Xu and Evans map four LLM roles in computational social science

Large language models now function in computational social science as four distinct kinds of instrument: subjects that simulate humans, assistants that automate analysis, collaborators that generate hypotheses and evaluators that audit research quality. That taxonomy comes from a Nature Human Behaviour review published 26 October 2026, led by Fengli Xu and James Evans with coauthors at Tsinghua University, HKUST and Microsoft Research Asia.

The authors argue the models are already doing more than analysis. Prompted with demographic or ideological profiles, they "reproduce survey distributions, economic games and polarization dynamics," and AI interviewers deliver protocols "with quality comparable to human interviews." When set to code qualitative data, the review calls them "alien readers," surfacing patterns a human coder would miss.

The caveats sit in the same abstract. LLMs "struggle with causal inference and raise urgent ethical concerns about privacy and misinformation," the authors write, and their ability to reproduce social complexity "remains uneven and unproven." Evans posted on X that the thread covers "where the field is fooling itself."

The review spans sociology, political science, psychology, anthropology and economics. It closes on the recursion the authors keep returning to: "As artificial intelligence reshapes human society, understanding this transformation requires the very tools driving it."