’t tell the 159 questions a human got right from the 159 questions an AI got right.

So would switching to essay-format questions fix it? This case answered that too. The second-round exam in which Mr. Ahn placed first was essay-format. In an interview published on November 14, the day before he turned himself in, Mr. Ahn cited this second-round score to push back against the allegations — holding up a high essay score as proof of his innocence. One day later, he admitted he had cheated on the second round as well.

The reason is simple. Generative AI is actually better suited to writing long passages than to picking the right number. If you can photograph a question and send it out, and get an answer back, an essay response just takes a little longer to copy down. The paper answer sheet only shows what was written — not who thought of it. The agency was having Mr. Ahn’s second-round answer sheet re-reviewed by grading experts, but his own confession came before that review ever reached a conclusion.

Oswarld’s Lens

As someone who teaches at a university, I’ve consistently argued that we should adopt essay exams and interview-style oral exams, even if they’re more burdensome for the person grading them. That’s exactly how I’ve evaluated students in my own courses. I often give students a written exam, then follow up by asking them to explain, in person, whether they actually understood what they wrote. In essays and oral exams, you can immediately tell the difference between a student who can articulate what’s in their head on the spot and one who can’t.

In a piece I wrote in January 2024, I noted that as AI writing tools spread, it was becoming harder to tell whether a student understood the material just from the polish of their writing — and that we’d need an oral defense³, where students explain in their own words what they wrote. In my book Those Who Outsource Their Thinking: Homo Brainless, I covered how AI was already passing medical, legal, and accounting exams.

This incident narrowed that judgment a bit further. A written essay alone isn’t enough. For an essay to do its job as an assessment tool, whoever wrote the answer has to be able to answer follow-up questions about it.

What makes a follow-up exam different from a written one is that the questions aren’t fixed in advance. The grader listens to the answer and asks, on the spot, why the candidate reached that judgment, and how the conclusion would change if the conditions were different. Because each next question depends on the previous answer, the candidate can’t know in advance what will be asked. An answer someone else produced for you might survive the first question, but it tends to break down under the ones that follow.

The downsides are real. Grading takes much longer, and grade appeals spike during the correction period. I think of this as a cost that educators simply have to bear. If we insist on formats that are convenient to grade, that same convenience ends up benefiting candidates who use AI, too.

National exams operate at a different scale than university courses. In an exam where 16,856 people sit for a single subject, you can’t put every candidate through an oral evaluation. Even so, there’s already a model where oral exams come at the final stage. Korea’s gisulsa (professional engineer) exam⁴ requires candidates who pass the written essay portion to sit for a roughly 30-minute oral interview. It’s worth considering a similar procedure for professional licensing exams — having candidates who clear the written essay stage of the second-round exam explain their own answers. Oral exams won’t catch every instance of cheating. But they let the testing process double-check, at least once, whether what’s on the answer sheet actually came from the candidate’s own head.

This step is also what’s needed for high scorers currently under suspicion. Right now, about all someone with a perfect score can do in the face of community suspicion is explain their study methods in an interview. If the exam itself included a follow-up stage, a high scorer could demonstrate, through the official process itself, that the score is genuinely theirs.

This isn’t something you fix by changing the exam alone. When the abacus arrived, people learned how to use the abacus; when the calculator arrived, people learned how to use the calculator. The same happened with computers and Excel. Now test-prep academies are busy figuring out how to use AI well. I find that striking.

The fact that AI can pass professional licensing exams can also be read the other way around. It means that even someone without professional training can now, with AI’s help, handle work that requires at least the level of expertise the exam is testing for. If that’s true, we need to reconsider where the boundaries of professional expertise actually lie. The knowledge that exams have traditionally verified is different from the experience and know-how that accumulate through actually doing the work. The latter is often called tacit knowledge. I suspect that, going forward, who’s memorized more statutes will matter less and less.

Teaching people to use AI well and asking students to demonstrate their own understanding aren’t in conflict. To judge whether an answer AI produced is correct, you have to understand the content yourself — and that understanding is exactly what a follow-up exam is designed to check.

Closing

Surveillance equipment has to change every time the methods of cheating change. But the follow-up questioning works the same way no matter what equipment is involved — it checks the same thing regardless. And that check isn’t only useful for catching cheaters. It can also clear someone who earned a high score honestly from suspicion.

---

💬 Has anyone ever asked you, after a test, “Why did you answer it that way?” If so, tell us in the comments which question was hardest to answer.

📨 If you know someone preparing for a certification exam, or a colleague who designs tests and assessments, please share this piece with them.

References & Further Reading

News coverage

• Financial News, report on the top-scoring tax accountant candidate turning himself in, November 15, 2026. Link ··· Covers the confession details (micro camera, generative AI, smartphone), his failure on the first round of the CPA exam, the seating arrangement and proctor’s responses, and the government’s countermeasures.
• The Kyunghyang Shinmun, reporting on the first- and second-round tax accountant exam scores, November 12–13, 2026. Link 1 · Link 2 ··· Source for the 159 questions, the 30.18-point average on the Introduction to Tax Law section, and the 663-point second-round score.
• Herald Corp, report on proctor training materials and the re-examination of second-round answers. Link ··· Shows how smart-glasses warnings ended up in the training materials.
• Ajunews, report on allegations raised in online communities, November 15, 2026. Link
• Tax Times, “Tax Accountant First-Round Pass Rate: 18.09%.” Link ··· Source for this year’s first-round applicant count and pass rate.
• Tax Times, “62nd Tax Accountant Exam: 728 Final Passers, 10.48% Pass Rate.” Link
• JoongAng Ilbo (published via Korea Daily), interview with the first perfect scorer on the LEET, September 2026. Link ··· Covers the first perfect score in the exam’s 19-year history and the community’s initial reaction.

Research on AI and professional licensing exams

• Kung et al., PLOS Digital Health, 2023. PLOS press release ··· ChatGPT’s performance on publicly released USMLE questions.
• Eulerich et al., Review of Accounting Studies, 2024. Link ··· Results on multiple-choice questions from four accounting certification exams. This was a sample-question experiment, not an actual exam sitting.
• Martínez, Artificial Intelligence and Law, 2024. Link ··· A paper that recalculates the claim that GPT-4 scored in the top 10% on the bar exam.

The author’s previous writing

• Kwangseob Ahn, an essay arguing for supplementing writing assessment with oral defense, January 31, 2024. Link

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. HRD Korea (Human Resources Development Service of Korea): A public agency that administers, on the government’s behalf, the setting, execution, and grading of national qualification exams such as those for tax accountants and professional engineers. ↩

2. LEET (Legal Education Eligibility Test): The law school aptitude test used for admission to Korean law schools. It plays a role similar to the LSAT in the US, and is known as an exam whose scores are notoriously hard to improve through studying. ↩

3. Oral defense: A format in which a person explains their own writing or answers aloud before an evaluator and responds to questions. It’s widely used in thesis and dissertation review. ↩

4. Professional Engineer (gisulsa): The highest grade of national technical qualification in Korea. Those who pass the written essay-style exam must then sit a further oral interview. ↩