When you tell an AI to play a specific kind of person, which traits should you give it, and do they matter?

This explores how a researcher decides which traits to write into a persona prompt, and how much that choice matters. The corpus has no single recipe, but it does have a clear pattern: choose attributes from evidence about real people, not from intuition, and check that the choice does anything before trusting it.

The best-supported approach is to let data pick the attributes. In survey simulation, choosing persona attributes using existing survey data on related questions improves the match between predicted and real answers (Why do persona prompts show such mixed results for surveys?). In practice, you keep the traits that already separated respondents on neighboring questions. The same principle works in product testing. Personas built from anonymized behavioral data predicted the direction of A/B test results with 75 to 90 percent accuracy across 40 experiments (Can behavior-based personas predict A/B test outcomes?). For evaluation work, one system pulls stakeholder personas out of domain documents with semantic clustering. That grounds them in real perspectives instead of arbitrary roles, and the personas carry over across tasks like summarization and dialogue (Can personas extracted from documents generalize across evaluation tasks?).

The attribute people reach for most, expertise, is one of the weakest. Across six models on graduate-level science questions, in-domain expert personas had no significant effect on accuracy. Mismatched experts gave marginal gains, and low-knowledge personas made answers worse (Do expert personas actually improve LLM factual accuracy?). Attributes seem to earn their place when the goal is reproducing how people differ, not making the model smarter. How you specify them matters too. Thin character sketches failed to hold up over a conversation, while a layered persona with a scripted structure produced more human-like dialogue (Can layered persona architecture sustain coherent character behavior?). So the question is also how deep and how structured the attributes are, not only which ones.

Even good attributes only pay off where there is signal to capture. Persona prompting works best on contested questions where humans disagree widely (Why do persona prompts show such mixed results for surveys?). Persona simulations replicated 76 percent of published experimental main effects, but their accuracy tracked how strong the original evidence was, and marginal effects produced both false positives and false negatives (Can AI personas reliably replicate human experiment results?). The behavioral-data personas were likewise least trustworthy for near-zero effects (Can behavior-based personas predict A/B test outcomes?).

Two checks are worth running before you tune attributes. First, when the same persona prompt is rerun, the variation between runs can match or exceed the variation between different personas (Why do LLM persona prompts produce inconsistent outputs across runs?). If swapping attributes moves outputs less than rerunning the identical prompt, the attribute choice is mostly noise. Second, persona conditioning steers what the model says without removing its underlying bias. Between-group sentiment gaps stayed put even when models followed trait instructions (Can persona prompts actually reduce bias in language models?). So attributes are a way to steer outputs, not to correct the model, and you should measure the result instead of assuming it.