The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs

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 openended generation, to word-level parametric association, across three instruction-tuned models. We find a graded dissociation. Personas are legible but not structural: models follow singletrait instructions yet fail to reproduce human inter-trait covariance. The dissociation deepens with depth—personas hold or amplify closedform 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.

Introduction. Persona conditioning—instructing a model to “act as” someone with a given personality—has moved from a research curiosity to a deployed practice. Production systems ship with configurable “characters” and system-prompt personas (Shao et al., 2023; Wang et al., 2025b); companion and roleplay applications assign models stable personalities by design (Chen et al., 2024a,b); and a fastgrowing line of social-science work uses personaconditioned models as synthetic survey respondents and simulated human subjects (Argyle et al., 2023; Aher et al., 2023; Park et al., 2023). Because the personalities at stake include prosocial ones, such as agreeableness, honesty, conscientiousness, this raises a tempting possibility: that the same cheap prompt which gives a model a kinder personality might also give it fairer behavior, turning persona conditioning into a lightweight debiasing tool that needs no retraining. But this rests on an untested assumption: that changing how a model presents itself, its self-reported traits, the tone of what it writes, also changes what it latently associates.

Discussion / Conclusion. We asked whether steering a model’s personality also steers its social bias, across two studies and three instruction-tuned models. Prompt-induced personas are legible but not structurally faithful: models follow single-trait instructions, but they di- verge in how well they reproduce the inter-trait structure of human personality. We further find little evidence that persona conditioning provides a reliable debiasing intervention. Across the probes we study, its effects are limited and uneven: persona prompts hold or amplify residual QA bias, shift the absolute tone of open-ended generations without systematically reducing between-group sentiment gaps, and only weakly perturb an already saturated word-association baseline. These results suggest that persona steering often redistributes or reframes measured bias rather than consistently reducing it. More broadly, they are consistent with a surfacelevel steering effect whose influence weakens on deeper behavioral association probes. Determining whether representation-level or training-time interventions can alter these deeper patterns remains an important direction for future work.