Nature review maps four roles for LLMs in social science

A new review in *Nature Human Behaviour* sorts large language models in computational social science into four distinct roles: "subjects that simulate human behaviour, assistants that automate analysis, collaborators that generate hypotheses and evaluators that audit and improve research quality." The authors, led by Fengli Xu with James Evans as senior author, argue the four roles are "transforming computational social science" and creating "new methodological paradigms."

The review is less celebratory than the taxonomy suggests. The authors write that LLMs "struggle with causal inference and raise urgent ethical concerns about privacy and misinformation, and their ability to reproduce social complexity remains uneven and unproven." Where LLMs stand in for human participants, the paper documents cases where models "can harmfully misportray and flatten identity groups," alongside "persistent anti-Muslim bias" and gender-related opinion exaggeration in reviewed studies.

Simulation results are mixed. The paper reports LLMs reproducing "human trust behavior" in certain experimental games, while other human-like reasoning patterns "disappeared in newer versions." The authors call validation "the central challenge for generative social simulation."

They close on a recursive note: understanding how AI reshapes society "requires the very tools driving it," a challenge the authors say "will define the future of social science."