Are AI language models scientific models of language?

Abstract

Abstract There has been a recent surge of philosophical interest in AI large language models (LLMs) and their relevance for the understanding of human language. In this article, we take the perspective of the philosophy of science on modelling in general to consider whether LLMs are (good) models of language. We establish that LLMs can be considered scientific models, and that as such they are good scientific models of certain aspects of human language, in so far as they can capture the modal structure of natural language. We argue that they can do this even if they do not reveal (all) its causal structure, nomological structure or mechanisms.