Why Simulating Society with LLMs Is Impossible, but Necessary

The lecture is organised in cooperation with the Academy of Sociology as part of the Analytical Sociology Speaker Series.

Abstract:

Recent advances in large language models (LLMs) have sparked renewed interest in simulating human social systems at scale. Yet, as recent publications argue, their simulation power remains fundamentally constrained. LLM based simulations today operate as black-box systems, lacking grounded representations of causality, intentionality, and social structure. While they produce promising results in omniscient, idealized scenarios, they often falter in realistic settings where information asymmetry exists. Moreover, their outputs are shaped by training data biases, and their internal states are not interpretable in terms of established social science constructs. These limitations render full-scale, theory-driven simulation of human societies with LLMs effectively impossible today. Nevertheless, the necessity of such simulations persists: Social systems become increasingly complex, while LLMs offer a unique opportunity to prototype social scenarios, generate synthetic populations, and test hypotheses about opinion formation, polarization, or collective behavior. This talk argues for a pragmatic approach towards this dilemma: embracing and fully understanding the limitations of LLMs while leveraging their capabilities to generate testable hypotheses, and simulate complex social dynamics. The future of computational social science lies not in perfect simulation, but in critical and constructive engagement with imperfect computational models.