A study has found that artificial intelligence (AI) recommendation systems utilizing large language models (LLMs) can lead to bias and repeated errors when they retrain on their own recommendations.
According to the Ulsan National Institute of Science and Technology (UNIST) on November 8, a research team led by Professor Lee Yeon-chang developed a diagnostic tool called 'EchoTrace' to track the processes of bias and error accumulation in LLM-based recommendation systems. The team analyzed three methods in which LLMs generate recommendation lists or create user profiles and training data.
When LLMs directly generated recommendation lists, the rate of suggesting items not included in the recommendations was 4.07% for movies, 7.4% for books, and 13.14% for stores. The estimated user occupations based on movie viewing records differed from actual records 93.16% of the time. The generated training data also showed a bias toward specific genres or industries. In the final recommendations, popular items appeared more frequently compared to actual usage records.
Even after repeating the retraining process five times, initial biases and errors persisted. In some systems, the bias toward popular items intensified. Notably, in models where LLMs directly recommended movies, 86 to 92 out of 100 users relied heavily on previously recommended films when selecting their next movie. This 'feedback loop,' where AI-generated recommendations are learned from user selection records and reflected in subsequent recommendations, perpetuated the initial biases and errors.
Professor Lee stated, "This research provides a systematic foundation for diagnosing how biases and hallucinations in LLM-based recommendation systems accumulate and amplify through repetitive recommendation processes, depending on the role of the LLM."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.