LLMs in Recommender Systems: 13 Opportunities, 18 New Risks

A team of 16 researchers led by Bohao Wang from multiple Chinese universities has published a comprehensive survey (arXiv:2606.00540) on the impact of large language models (LLMs) on recommender system trustworthiness. By analyzing over 200 recent studies, they find that while LLMs bring richer semantic understanding, stronger intent reasoning, and more flexible user interactions, they also introduce novel forms of bias and hallucination-induced issues. The survey systematically categorizes 13 opportunities and 18 challenges across six fundamental dimensions of trustworthiness: robustness, fairness, privacy preservation, accountability, transparency, and ethics.

Concretely, LLMs enable more natural conversational recommendations and can leverage world knowledge to improve relevance, but they also make systems vulnerable to adversarial attacks through prompt injection, propagate subtle biases from training data, and raise new privacy concerns due to their need for extensive user data. The paper also reviews commonly used datasets and evaluation metrics, and outlines critical open challenges such as real-time trustworthiness monitoring and cross-cultural fairness. For professionals building recommendation pipelines, this survey provides a structured checklist of where LLMs help versus hurt trust, and where future research must focus to mitigate new risks before wide deployment.