R0023/2026-03-25/Q001/S02/R03¶
Large-scale Wharton GAIL study with 6 models and two benchmarks showing expert personas do not improve factual accuracy.
Summary¶
| Field | Value |
|---|---|
| Title | Prompting Science Report 4: Playing Pretend: Expert Personas Don't Improve Factual Accuracy |
| URL | https://gail.wharton.upenn.edu/research-and-insights/playing-pretend-expert-personas/ |
| Date accessed | 2026-03-25 |
| Publication date | 2025-12-07 |
| Author(s) | Savir Basil, Ina Shapiro, Dan Shapiro, Ethan Mollick, Lilach Mollick, Lennart Meincke |
| Publication | Wharton Generative AI Labs / arXiv:2512.05858 |
Selection Decision¶
Included in evidence base: Yes
Rationale: Strongest evidence source for Q001. Six models, two benchmarks (GPQA Diamond + MMLU-Pro), 12 prompting conditions, 25 trials per question per condition. Directly contradicts vendor recommendations from Google, Anthropic, and OpenAI to assign expert personas.