EXPLORINEER
en

KR — 전자신문

Using AI, Tax Accountant Exam Scores Perfect; Hwang Jeong-ah Calls for Full Investigation

AI2 min read

A Korean lawmaker tested AI models on this year's tax accountant exam and law school aptitude test (LEET). Top AI models scored perfectly, which raised questions about high scores on the real exams.

What happened

Rep. Hwang Jung-a of the Democratic Party fed the 2026 tax accountant first-round exam papers into Anthropic's AI model Claude. She gave only the exam images, with no answer keys or explanations. The AI's answers were then graded against the official answers from the Korea Human Resources Development Service. A new model, Opus 4.7, released 9 days before the exam, scored a perfect 100 on the Accounting exam. In the real exam, the number of people scoring 90 or higher jumped more than 30 times over last year.

Key numbers

  • Opus 4.7 scored 100 on Accounting (40 questions, all correct). Earlier models scored 62.5 (Sonnet 4.5), 25 (Haiku 4.5), and 75 (Opus 4.5).
  • People scoring 90+ on Accounting: 1 in 2024, 2 in 2025, and 66 this year.
  • Highest Accounting score rose from 90 and 95 in earlier years to 100 this year.
  • Only 1 person scored 90+ on Tax Law this year, a perfect score.
  • On LEET (law school aptitude test), Fable 5 got all 30 Language questions right and 39 of 40 Reasoning questions right. After extra study on past exams, it got a perfect score on Reasoning too.

Some context

The tests were given to the AI models in different ways. For Tax Law, the AI was allowed to read the original law texts for each question before answering. No answer keys or explanations were given. The article says the AI's results improved with each newer model, and that the exam scores rose at the same time.

Why it matters

The lawmaker says the jump in top scores is unlikely to be a coincidence and wants a full check for cheating. The case also shows how fast AI is improving. In seven months, AI went from weak results to perfect scores on tax and law exams. She says Korea needs fairness measures for all national qualification exams, like the supervision used for the college entrance exam.

What to watch

  • Whether authorities carry out the full check of high scorers that the lawmaker requested.
  • Whether the agency changes how national qualification exams are supervised.
  • Whether the ability of AI models on school and professional exams keeps rising with each new release.

Glossary

Claude
An AI model made by the US company Anthropic. It answers questions and writes text.
Opus 4.7
A newer AI model from the same family. It was released 9 days before the tax exam.
LEET
The Legal Education Aptitude Test, a Korean exam used to apply to law schools.
Fable 5
An AI model named in the article. It was tested on the LEET exam.
Korea Human Resources Development Service
A Korean public agency that runs national exams and publishes the official answers.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

AITechCrunch

Is selling an AI version of your face to this microdrama startup the future of Hollywood?

Microdrama startup Shortical, founded last January, makes short two-minute romance episodes and is now producing AI-generated content. Its latest series, "Inevitable," stars Israeli actor Aki Avni, who licensed his face to the company to create a digital avatar of himself for the wholly AI-made project, which premiered in Los Angeles. Shortical says the offering is in its early days and has not yet resulted in any signed contracts beyond the Avni deal.

AITechCrunch

Here are the top AI agents that can live in your text messages

A growing number of AI agents can be texted like a normal contact, remembering context, connecting to everyday apps and services, and completing tasks such as scheduling, booking, shopping, and reminders. The article highlights Instinct, which reportedly has a $10 billion valuation after raising $1 billion, along with other options like Ask Proxy, Caddy, Comma, and Fambot, which target personal, work, and family use.

AIThe Register

ChatGPT shrunk down to run on a Super Nintendo

Tech journalist Benj Edwards has released ChatGPT for SNES, a tiny chatbot that runs in an emulator or as a ROM on a Super Nintendo cartridge. It is a 2.2 million-parameter distilled version of OpenAI's GPT-OSS-120b, with an optional fact book drawn from simple English Wikipedia, and Edwards says it knows about 8,000 words and often gets things wrong.