CSAT English Passages to Be AI-Generated, Pilot Run Targeted for June 2027

Plans to complete development by next May with a 3.15 billion won investment
Separate oversight project to conduct a three-stage review

The Korea Institute for Curriculum and Evaluation (KICE) is moving to develop an AI-powered test-setting support system that generates English questions for the College Scholastic Ability Test (CSAT) and predicts their expected correct-answer rates and the proportions of test-takers in each grade. After the proportion of test-takers receiving the top grade in last year's CSAT English section was just 3.11%, sparking controversy over a failure to control the difficulty, the institute plans to establish a data-based system for predicting and validating question difficulty.

On the morning of the 14th, students wait for the 2025 College Scholastic Ability Test to begin at Yeouido Girls' High School in Yeongdeungpo-gu, Seoul. Photo by the Joint Press Corps.

According to the request for proposals for KICE's "AI-Based Automatic Question Generation for CSAT English" project, released on Oct. 21, the institute plans to invest 3.15 billion won to build the system by the end of May 2027. AI company Upstage was selected as the preferred bidder for negotiations, and a formal contract is expected to be finalized in October.

KICE plans to have AI generate CSAT English passages and answer choices based on factors including curriculum achievement standards, target difficulty, question types, and vocabulary use based on curriculum lists. The system will analyze similarities to existing questions, including past exam questions and EBS textbooks, to prevent plagiarism disputes. It will also automatically generate appealing distractors that engage high-performing students' reasoning by applying principles for arriving at correct answers and types of distractor design.

In particular, the system is designed to use an AI model trained on historical data about test-takers' correct-answer rates to calculate the expected correct-answer rate and discriminatory power of each question. It will also simulate the expected average score for the entire test and the proportion of test-takers in each grade, with the aim of supporting adjustments to question difficulty during the test-setting process.

KICE is developing this AI test-setting support system because it believes the existing approach has limitations in ensuring continuity in question quality, systematic management, and standardization. In the request for proposals, the institute noted that it currently relies on an intensive, labor-heavy process in which test setters are isolated from the outside world for an extended period and work together in a residential setting to maintain CSAT security. However, it said the process of manually selecting passages, developing questions, and reviewing them within a limited time has its limitations.

In particular, regarding the scope of the project, KICE stated that "this project targets the English section of the CSAT," while adding the caveat that it "must be carried out with a view to expanding the system's scope of application to all sections of the CSAT." This leaves open the possibility that AI-assisted question development could later be expanded to subjects beyond English.

Separately from the development project, KICE is also pursuing a supervision project worth 246 million won. According to the request for proposals for the supervision project, the system's development process will be reviewed at three stages: requirements definition, design, and completion.

A KICE official said, "The goal is to pilot the system in the June 2027 mock exam," while adding, "We need to assess the situation carefully before making a decision." The official continued, "If we determine that the development results could cause confusion or are difficult to implement, the stability of the CSAT will take priority."

This content was produced with the assistance of AI translation services.