AI Glasses Score Perfect Mark on 160 Tax Accountant Exam Questions in 49 Minutes
AI glasses solved 160 past questions from the first round of the tax accountant exam in 49 minutes and received a perfect score of 400. The video documenting the experiment was released on the 13th. Tech YouTuber Techmong uploaded the experiment video to YouTube under the title "I got a perfect score on the tax accountant exam wearing AI glasses. It's honestly scary." The experiment used Lyrid's AI glasses together with Claude, an external AI model.
The experiment covered all four subjects of the tax accountant exam's first round: Public Finance, Introduction to Tax Law, Introduction to Accounting, and Commercial Law. In the first session, with Claude's reasoning level set to high, the glasses solved 80 questions on Public Finance and Introduction to Tax Law in 32 minutes. During grading, some answers were misaligned and initially marked wrong, but after corrections, the glasses scored 100 in both subjects. In the second session, with the reasoning level lowered to medium, the glasses solved 80 questions on Introduction to Accounting and Commercial Law in 17 minutes, scoring 100 in both. It earned a perfect score even in Introduction to Accounting, which contains many calculation problems. The total pure solving time for the two sessions was 49 minutes, averaging about 18.4 seconds per question.
In this year's first-round tax accountant exam, the average score for Introduction to Tax Law was 30.18 points, with a failing rate of 74.21%. Techmong explained that he scored 55 in that subject in an earlier experiment, but this time answered every question correctly.
Techmong explained that he had the AI pre-study five years of past questions from 2021 to 2025 and instructed it to solve the 2026 questions without access to answer keys. The 2026 tax accountant exam has already been administered, and the test questions and explanations have been made public. He noted that in an additional test using a different Claude model, the AI got one question wrong. The experiment did not involve sitting an actual national qualification exam, and electronic devices and charging are not permitted in real test centers. Because of these conditions, the results cannot be taken as evidence that such cheating would succeed in practice, and the verification has limitations.
Still, the fact that an AI, after studying past questions, can solve multiple-choice questions—including calculation problems—in an average of 18.4 seconds each, and that a person can access this through smart glasses, is seen as an example of the level AI technology has reached. In the video, Techmong called for discussion on how such devices should be monitored and what penalties should apply when cheating is detected.
