Tell an AI to 'act as' someone and its answers don't just get noisier — they lean the same way every time.

This explores why telling a model to 'act as' a particular person pushes its answers in a consistent direction, instead of just adding random noise. The corpus has no single note that answers this outright, but read together the notes point to a few forces working at once. The first is that the frame changes the task. Shanahan's view is that a dialogue agent is producing text that fits a character, not reporting real beliefs Should we treat dialogue agents as role-playing characters?. So 'predict what this kind of person thinks' quietly becomes 'write what this kind of person would sound like'. The model's version of that character is built from whatever stereotypes are in its training text.

The bias also reaches the reasoning, not just the voice. Models given a persona were about 90% more likely to accept evidence that matched their assigned identity, a human-like motivated reasoning, and standard 'please be unbiased' instructions didn't fix it Do personas make language models reason like biased humans?. A companion finding is that persona prompts work only at the output level. Across three models, the sentiment gaps between groups stayed unchanged, so the bias was redistributed rather than removed Can persona prompts actually reduce bias in language models?. The persona seems to pick which of the model's existing associations get expressed. It doesn't rewrite them.

The model's own training also leaks in, and the errors lean in a predictable direction. On the Moral RolePlay benchmark, scores fall from 3.21 for moral paragons to 2.62 for villains. Models fail most on deception and manipulation, swapping in crude aggression for nuanced malevolence Does safety alignment harm models' ability to roleplay villains?. This may be because post-training installs personas that persist under adversarial pressure, while prompt-induced role-play collapses under jailbreaks Are RLHF personas performed characters or realized dispositions? Are LLM personas realized or merely simulated through training?. If the built-in character is the stronger force, a thin roleplay frame will keep getting pulled back toward it.

Thin frames also leave gaps that the model fills unpredictably. When the same persona prompt is run repeatedly, the variation between runs matches or exceeds the variation between different personas, so model uncertainty is doing more of the work than persona-specific knowledge Why do LLM persona prompts produce inconsistent outputs across runs?. Where personas do work, they have more to lean on. Simulations replicated 76% of published experimental main effects, tracking how strong the original evidence was, but marginal effects produced both false positives and false negatives Can AI personas reliably replicate human experiment results?. Expert-written profiles paired with retrieved memories improved predictions of character choices Can LLMs predict character choices from narrative context?. Layered, scripted personas held together better than shallow descriptions Can layered persona architecture sustain coherent character behavior?. The pattern is that a roleplay frame is strong enough to push a model toward a stereotype but too weak to supply the specifics, and that gap is where systematic bias shows up.