Misguided Recommendations Reveal Blind Spot in "AI Algorithms"... UNIST Develops "EchoTrace" to Diagnose Recommendation System Risks

Lee Yeonchang's team diagnoses a blind spot in AI algorithms that entrenches flawed recommendations

From recommending nonexistent products to amplifying bias... Accepted for ACM CIKM 2026

A study has found that product recommendation services powered by large language models (LLMs) relearn from their past errors and repeatedly produce distorted recommendations.

Researchers in South Korea have developed a diagnostic tool to track persistent flaws in LLM recommendation systems, such as recommending nonexistent stores or making up unrelated occupations.

Consumers increasingly trust and choose products recommended by artificial intelligence (AI). Yet AI has been relearning its own errors and continuing to recommend inappropriate products.

A research team led by Professor Lee Yeonchang of the UNIST Graduate School of Artificial Intelligence said on November 8 that it had developed EchoTrace, a diagnostic tool for tracking bias and errors in recommendation AI based on large language models (LLMs).

Researchers Professor Lee Yeonchang (left) and researcher Park Donguk. Provided by UNIST.

The research team’s analysis revealed a critical flaw when AI generated recommendation lists directly. The rate of recommendations for unrelated items not on the lists at all reached 13.14% for stores, 7.40% for books, and 4.07% for movies.

The “hallucination” of fabricating user profiles was also serious. Of the occupations AI inferred solely from users’ movie-watching histories, 93.16% did not match their actual occupations. It also arbitrarily generated occupational categories that were not included in the reference data.

The biggest problem is that these errors do not disappear. Even after the recommendation results were used for retraining five times, the initial bias persisted or recommendations gravitated toward certain popular items. For roughly nine out of 10 users, the AI relied primarily on the work it had just recommended when making its next recommendation. In effect, the AI became trapped in a cycle of errors of its own making.

The research was accepted for ACM CIKM 2026, an international academic conference in the field of data mining.

“This study lays the groundwork for systematically diagnosing how bias and hallucinations in LLM-based recommendation systems accumulate and intensify through repeated recommendation cycles, based on the role played by the LLM,” Professor Lee Yeonchang said. “The findings could be used to design and evaluate safer and more reliable recommendation systems in the future.”

Park Donguk, a researcher at the UNIST Graduate School of Artificial Intelligence, participated in the study as first author. The team conducted the research jointly with Professor Lee Dongwon of Pennsylvania State University in the United States.

The research was accepted for ACM CIKM 2026 (Conference on Information and Knowledge Management), a leading international conference in data mining. The conference began on November 7 and runs through November 11 in Rome, Italy.

Research Figure: EchoTrace's LLM-Based Recommendation System Risk Diagnosis Process.

The research was supported by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation (IITP) through the AI Star Fellowship, Graduate School of Artificial Intelligence, Generative AI Leading Talent Development, Regional Intelligence Innovation Talent Development, and InnoCORE programs, as well as by the U.S. National Science Foundation (NSF).

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