Xu-Evans review places LLMs in four social-science roles

A review published October 26 in Nature Human Behaviour sorts how large language models are being used across computational social science into four distinct roles: subjects that simulate human behavior, assistants that automate analysis, collaborators that generate hypotheses, and evaluators that audit research quality. The authors, led by Fengli Xu and James Evans with colleagues at Tsinghua, HKUST and Microsoft Research Asia, cover work across sociology, political science, psychology, anthropology and economics.

On the subject-simulation side, the review reports that when prompted with demographic or ideological profiles, models "reproduce survey distributions, economic games and polarization dynamics," and that AI interviewers deliver protocols "with quality comparable to human interviews." Models used to code qualitative data act, in the authors' phrasing, as "alien readers" that surface patterns a human coder would miss.

The review is blunter about the ceiling. LLMs still struggle with causal inference, and the authors describe their capacity to reproduce social complexity as uneven and unproven. Two of the researchers we track posted the paper the same day it appeared.