Bias in LLM Agents: The Impact of Personas

Large Language Models (LLMs) are increasingly deployed as autonomous agents, capable of actions with real-world impacts beyond simple text generation. A recent study has highlighted how assigning 'personas', i.e., demographic profiles, to these agents can introduce significant biases and negatively influence their performance.

The research has shown that variations in performance can reach 26.2% due to irrelevant cues related to the assigned 'persona'. This phenomenon has been observed in various types of tasks, including strategic reasoning, planning, and technical operations, and manifests across different model architectures.

Vulnerabilities in Agentic Systems

The study's findings reveal a previously underestimated vulnerability in LLM-based agentic systems: the assignment of 'personas' can introduce implicit biases and increase behavioral volatility. This raises significant concerns for the safe and robust deployment of LLM agents in real-world contexts.