Give an AI character a richer backstory and personality — does that scrub out its built-in stereotypes, or just disguise them?

This explores whether giving a generative agent a more detailed, layered persona (more traits, backstory, context) cleans out the biases the underlying model picked up in training, or only changes how those biases show up. The corpus points to a fairly clear no. Richer personas buy realism and consistency, but nothing here shows them removing bias, and some notes show them adding it. No note directly compares thin personas with rich ones on bias, so the closest evidence is about persona prompting in general.

The most direct evidence is that persona prompts work at the output level. Across three models, persona conditioning made models follow the assigned traits, but the between-group sentiment gaps stayed unchanged. The bias was persona-steering-operates-in-the-output-channel-and-redistributes-measured-bias|redistributed rather than reduced. A second finding is more unsettling: when an LLM is assigned an identity, it can pick up new biases. persona-assigned-llms-exhibit-human-like-motivated-reasoning-that-prompt-based-d|Persona-assigned models were about 90% more likely to accept evidence matching their assigned identity, and standard debiasing prompts did not fix this. So a richer persona can give the model more identity to be biased on behalf of.

What richer input does improve is fidelity. Realistic synthetic dialogue synthetic-dialogue-diversity-requires-persona-subtopic-contextual-characteristic|needs subtopic, Big Five personality, and situational context stacked together. A shallow-character-descriptions-fail-to-sustain-coherent-persona-behavior-deep-pe|three-layer persona architecture produces conversation closer to human patterns than a shallow description. Grounding personas in automated-stakeholder-persona-extraction-from-domain-documents-enables-cross-tas|real stakeholder documents makes evaluation transfer across tasks. These are gains in variety, consistency, and reproducibility, not neutrality. The population-level picture is mixed too: llm-persona-simulations-replicate-76-percent-of-published-experimental-main-effe|persona simulations replicated 76% of published experimental effects, but were unreliable on marginal effects, with both false positives and false negatives. A persona can look human without being an unbiased human.

The corpus offers two reasons for this. First, on the philosophical side, post-training realizationism-holds-that-rlhf-trained-personas-are-realized-quasi-psychologies|installs stable dispositions that persist under adversarial pressure, and the llm-interlocutors-are-best-understood-as-virtual-model-instances-that-realize-pe|virtual-instance account treats these as substrate-level, not a costume. A prompt-supplied persona sits on top of that layer without rewriting it. Second, there is a statistical angle. generative-models-transcend-their-training-experts-through-implicit-majority-vot|Models trained on many imperfect sources cancel out uncorrelated errors through implicit majority voting. My inference, not the note's claim, is that biases shared across most of the training data are the ones this averaging cannot cancel, and low-temperature sampling would concentrate them.

The practical takeaway: if you want a simulated population to be diverse, rich persona input helps. If you want it to be unbiased, the corpus says to measure the outputs directly instead of trusting the persona description to do that work.