← News

ResearchSources: arXiv · 18 Nov 2026

Rich self-report data make AI person simulations more accurate, arXiv study finds

A research paper reports that simulations built from interviews and surveys of 1,052 U.S. adults outperform demographic-prompted simulations, though gaps to human consistency remain.

In plain words

Researchers tested whether AI copies of people become more accurate when fed detailed personal data instead of just basic demographics. They interviewed and surveyed over a thousand U.S. adults, then asked AI models to predict those people's answers. The richer data helped, but the AI copies still fell short of how consistently the real people answered.

What was announced

The paper, posted on arXiv, describes an AI voice interviewer that conducted two-hour interviews with 1,052 U.S. adults following the American Voices Project protocol; participants also completed surveys and behavioral experiments. The authors compared simulations prompted with demographic attributes, self-authored personas, survey responses, interview responses, or a combination, testing accuracy on held-out responses and measuring eight sample-level distortions such as biased averages, compressed variance, and over-correlated items.

On General Social Survey items, the authors report demographic simulation reached 60.4% accuracy, interview and survey simulation about 69%, and combined simulation 72.8%, against 79.5% for participants' two-week test-retest agreement. An open-weight model fine-tuned on other participants' data reached 75.2%. Combined simulation closed a median of 53.5% of the distortion gaps, and the fine-tuned model 69.3%. The authors note both model types reduced distortions without erasing them, did not improve prediction of incentivized economic choices, and left hyper-rationality in economic games unchanged.

What it does not show

  • The source is the paper's abstract only; methods, peer review status, and full results are not available for verification.
  • All figures and findings come from the paper's authors; no independent confirmation is included in the evidence.
  • The study covers U.S. adults and specific GSS items and economic games; accuracy on other populations or tasks is not established.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.