This research investigates how Large Language Models (LLMs) can perpetuate stereotypes and biases in recommendation systems, and proposes novel methods to detect and mitigate these issues.

Key findings:

• LLM-based recommendation systems can inherit stereotypes from training data
• These biases create fairness vulnerabilities and diminish trustworthiness
• The paper presents a framework to identify and reduce stereotype-aware unfairness
• Mitigation techniques improve fairness without significant performance degradation

Why it matters: Security and fairness in AI systems are essential for building trustworthy recommendation platforms that don't discriminate against users based on sensitive attributes or perpetuate harmful stereotypes.

Investigating and Mitigating Stereotype-aware Unfairness in LLM-based Recommendations