This extended preprint argues that the claim that large language models have no bodies rests on a category mistake: it reduces the condition for embodiment to the possession of a biological body. The paper defines a body as a functional boundary between system and environment, distinguishes three levels of Markov blanket, and argues that an LLM-in-operation satisfies the conditions for a functional Markov blanket during a session. It then shifts the question from whether an LLM has a body to what type of body it has and how thick that body is, introducing Markov blanket thickness as a secondary measure and applying it to 472 LLM session logs. This version has not been peer reviewed and is not the version of record.
Tolmetes (Fri,) studied this question.