'APEX' co-developer Ahn Hee-gap: 'This exam scores good judgment, not right answers'

The rise of generative AI has transformed university classrooms — and nowhere more visibly than in computer science and AI graduate programs. Ahn Hee-gap, chair of the AI Graduate School at Pohang University of Science and Technology, said that where courses once revolved around lectures, written assignments, paper exams and seminars, AI now weaves through virtually every stage of teaching and research.

Ahn is recognized as an authority in algorithms — last year he received a ministerial commendation from the Ministry of Science and ICT for his research on optimal data structures and applied algorithms in dynamic geometric environments. More recently, he joined the team that developed APEX, a domestic AI competency certification exam designed to comprehensively measure and certify AI job skills. He now finds himself fielding a pointed question: can a test really measure how well someone uses AI?

"Saying that using AI well is just a matter of intuition is not entirely wrong, but it is not the whole answer either," Ahn said in an interview Friday. "What looks like intuition can be converted into observable performance and evaluated."

APEX, which Ahn helped develop, is an AI job-competency certification exam co-developed by Day1 — in partnership with OpenAI technology — together with researchers from Seoul National University, POSTECH and Korea University. Rather than asking whether candidates know about AI, it asks whether they can complete real work tasks with AI. The 50-question, 140-minute exam assesses six competency areas: problem definition, precise instruction, output refinement, integrative thinking, security and ethics, and business judgment.

"Evaluating AI utilization competency requires looking not only at foundational knowledge but also at multidimensional abilities — problem-definition skills, verification ability, robustness and reproducibility, an understanding of safety, ethics and security, and the capacity for collaboration and explainability," Ahn said. The exam, he added, is not a simple extension of conventional paper-based testing; its content and format had to be designed from scratch.

Ahn said the ideal scoring approach uses a rubric that evaluates candidates on: structuring problems and defining requirements and constraints; presenting evidence-based solutions step by step; verifying results and correcting errors across varied data and conditions; carrying out safety, security and ethical checks; and seeing the work through to responsible completion, including performance improvement.

Because APEX ventures into unfamiliar territory with a format that has no real precedent, some test-takers have expressed skepticism about trusting scores generated by AI. Ahn said it is essential to explain clearly what standards and methods underpin the scoring. "APEX does not simply present a score — it provides the evaluative basis and feedback for each criterion so that candidates can understand their results and improve," he said. He added that the system continuously and systematically checks whether AI scoring aligns with the rubric to maintain consistency and fairness, and that safeguards including feedback mechanisms and improvement procedures are built in to make the scoring system verifiable.

Ahn has set exam questions for the International Olympiad in Informatics and for major corporate recruitment tests — work that always involved problems with definitive right answers. Designing a test that scores judgment rather than correctness was a different challenge entirely. "The biggest difference in an AI utilization competency exam — where you are scoring open-ended judgment — was that the task was not to write questions with correct answers, but to design criteria that translate good judgment into scorable performance," he said. "It was not simply a matter of right or wrong. We had to evaluate how a candidate defined the problem, what assumptions they made, what reasoning drove their decisions, and what verification, error-correction and safety procedures they followed."

APEX is an exam taken with AI — candidates solve problems using AI tools at their side. Ahn said this reflects where the workplace is heading. "Real work going forward will not be an environment where AI is banned, but one where AI is used to raise productivity and quality," he said. "APEX does not evaluate people who can beat AI — it evaluates people who use AI to extend their own problem-solving capacity."

The goal, Ahn said, is not merely knowing how to use AI but connecting that knowledge to job outcomes. He defined that "connection" as "the ability to go beyond producing plausible outputs and to link AI use — in a real job context — to problem definition, execution, verification, risk control and reproducible results." It is not about crafting clever prompts, he explained, but about taking full operational responsibility for the entire workflow, AI included.

Throughout the interview, Ahn returned repeatedly to two themes: ethics and continuous improvement. Both are central to what he called "the ability to operate responsibly." They are also why APEX places equal weight on what he described as "thinking muscles" — the capacity to analyze, structure and decide — and "adaptive muscles": the ability to verify outputs against AI error and uncertainty, control risk from safety, security and ethical perspectives, and drive improvement. Those two capacities, he said, are the core of what APEX is designed to test.