Xu-Evans Nature review maps four LLM roles in social science

A new review in Nature Human Behaviour sorts the current wave of LLM use in computational social science into four distinct jobs. The authors, led by Fengli Xu at Tsinghua with James Evans of the University of Chicago as senior author, argue the models now function as "subjects that simulate human behaviour, assistants that automate analysis, collaborators that generate hypotheses and evaluators that audit and improve research quality."

As simulated subjects, LLMs prompted with demographic or ideological profiles "reproduce survey distributions, economic games and polarization dynamics," the authors write. As assistants, AI interviewers can deliver protocols "with quality comparable to human interviews," and models that code qualitative data surface patterns a human coder would miss.

The review, published 26 October 2026, spans sociology, political science, psychology, anthropology and economics, pulling in work from groups at Tsinghua University, HKUST, Microsoft Research Asia and the University of Chicago. It also catalogues the limits. The paper flags that LLMs struggle with causal inference, that their capacity to reproduce social complexity remains uneven and unproven, and that privacy and misinformation concerns are not hypothetical.

Two researchers we track in our Who's Who directory circulated the paper this week.