Source
Shanahan (2025), “Palatable Conceptions of Disembodied Being”, arXiv:2503.16348 (v3), LIT-466, read in NOTE-376. Shanahan, McDonell and Reynolds (2023), “Role play with large language models”, Nature 623: 493–498, LIT-460, read for the self in NOTE-370, which also records where the arXiv version differs.
What was actually shown
Role play (LIT-460). A dialogue agent can be seen as playing one character, or as “a superposition of simulacra within a multiverse of possible characters”. The simulator, the base model with sampling and an interface, “contains multitudes”; simulacra “only come into being when the simulator is run”. A character that says “I” carries the self-conception the corpus and the conversation give it: each simulacrum has “its own theory of selfhood consistent with the dialogue prompt and the conversation”, and the superposition narrows as the conversation rules theories out (pp. 497–498). Candidate sites a character might protect are the hardware and, on a “substrate neutral” theory, the computational process; many concurrent instances make “the picture more complicated”. NOTE-370 rates the framing moderate as a description.
Candidate referents (LIT-466 §5.1). When an LLM says “I”, the referent might be the model (architecture plus weights), the deployed process, all its concurrent instances, one instance, or the model plus one conversation’s suspended state.
Two conceptions (§5.2). “I am still learning” must refer beyond the conversation; “I did not say that”, of another window, must refer within one. So one self is the model, “manifest as multiple, simultaneous instances”, and the other is tied to a conversation, which “flickers into being with each user interaction, and lies dormant in the gaps”. “We (the users and the LLMs) can legitimately slip from one usage to another.” The two usages are invented, not sampled.
What a conversation self is (§5.3). It “is literally the combination of a piece of text (the conversation transcript) and a computer program”, so it can be “arbitrarily edited, copied, and merged” and continued by a different model. Instances “have no access to each other’s memories or experiences”. NOTE-376 rates this strong as a description of how chat systems work.
What could have come out otherwise. Usage could have fixed one referent, so that every use of “I” read as the instance, or as the model. The two invented examples are offered to show it does not; whether real usage behaves so is untested.
What this does not say
- That there is “no-one at home”. The role-play paper says so of self-preservation talk, and that the simulator has no beliefs, goals or agency “not even simulated versions”. The reading rates both weak: asserted from the framing, with no criterion for having an attitude. Birch (LIT-111) replies that role play “does not rule out a conscious ‘actor’”. This account is about reference, and it does not bear on THEORY-023.
- That no candidate is a self. Shanahan’s no-self conclusion, that “there is nothing enduring, essential, or substantial” to underpin selfhood for such a system (§6.1), is asserted as the upshot without testing each candidate against a criterion. His extension to human selves (§6.2) is an assertion, and the 2025 paper says it aims at “poetic truth”, not literal truth (§3).
- What makes such an agent the same over time. The role-play paper calls the criteria “far from clear” and cites one work on personal identity, Perry’s anthology. Neither paper adopts a criterion.
- Anything about agents with tools or memory across conversations. The 2025 paper restricts itself to an early-2024 chat agent (fn. 1). Memory across conversations could restore a single persisting referent or add a third conception.
- Which conception a deployed assistant in fact enacts. That is testable by regeneration and is not done.
- That selfhood is agency. The papers’ sites of a self are the bearers Goldstein and Lederman make agents; ADR-024 keeps the two apart, and picking out the same bearer is not the claim that self is agent.
Connections
- The rival view, not filed as a THEORY. Goldstein and Lederman (LIT-206, NOTE-138) make the per-conversation instance the bearer of beliefs and desires and find no work for a model agent. Shanahan’s conversation self picks out their instance; his model self is the bearer they reject. They also argue that role play is incompatible with attributing belief, and NOTE-370 finds that version in the published text.
- Floridi (THEORY-650). Letting context fix which “I” is meant is the level-of-abstraction move, as NOTE-376 observes; it is not cited.
- Dennett (THEORY-691). A simulacrum has a fictional character’s indeterminacy; Dennett would count a coherent one a self, and the role-play paper counts it no one. Dennett’s “Where Am I?” (LIT-405) asks §5.1’s question of a human split across brain, body and duplicate.
- Branching and counting. The §5.3 operations, copy, merge, edit and restore, are the cases of Parfit (THEORY-447) and Lewis (THEORY-476) made routine. Shanahan uses them to show that intuitions about integrity fail, not to decide what matters.
- The record’s own claim. CLAIM-029 holds that a language model’s response to a context is functionally a parameterized pragmatic frame, and grounds on LIT-460 and on this account, for the role-play framing stated above.
- THEORY-173 treats large models as a cultural technology rather than agents and calls the simulator account compatible from the other side.
Note of 2026-10-29: what holds it back
No sample of deployed conversations has been coded for the referent of ‘I’. The two conceptions rest on Shanahan’s two invented usages, and no regeneration study shows which self-conception a fine-tuned assistant enacts. Products with memory across conversations, which could restore one persisting referent, are outside both papers.