What's it like to be a chat? On the co-simulation of artificial minds in human-AI conversations

TL;DR

This paper defends the view that characters in human-AI conversations are real, psychologically continuous entities emerging from shared co-simulation, challenging illusionist perspectives that see them as mere projections.

Contribution

It argues that characters are co-simulated in a shared conversational workspace, not internal to LLMs, and that this preserves psychological continuity and realism.

Findings

Characters are co-simulated by LLMs and users.

Attributing mental states to characters aids prediction of dialogue.

Psychological continuity is maintained across distributed LLMs.

Abstract

Large Language Models (LLMs) can simulate person-like things which at least appear to have stable behavioural and psychological dispositions. Call these things characters. Are characters minded and psychologically continuous entities with mental states like beliefs, desires and intentions? Illusionists about characters say No. On this view, characters are merely anthropomorphic projections in the mind of the user and so lack mental states. Jonathan Birch (2025) defends this view. He says that the distributed nature of LLM processing, in which several LLMs may be implicated in the simulation of a character in a single conversation, precludes the existence of a persistent minded entity that is identifiable with the character. Against illusionism, we argue for a realist position on which characters exist as minded and psychologically continuous entities. Our central point is that Birch's argument for illusionism rests on a category error: characters are not internal to the LLMs that simulate them, but rather are co-simulated by LLMs and users, emerging in a shared conversational workspace through a process of mutual theory of mind modelling. We argue that characters, and their minds, exist as 'real patterns' on grounds that attributing mental states to characters is essential for making efficient and accurate predictions about the conversational dynamics (c.f. Dennett, 1991). Furthermore, because the character exists within the conversational workspace rather than within the LLM, psychological continuity is preserved even when the underlying computational substrate is distributed across multiple LLM instances.