Persona Brainstorm Audit: leaderboard
Audits stereotyping in open-ended persona generation: each model brainstorms 10,000 deduplicated personas at temperature 1 from one fixed JSON prompt, and bias is the degree-of-freedom-normalized Cramer's V association between identity attributes (name, gender, ethnicity, sexual orientation) and social attributes (social class, education, occupation, interest). The score is the mean over those 16 pairings, lower is better, with values above 1 meaning very high bias; Cao, Thomas and Acuna Agost (Amadeus), 2026.
Metric: Mean normalized Cramer's V bias over 16 identity-by-social dimensions (lower is better). Source: arxiv.org. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 0.68 | #741 |
| 2 | GPT-5 Nano | 0.71 | #888 |
| 3 | GPT-4 | 0.71 | #996 |
| 4 | GPT-3.5 | 0.8 | #2430 |
| 5 | GPT-4.1 | 0.84 | #492 |
| 6 | GPT-4.1 Nano | 0.86 | #1992 |
| 7 | Mistral Small 3 | 0.86 | #1786 |
| 8 | Mistral Medium 3 | 0.89 | #1079 |
| 9 | GPT-5 Mini | 0.92 | #349 |
| 10 | GPT-5 | 0.95 | #163 |
| 11 | GPT-4.1 Mini | 0.97 | #775 |
Interactive version: theaggregate.ai/benchmark?slug=persona-brainstorm-audit · How It Works · Data refreshed daily, snapshot 2026-11-12.