First survey maps fairness risks in LLM-powered recommender systems

A team from National Taiwan University and IBM Research has released the first comprehensive survey focused specifically on fairness in LLM-based recommender systems (LLM4Rec). The paper, now on arXiv, systematically reviews how large language models introduce new fairness challenges beyond traditional collaborative filtering biases. The authors identify five core bias mechanisms: pretrained knowledge (models inheriting societal stereotypes), prompt engineering (biased instructions skewing outputs), generated explanations (rationales that amplify disparities), decoding strategies (sampling methods favoring certain groups), and feedback loops (user interactions reinforcing initial biases).

To structure the landscape, the survey uses a two-dimensional framework mapping bias mechanisms against fairness targets (e.g., user groups, item categories). It also catalogs existing evaluation metrics and mitigation techniques, from data debiasing to adversarial training. Importantly, the paper links fairness to broader trustworthy AI concerns—explainability, privacy, robustness, and controllability—arguing that these dimensions are interdependent. This work provides a foundational reference for researchers and practitioners building equitable LLM-driven recommendation pipelines.