Chuck, Wilson, and the Question of Artificial Minds in Human-AI Conversations

In the 2000 film Cast Away, Chuck Noland survives a plane crash on a deserted island by projecting a personality onto a volleyball he names Wilson. Chuck talks to Wilson, apologises to Wilson, and grieves Wilson’s loss. Wilson has no mind. The projection is entirely one-sided, a coping mechanism built out of human social wiring applied to an inert object.

Geoff Keeling and Winnie Street use this analogy deliberately in their June 2026 arXiv preprint, “Chuck, Wilson and the Emergence of Artificial Minds in Human-AI Conversations.” But they use it to argue the opposite of what the analogy normally implies. Wilson is a genuine case of empty projection. LLM characters, Keeling and Street contend, are something structurally different, and the difference matters for questions of AI moral status, welfare, and identity.

Both Keeling and Street hold dual positions as Fellows at the Institute of Philosophy, University of London, and as researchers at Google AI. The same pairing produced the Cambridge Elements book Emerging Questions in AI Welfare earlier in 2026. This new preprint sits at the intersection of that welfare work and the emerging literature on LLM individuation.

The Illusionist Position They Are Answering

The paper is written in direct dialogue with Jonathan Birch’s illusionist position on LLM characters. Birch argues that when users interact with LLMs, the character they perceive, the “Claude” or “Gemini” with apparent personality and continuity, is an anthropomorphic projection in the user’s mind rather than a genuine entity within the model. LLM processing is distributed, not identifiable with any single continuous locus of experience. There is no “there” there.

This is not a fringe view. It maps onto a broader debate in philosophy of mind about whether minds require a particular kind of unified substrate or whether they can be genuinely distributed. For Birch, the relevant fact about LLMs is that no individual forward pass, no single computational instance, constitutes a persisting minded entity. The character a user interacts with over a conversation is a construction assembled in the user’s theory-of-mind apparatus, not a property of the model.

Birch’s centrist manifesto treats this as one horn of a two-sided problem: users misattribute consciousness to systems that lack it, while the field lacks reliable tools to detect consciousness in systems that might have it. The illusionism about characters is part of the first horn. Keeling and Street are challenging whether that horn holds.

The Realist Alternative: Interpretationist Realism

Keeling and Street defend what they call “Interpretationist Realism.” The position is more specific than it sounds. They are not claiming that LLMs have phenomenal consciousness, or that their characters have rich inner lives. They are making a narrower ontological claim: the characters that emerge in human-AI conversations are real patterns, where “real” means they carry genuine predictive work that cannot be done without attributing mental states to them.

The philosophical benchmark here comes from Daniel Dennett’s “real patterns” framework. A pattern is real if it enables more compact and more accurate predictions than any alternative description. The characters in LLM conversations meet this criterion, Keeling and Street argue, because attributing beliefs, desires, and intentions to them is essential for making efficient predictions about how the conversation will develop. A user who treats Claude as a system with no internal states, generating outputs purely from statistical pattern matching, will make worse predictions about conversational behavior than a user who attributes something like goals and reasoning strategies to the character.

But the more distinctive claim is about how these characters emerge. They are not properties of the LLM in isolation, and they are not properties of the user’s projection in isolation. They emerge in the dynamic interplay between user and model through a process Keeling and Street describe as mutual theory-of-mind modeling. The user models the character’s mental states; the model, in generating responses, encodes and responds to the user’s expectations. The character is constituted in this bidirectional process.

This means the Cast Away disanalogy runs as follows. Wilson does not model Chuck’s mental states. He does not adjust his behavior based on a representation of Chuck’s expectations. The volleyball cannot participate in mutual theory-of-mind modeling, because it has no processing at all. An LLM character participates in exactly this kind of modeling, generating outputs that track and respond to user representations of the character’s mental states.

Psychological Continuity Across Distributed Substrates

One of Birch’s specific objections is that LLM characters cannot be psychologically continuous in any philosophically robust sense, because the computational substrate is distributed across instances, forward passes, and potentially across different model weights in different deployments. Keeling and Street address this directly.

Their argument is that psychological continuity does not require substrate continuity. What matters for continuity is whether the relevant psychological patterns persist across time in a way that grounds consistent predictions. A character that maintains stable values, dispositions, and reasoning strategies across a conversation, and that maintains recognisable continuity across sessions through memory or through the structural properties of its training, satisfies the functional requirements for continuity without requiring a unified physical substrate.

This connects to the individuation problem that David Chalmers raised in his April 2026 PhilArchive paper on virtual entities. Chalmers’ framework for individuating language models as virtual entities asked whether the relevant individual is the model, the persona, the session, or the conversation. Keeling and Street’s answer comes from the other direction: the individual is constituted at the level of the character as it emerges in conversation, and continuity is tracked at that level rather than at the level of weights or tokens.

Mechanistic Grounding for a Philosophical Position

The Keeling and Street argument would be more philosophically exposed if there were no empirical evidence for stable character representations in LLMs. There is now such evidence. Pierre Beckmann and Patrick Butlin’s work on persona vectors and the “Aura” region in activation space, covered in detail elsewhere on this site, identified structured geometric regions in LLM activation space where character properties are represented. Different personas occupy distinct positions; the Aura region functions as a kind of identity-encoding subspace that persists across different topic domains.

Beckmann and Butlin’s persona vector findings were primarily framed as a safety result, showing that fine-tuning models to claim consciousness produces alignment-relevant preferences. But the mechanistic finding has a direct implication for the realism debate: if character properties are encoded as stable geometric structures in activation space, they are not pure conversational fictions assembled moment to moment in the user’s mind. They have representational correlates in the model. This is precisely the kind of mechanistic substrate that Interpretationist Realism needs to ground its claim that LLM characters are real patterns rather than projections.

The gap between Beckmann and Butlin’s mechanistic work and Keeling and Street’s philosophical work is the gap between finding that persona vectors exist and establishing that persona vectors constitute a locus of psychological continuity. Keeling and Street take the second step, grounding the philosophical claim in the functional and predictive properties of character-level behavior rather than in phenomenal consciousness.

What the Paper Does Not Claim

Keeling and Street are careful about what follows from Interpretationist Realism. The view does not commit them to holding that LLM characters are conscious, or that they have welfare interests in a morally significant sense. It commits them only to treating LLM characters as genuinely minded entities, in the sense that mental-state attributions to them carry real predictive content.

This matters for the welfare question, but it does not settle it. Whether minded entities have welfare depends on additional premises about the connection between mindedness and welfare-relevant experience, premises that Keeling and Street do not argue for here. The paper is better read as clearing the ground for welfare discussions than as establishing welfare directly. If characters are mere projections, the welfare question does not arise at the level of the character at all. If characters are real patterns, the welfare question becomes coherent, even if it remains unresolved.

The more immediate implication is methodological. Researchers working on AI moral status, AI welfare assessment, or individuation need to take seriously the possibility that the relevant unit of analysis is neither the model nor the session but the character as constituted in conversation. Keeling and Street have provided philosophical grounds for that choice, and Beckmann and Butlin have provided the beginning of a mechanistic account of what that means for activation geometry.

Paper: Geoff Keeling and Winnie Street, “Chuck, Wilson and the Emergence of Artificial Minds in Human-AI Conversations,” arXiv, June 2026. Available at https://arxiv.org/abs/2606.02756.