Persona prompts change model tone but not core biases in language models
The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
Computation and Language
Summary
People try to reduce bias in large language models (LLMs) by giving them different personas or roles in their prompts. This paper shows that while these personas can make the models sound different or express certain traits, they don't actually change the deep internal biases of the models. Instead, the personas mainly tweak the model’s final outputs without fixing persistent bias patterns. The authors measured this across various ways of testing and found that the biases inside remain even when the model talks differently.
What this means in practice
- •For ai system developers: Design bias mitigation strategies that target model internals beyond simple persona prompts to achieve deeper debiasing.
- •For content moderation teams: Recognize that prompt-based persona changes in LLMs can alter tone but may not reduce underlying biased outputs when monitoring content.
Authors
Ziyue Feng, Hongbo Fang, James A. Evans
Abstract
Prompt-based interventions: system prompts, personas, role instructions, reliably reshape what a language model says, but it is unclear which layer they reach. Do they reconfigure internal structure, or only modulate the output channel? We use persona conditioning as a controlled probe, measuring its effects along a depth axis from self-report, through open-ended generation, to word-level parametric association, across three instruction-tuned models. We find a graded dissociation. Personas are legible but not structural: models follow single-trait instructions yet fail to reproduce human inter-trait covariance. The dissociation deepens with depth: personas hold or amplify closed-form QA bias, shift absolute tone while leaving between-group disparity unchanged, and barely perturb an already saturated associative baseline. Prompt-based steering thus operates in the output channel and has a structural reach limit that surface manipulability can mask.