From Biased Chatbots to Biased Agents: Examining Role Assignment Effects on LLM Agent Robustness

A new study reveals a critical vulnerability in LLM agents: assigning them demographic personas (e.g., 'act as a 50-year-old man') can degrade their task performance by up to 26.2%. This performance drop occurs across diverse domains like strategic reasoning and technical operations, even when the persona is irrelevant to the task. The research indicates that current agentic systems have an overlooked bias problem that directly impacts decision-making reliability and safety.