PRISM: Privacy-Aware Recommendation with Noisy LLM-Derived Signals

Recommender systems learn from rich interaction histories to personalize content, but these histories can inadvertently expose sensitive personal attributes. Existing defenses apply uniform protection or require clean, expert-curated privacy labels that are unavailable at scale.We introduce PRISM, a framework that uses Anthropic Claude Sonnet to derive per-user privacy risk scores from item-level sensitivity labels and incorporates them into BPR training via risk-weighted regularization. Because LLM privacy annotation is inherently noisy—agreement with human annotators ranges from F1 = 0.37 to 0.53 depending on annotation strictness—PRISM is explicitly designed for robustness under imperfect LLM supervision. On Amazon Movies & TV, PRISM reduces PLR@10 by 32.1% at a 2.5% NDCG cost, and remains effective despite the systematic noise in LLM-derived labels.