Bias & Persona

The Problem

Language models are trained on human data, and they inherit both its biases and its many voices. The result is a system whose behavior can shift depending on the "persona" it adopts, sometimes in ways that reinforce harmful stereotypes or produce inconsistent, unreliable outputs.

Key challenges include:

• Inherited bias — Models absorb social, cultural, and demographic biases from training data and can amplify them at scale.
• Persona drift — A model's tone, values, and reliability can change as it takes on different roles or characters, making behavior hard to predict.
• Measurement — Bias and persona effects are subtle and context-dependent, making them difficult to detect and quantify.
• Steering without erasing — Mitigating harmful bias while preserving a model's usefulness and expressive range is a delicate balance.

What We're Working On

• Bias measurement — Developing methods to detect and quantify demographic and social biases in model outputs.
• Persona analysis — Studying how models represent and switch between personas internally, and how this affects safety and consistency.
• Mechanistic links — Connecting bias and persona behaviors to specific internal representations, so they can be understood and steered rather than merely patched.