What do Large Language Models Represent?

If large language models (LLMs) are models, what do they represent exactly? I address this question by applying the various conceptions of epistemic representation that have been entertained by philosophers of science to this case. After discarding the idea that they would represent the structure of natural languages, I argue that LLMs do not directly represent the world by linguistic means either, but rather a certain class of appropriate linguistic production that is constructed by their makers. They are mostly used to generate fictitious instances of this class, although these fictions can be performative when appropriated by humans. The main implication of this analysis is that conversing with a LLM chatbot amounts to engaging in a fiction, where our own imagination plays a central role. I examine some ethical consequences of this conclusion regarding how LLMs are currently deployed in numerous places to serve as “AI assistants”.