GPT-4 Accurately Predicts Social Science Experiment Results, Nature Study Shows
RobbWiller · x · 2026-07-23
A new paper published in Nature reveals that GPT-4 can highly accurately predict the outcomes of social science experiments by simulating participants.
- Study Scale: The team evaluated 70 preregistered experiments (469 effects, 119,330 participants) by prompting GPT-4 with study materials and demographic profiles.
- Key Finding: There was a strong correlation (r=0.85; adjusted r=0.92) between LLM-predicted and actual observed effects.
- Anti-Data-Leakage: Correlation remained extremely high for studies unpublished by its training cutoff (r=0.90) and even unpublished by June 2025 (r=0.92), proving the model isn't merely retrieving memorized data.
- Human vs. AI: On a harder archive, GPT-4's accuracy (r=0.39) matched pooled forecasts from 2,659 humans (r=0.31). Averaging human and GPT-4 predictions improved accuracy beyond either alone (r=0.89), as they offered partly independent information.
Related event: Nature Study: GPT-4 Can Predict Social Science Experiment Results(18 posts)→
More from Research
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B — dlwh · 2026-10-27