Generative-agent social simulation
Simulating populations of individually modeled people with AI agents to predict how a society responds to policies, events and information.
State of the art
Park et al. — Generative Agent Simulations of 1,000 People (2024) reproduce individual attitudes about as well as people's own two-week retest consistency.
- Agents mirror LLM biases rather than real behavioral variance
- No validation against real outcomes over months or years
- Simulating millions of agents with interactions is costly
Who’s working on it
- Joon Sung ParkResearcherStanford University (founder and CEO, Simile)
Built the Smallville generative agents (UIST 2023 best paper) and then agents from two-hour interviews with 1,052 people that replicated those individuals' survey answers
- SimileStartup
Commercialising interview-grounded agent populations for market and policy research; raised $100M Series A in February 2026 and $200M Series B at a $2B valuation in July 2026
- Michael BernsteinResearcherStanford University, Computer Science
Co-author of the generative-agents and agent-bank work and the main voice on what these simulations can and cannot validly predict
- Percy LiangResearcherStanford University (Center for Research on Foundation Models)
Co-founder of Simile and the evaluation methodologist pushing for held-out behavioural benchmarks rather than plausibility demos
- Stanford HAI ↗Academic lab
Hosts and funds the agent-simulation work and the survey-replication datasets that make external validation possible
- Matthew SalganikResearcherPrinceton University
His Fragile Families Challenge (160 teams, PNAS 2020) is the benchmark result showing how poorly individual life outcomes are predicted, and the bar any agent simulation must beat
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| research | Prospective forecasting tournaments pitting agent simulations against real policy outcomes | $$ $10M – $100M |