AI Models Ace South Korea’s College Entrance Exam in Under Three Minutes

Global Economic Times Reporter

korocamia@naver.com | 2026-11-19 18:04:48


Latest iterations of ChatGPT and Gemini achieve perfect scores across Korean, math, and reading sections of the 2027 CSAT, sparking renewed focus on anti-cheating protocols.

SEOUL, Nov. 19 — Leading artificial intelligence models from OpenAI and Google have achieved perfect scores on the core subjects of South Korea’s notoriously rigorous 2027 College Scholastic Ability Test (CSAT), completing the exams in less than three minutes combined.

According to testing conducted on Thursday following the release of the official exam papers by the Korea Institute for Curriculum and Evaluation (KICE), OpenAI’s GPT-6.2 Astra and Google’s Gemini 4 Hydrogen answered every question correctly in the Korean, mathematics, and English reading sections.

In the evaluation, researchers uploaded the exam PDFs directly to the models without web search access and prompted them to list the correct answers. Both AI systems scored a full 100 points each in Korean and math, and 63 out of 63 points on the 28-question English reading comprehension section (excluding the audio-based listening portion).

Notably, the AI models solved a heavier workload than human test-takers. While students choose one elective track per subject—answering 45 Korean and 30 math questions—the models were tasked with solving all elective tracks simultaneously, tackling 56 Korean and 46 math problems.

Operating on a high reasoning setting and utilizing code execution tools for certain calculations, GPT-6.2 Astra generated its answers in just 2 minutes and 44 seconds total: 44 seconds for Korean, 1 minute and 25 seconds for math, and 35 seconds for English. Gemini 4 Hydrogen also achieved a 100% accuracy rate, though it occasionally requested clarification on specific math problem conditions due to difficulties parsing certain complex graphs and equations.

Performance varied slightly depending on the operational environment and model version. In additional runs using different chat modes and reasoning tiers, GPT models occasionally missed a single Korean language question, scoring 98 points. Independent tracking on a public LLM benchmark dashboard operated by developer "hehee9" similarly confirmed that flagship models—including GPT-6.2 Astra, Sol, and Gemini 4 Hydrogen—swept all common and elective questions for this year's Korean and math exams.

Rising Concerns Over Exam Security

The rapid advancement in AI problem-solving capabilities has intensified scrutiny over test security and academic integrity in South Korea. Because examinees could theoretically capture images of test papers and receive answers within seconds, regulating cameras and communication devices has become a top priority for exam administrators.

Earlier this year, the top-scoring candidate on the national Certified Public Tax Accountant exam admitted to cheating using a miniature camera paired with generative AI. Police have also launched an investigation into allegations that a perfect scorer on the Legal Education Eligibility Test (LEET) utilized AI during the exam.

In response to these incidents, the South Korean government has tightened oversight ahead of national exams. During an interagency meeting on Nov. 16, Prime Minister Han Seong-sook ordered strict inspections to ensure smart devices are confiscated before testing begins and called for a comprehensive review of state-administered exam protocols. Authorities announced they will continue investigating suspected irregularities even after exams conclude, warning that confirmed violators face testing bans and criminal referrals.

[ⓒ Global Economic Times. 무단전재-재배포 금지]

WEEKLY HOT