Three AI models released before exam achieve perfect score in accounting

AI also aces tax law section when given access to legal statutes

Lawmaker calls for full probe of high scorers' statistical anomalies

AI models released before this year's first-stage tax accountant licensing exam scored perfect marks on the introductory accounting section after being shown only the question paper, according to a lawmaker's office.

The number of test-takers who scored 90 or above on that subject surged from two last year to 66 this year, an unusual distribution that has prompted calls for a comprehensive investigation into potential AI-assisted cheating.

Education Minister Choi Kyo-jin reviews materials on AI cheating cases involving miniature cameras at the tax accountant exam during a plenary session of the National Assembly's Education Committee on Monday. [Yonhap]
Education Minister Choi Kyo-jin reviews materials on AI cheating cases involving miniature cameras at the tax accountant exam during a plenary session of the National Assembly's Education Committee on Monday. [Yonhap]

According to the office of Democratic Party of Korea lawmaker Hwang Jeong-a, Anthropic's AI model "Opus 4.7," released April 16, answered all 40 questions on the introductory accounting section of the 63rd tax accountant first-stage exam correctly, scoring 100 points, when given only the question paper. The exam was held April 25 — just nine days after the model's release.

Hwang's office said it presented image files of the question paper, without answer keys or explanations, to models in Anthropic's Claude family and had each model attempt the paper once, then graded the results against the official answer key published by the Human Resources Development Service of Korea.

AI models released last year scored lower on the introductory accounting section: Sonnet 4.5 received 62.5 points, Haiku 4.5 received 25 points, and Opus 4.5 received 75 points. Opus 4.6 and Sonnet 4.6, both released in February this year, each scored 100 points, and Opus 4.7 — released just before the exam — also achieved a perfect score.

High scorers on the introductory accounting section surged in the actual exam as well. According to the Human Resources Development Service of Korea, the number of test-takers scoring 90 or above rose from one in 2024 and two in 2025 to 66 this year. The top score also climbed over the same period — from 90 to 95 to 100 points. This year marked the first time in three years that anyone achieved a perfect score.

The introductory tax law section told a different story: the number of test-takers scoring 90 or above was one in 2024, zero last year and one this year. Opus 4.7, when given only the question paper, scored 72.5 points on that section. The surge in high scorers was concentrated precisely in the subject where AI could achieve a perfect score from the question paper alone.

Results of AI models' attempts at the tax accountant first-stage exam [Office of Rep. Hwang Jeong-a]
Results of AI models' attempts at the tax accountant first-stage exam [Office of Rep. Hwang Jeong-a]

However, it also emerged that a perfect score on the introductory tax law section was achievable using AI models available at the time of the exam.

When Hwang's office instructed Opus 4.7 to look up the relevant statutory text from sources such as the National Law Information Center before answering each question, the model answered all 40 questions correctly. No answer keys, explanations, or post-exam materials were provided during the process.

A similar result appeared in the Legal Education Eligibility Test (LEET) held July 19 for the 2027 academic year. "Fable 5," released June 9 — 40 days before the exam — answered all 30 questions in the verbal reasoning section correctly when given only the question paper, and got 39 of 40 questions right in the logical reasoning section. When trained on the past three years' exam questions and allowed to work through items in batches, it answered every logical reasoning question correctly as well.

AI problem-solving ability continued to improve rapidly even after the exam. In tests conducted by Hwang's office, all nine AI models released after the tax accountant first-stage exam scored 90 or above on the introductory accounting section, and six achieved perfect scores. On the introductory tax law section, five models scored 90 or above from the question paper alone.

"Fable 5.2," a top-tier model released Nov. 9, scored perfect marks on both the introductory accounting and introductory tax law sections without any additional training — just the question paper. It also answered all 70 LEET verbal and logical reasoning questions correctly. The latest AI model took about one minute and 30 seconds to work through all 40 introductory accounting questions.

"It is hard to regard as coincidence the fact that high scorers increased more than 30-fold as AI models' problem-solving capabilities advanced," Hwang said. "We must conduct a comprehensive investigation into the statistical anomalies among high scorers and clearly determine whether cheating occurred."

Hwang added that AI performance had reached the level of achieving perfect scores on tax law and the LEET within seven months of the exam and would continue to improve. Hwang called for reviewing the oversight and supervision measures applied to this year's College Scholastic Ability Test and expanding them to all national licensing exams. "As President Lee Jae Myung has stated, we must mobilize the full capacity of the state to block cheating and make an example of those responsible," Hwang said.


y2k@heraldcorp.com