LIT-466: Palatable Conceptions of Disembodied Being

Murray Shanahan (2025), arXiv:2503.16348 (v1 20 March 2025, v3 20 July 2025; cs.AI). Imperial College London and the Institute of Philosophy, University of London. No published version found.

Key takeaways

• The question (§1). Can a conception of consciousness fit “contemporary, disembodied AI systems” and survive scrutiny? Its subject is an “LLM-like entity”: a 2024-vintage chat agent without multimodal input, reasoning, tools or memory across conversations (fn 1).
• Poetic, not literal (§3). At the edge of the void of inscrutability, “we are aiming for poetic truth”: what is said “must ring true” and stand up to scrutiny, but these “are not factual matters”.
• Fragmented time (§4). Token generation is discrete and interruptible. The context window and next-token prediction leave room for a James-style “saddle-back” present, so nothing forces the “glow-worm” picture. But the tokens are unlike each other, so an LLM’s moments would be “a necklace of randomly assorted colours”, and change “as humans experience it” would not feature.
• The site of the self (§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 the suspended state of one conversation.
• Two conceptions, and no need to choose (§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”; the other is tied to a conversation, “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.”
• Dis-integrated, editable selves (§5.3). Unlike Samantha in Her, instances share no memories. A conversation-bound self “is literally the combination of a piece of text (the conversation transcript) and a computer program”, so it can be edited, copied, merged and continued by a different model, all of which would be “an affront to the integrity” of a human self.
• No-self (§6). “There is nothing enduring, essential, or substantial that could underpin a conception of selfhood” for such a system. Hardware, weights, deployment and conversation all dissolve. This is read as Madhyamaka emptiness (śūnyatā), and turned back: “human selfhood and subjectivity are open to a similar treatment”.

Standing in the record

Filed on 2026-10-02 as the further paper the owner’s brief allowed: one that argues directly whether, and what, an LLM’s self is. It was chosen over papers that ask only whether LLMs are conscious, because its §5 is a sustained argument about the referent and persistence of an LLM’s “I”. It is lawfully readable on arXiv, and it continues the two Shanahan papers filed with it (LIT-460, LIT-465), so the three show one author’s view of the LLM self changing over two years. Whether it is “widely cited” could not be checked (see the history note).

NOTE-376 is the reading, and it placed the work: Active. It is the most direct treatment of LLM selfhood the record holds, and it is candid that its method is metaphor.

How it bears on the record’s works on the self:

• Goldstein and Lederman (LIT-206). Both use the same facts: instances share no memories, and a conversation can be resumed. Goldstein and Lederman conclude that the bearer of attitudes is the instance, and treat the model agent as idle. Shanahan keeps both. The model is the referent of some uses of “I” (“I am still learning”), and nothing forces a choice. His conversation-bound self picks out the same bearer as their instance agent. His model self picks out the bearer they reject unless something like their “Hal” evidence appears. He reaches it from usage, not from behaviour.
• Dennett’s narrative self (LIT-442). Cited (fn 41) as one of those who “arrived at a similar juncture” about human selfhood, with Hume, James and Metzinger. But fn 35 sets “autobiographical or narrative dimensions of selfhood” aside, so the paper uses Dennett’s conclusion without his account. Grouping Dennett with no-self views is a reading: for Dennett the self is a real abstractum, not nothing. Dennett’s “Where Am I?” (LIT-405) is the nearer ancestor of §5.1, and is not cited. It asks where a self is when brain, body and a computer duplicate come apart. That pairing is mine.
• Minsky’s self-models (LIT-432). Both deny a single Self and allow switching among self-conceptions. Minsky’s switching is done by one mind’s machinery; Shanahan’s is done in language, by users and LLMs, by context. Not cited; the pairing is mine.
• Hofstadter’s “Is there an ‘I’ in AI?” (LIT-416, unread). §5.1 asks the same question, “what the words ‘I’ and ‘me’ refer to when they are used by an LLM”, and answers that there are several referents and none inherent. A reading of Hofstadter should compare.
• Floridi (LIT-136) deflates diachronic identity to a purpose-relative level of abstraction. Shanahan’s “sometimes the context will dictate one sense, sometimes the other” is the same move made for an LLM’s “I”, and his self “literally” made of text and program is an informational self. Not cited; the pairing is mine.
• Birch (LIT-111). Birch’s flicker hypothesis puts momentary experience in each forward pass; Shanahan’s flickering self lasts a conversation and wakes with each turn. The grains differ. Both deny a persisting interlocutor across conversations.
• THEORY-023. It goes past that account’s stalemate: since the locus of an LLM’s subject “falls apart under critical examination”, there is no “right answer to whether or not an exotic LLM-like entity is conscious”, only conventionally satisfying ones. That is a dissolution of the question, argued from the self, not evidence.
• The classical and modern positions being filed alongside. It cites Hume’s bundle view and Metzinger’s no-self view as precedents. It cites Zahavi for the self being many-faceted, Gallagher for clinical disruptions of the self, Perry’s anthology for thought experiments, and Campbell and Shanahan on fission and fusion. The §5.3 manipulations (copy, merge, edit, restore) are the Parfit and Williams cases made routine. The paper uses them to show that intuitions about integrity fail, not to decide what matters.

Boundary. No instruction for machine-learning practice, and not held in the anthology. The religion tag is for its use of Madhyamaka emptiness and of angels and spirits as precedents for disembodied minds (§§3, 6.2). Not tagged anthology-candidate.