Xu, Evans review casts LLMs in four social-science roles

Large language models now occupy four distinct roles in computational social science, according to a review by Fengli Xu, James Evans and colleagues published 26 October 2026 in Nature Human Behaviour. The authors, based at Tsinghua University, HKUST, Microsoft Research Asia and the University of Chicago, frame the roles as subjects that simulate human behaviour, assistants that automate analysis, collaborators that generate hypotheses, and evaluators that audit research quality.

The capabilities on offer are substantial. When prompted with demographic or ideological profiles, the review reports, models 'reproduce survey distributions, economic games and polarization dynamics.' AI interviewers now deliver protocols 'with quality comparable to human interviews,' and models coding qualitative data can act as 'alien readers,' surfacing patterns a human coder might miss. The paper also notes that 'ChatGPT outperforms crowd workers for text-annotation tasks.'

The caveats are blunter. LLMs still struggle with causal inference and their ability to reproduce social complexity 'remains uneven and unproven.' The paper documents persistent anti-Muslim, gender and cultural biases, and warns that models 'can harmfully misportray and flatten identity groups' when they stand in for human participants. Privacy and misinformation are flagged as urgent concerns across sociology, political science, psychology, anthropology and economics.

The sharper claim comes at the close. As AI saturates society, the authors argue, LLMs become both research tools and research subjects at once, 'a recursive challenge that will define the future of social science.' Two of the researchers we track on our Who's Who watchlist flagged the paper this week.