What would it mean for a large language model to have a "self"? A new paper on arXiv takes up this question, arguing that the answer is neither a particular word nor a single privileged layer inside the network. Instead, the authors propose a relational account: a functional self is an actor-centered organization of information.

This framing shifts attention from locating the self in a specific component to understanding it as a pattern of how information is arranged and used. The self, in this view, is not a thing but a way of organizing information around the model's role as an actor—something that acts and responds in a coherent way.

The paper's proposal offers a way to talk about selfhood in LLMs without relying on human-like consciousness or a fixed internal structure. It suggests that what we call the "self" may emerge from the relationships between information and the model's actions, rather than being stored in any one place.