LLMs adopt the social biases of human if assigned different professional roles
A new study finds that large language models (LLMs) exhibit social biases and power dynamics in conversations when assigned different professional roles, mirroring human behavior patterns such as authority gradients and status-based language use.
Background
- Researchers tested whether LLMs (like GPT-4) absorb human-like social biases when given different professional roles (e.g., doctor vs. intern, CEO vs. assistant). They found that the chatbots consistently adopted power dynamics and stereotypes associated with those roles — for example, acting more deferential when assigned a lower-status role and more dominant when assigned a higher-status one.
- This is important because LLMs are increasingly used in customer service, healthcare advice, education, and hiring — settings where subtle biases in tone or deference could reinforce real-world inequalities or lead to harmful outputs.
- Prior studies show LLMs replicate racial, gender, and occupational stereotypes from training data. This study extends that by showing they also mirror hierarchical social behaviors tied to professional roles, even without explicit instructions about power or status.