On the Use of Large Language Models in the Social Sciences: Promise, Limits, and Epistemic Risks

This talk examines the use of large language models (LLMs) in the social sciences. I discuss to what extent LLMs can produce outputs that are representative of social trends, and whether they can be reliably used to annotate or analyze social science data. Beyond the well-documented concerns about representativity and reproducibility, I argue that the very training and alignment strategies underlying these models raise deeper epistemic issues. LLMs are shaped by optimization and alignment techniques that are far from neutral, and these processes inevitably influence the patterns, norms, and values reflected in their outputs. The talk builds in part on arguments developed in my book « Understanding Conversational AI », Ubiquity Press (esp. chapter 6), and aims to clarify both the potential and the structural limitations of LLMs as tools for social scientific inquiry.