Golden Gate Bridge, as Always? Eliciting Prototypical Places From Autoregressive Large Language Models via Category Production

ABSTRACT Among the various theories of categorization, Eleanor Rosch's prototype theory stands out as both influential and contested. In contrast to the classical theory of concepts, prototype theory posits that humans conceptualize the world using category structures where exemplars vary in their degree of membership. While this theory remains well‐regarded among cognitive psychologists as an account of human cognition, we know little about whether artificial agents—particularly large language models (LLMs)—also use a similar type of categorization. Given that these models have shown to mirror human‐like cognitive abilities, an important question arises: Do they also encode categories with graded membership? This question also invites a broader ethical reflection: While substantial research has documented a lack of diversity in LLM outputs, few has realized that such bias can be understood as a graded structure in the model's internal representations of categories. In this study, we explore this possibility by examining the geographical conceptualizations of 11 autoregressive LLMs. Guided by prior work that advocated empirical, inductive testing for studying geographic categories, we employ a cognitive‐psychology experiment known as category production . It prompts a model to generate exemplars of a given category. This approach yields an extensive list of category norms (i.e., typical items associated with a category) across 297 geographic categories, revealing their graded structures both encoded within individual models and collectively. Notably, we discover a strong presence of prototypical places shared across different sampling temperatures, geographic categories, and models. Furthermore, we observe and explain a lack of geographic diversity in LLM outputs through the lens of prototype theory.