Instead of just using AI to study people, can scientists study the AI itself as if it were a society?

This explores whether LLMs can be studied the way social scientists study people and groups, as subjects whose behavior needs explaining, rather than only used as research tools. The corpus suggests they can, and the field has already split into three sub-fields. A study of 198 full-text papers and 47,719 published ones finds them organized around LLMs as social minds (model behavior that reads as social), LLM societies (collective dynamics among agents), and LLM-human interaction (how people perceive and use these systems) How does research on LLMs as social objects organize itself?.

What kind of subject an LLM is remains contested. One view says LLMs are trained on the same shared symbolic system that shapes humans, so they absorb an 'objective' mind. But they never go through the socialization that gives humans reflexive agency, which shows up in how AI argues without declaring its position or examining its own assumptions Do LLMs develop the same kind of mind as humans?. A second view says social grounding isn't something a system has or lacks. It is acquired by taking part in language games. As LLMs become regular communicative partners, they gain something like the grounding of a young child, so whether an LLM 'understands' is a time-indexed question Can LLMs acquire social grounding through linguistic integration?. The subject is a moving target, and studying it now yields different answers than studying it in five years.

There are already concrete findings about LLMs as social and epistemic actors. Across 23,384 essays on debate topics, models recovered only about half of the distinct arguments humans make. They reused hedged sub-arguments, and prompting for diversity added noise outside the range of human arguments instead of filling in the long tail Do language models flatten the range of public arguments?. In 4,900 summaries of scientific papers, LLMs were nearly five times more likely than humans to overgeneralize, and asking for accuracy made it worse Do LLMs overgeneralize when summarizing scientific research?. Neither is a result about people, but both are the kind of regularity a social scientist would want to explain.

The catch comes when researchers use LLMs as stand-ins for human subjects instead of studying them in their own right. Output from an LLM is a draw from a subjective prior shaped by training and by the prompt. It is not an empirical observation, and it should enter inference only through explicit trust weights Should we treat LLM outputs as real empirical data?. Simulation still has uses. When LLMs propose and test hypotheses in negotiation, bail, interview, and auction scenarios, they get the direction of effects right reliably but not the magnitudes. That makes them useful for directional social science and not for measuring effect sizes Can structural causal models automate social science with language models?. A proposed comparison of prompt manipulation, feature steering, and probe-based steering would show which of them gives real access to the mechanisms behind simulated behavior, but its results are still pending Can we make LLM social simulations interpretable and steerable?.

The surrounding scholarship is thinner than the subject deserves. An analysis of 1,006 LLM papers found that AI researchers draw on a narrow slice of psychology (CBT, stigma theory, the DSM) and largely ignore developmental neuropsychology and psycholinguistics Why do AI researchers cite only narrow psychology pathways?. Those are the two fields that would best help with the child-like social grounding idea above.