Social science is becoming executable, and the subjects are models

Social science is getting a new kind of labor force, and that labor force is also writing the study.

For the first time, the same runtime can host a large language model that reads the literature, designs the experiment, runs the survey, drafts the paper — and a second, separate population of large language models that answer as if they were the human subjects being studied. The result is a method that compresses a research cycle that used to take a field team and a panel company into a loop that can run overnight, and a finding that has to be read as if it came from people when it did not.

Tsinghua's FIB-Lab tested that loop across seven illustrative studies spanning micro-level lab experiments, meso-level social media dynamics, and macro-level urban scenarios, and the preprint claims the agent-run studies reproduce the qualitative patterns of prior human work while surfacing deviations the authors call "informative." The interesting question is not whether the simulation ran. It is what an "informative deviation" means when both the scientist and the respondent are the same kind of model, fine-tuned into different roles.

The payoff for a small team is speed: a graduate student can iterate on a hypothesis the way a quant once iterated on a trading signal. The cost is interpretive — when the participants are large language models behaving like people, a reproduced pattern says something about the models first and about the people they are standing in for only secondarily. Every AgentSociety-style result should be weighed against that order.