The Sociolinguistic Interpretability of Large Language Models: Directional Accuracy but Social Simplification

As rapidly improving language‐generation systems, Large Language Models (LLMs) represent an emergent site for studying sociolinguistic phenomena. We propose a framework for investigating sociolinguistic sensitivity in LLMs: the extent to which they capture not just general linguistic patterns but also the socially structured variation that underlies them. To illustrate this framework, which we label sociolinguistic interpretability , we analyze two new GPT‐generated corpora and compare them with established findings on lexical variation in the United States and on language variation and change in Toronto English. In the dialect survey comparison, the model reproduces well‐established geographical associations but compresses local competition into coarser categories. In the generated Toronto interviews, the model reproduces the apparent‐time rise of be like with striking fidelity, even as it smooths its onset and fails to reproduce robust social stratification. This suggests that LLMs may approximate large‐scale trajectories of variation and change while missing key important social mechanisms.