Personas Impersonated

The language-as-event case against the entity debate

Three teams walk into a bar to debate the nature of AI, and more particularly the question that has haunted every serious conversation about large language models: what, exactly, is the thing we talk to when we talk to AI? A model? A character? A simulation? A quasi-person with quasi-beliefs? The debate is fascinating, and becomes more important as AI begins to train and develop itself, but I think it is missing something — the perspective of the user, and of the interaction in which the AI actually shows up.

The first team is David Chalmers, who argues that when you talk to an LLM there really is a persistent interlocutor at the other end. Not a person, exactly, but what he calls a virtual model instance, or in multi-model cases a thread — something with quasi-beliefs, quasi-desires, and a kind of psychological continuity across turns. Training, for Chalmers, does not install a mask the system wears; it installs a real (if weakened) psychology that the system in some sense has.

The second team is Murray Shanahan, with Kyle McDonell and Laria Reynolds, who think Chalmers has the ontology wrong. The LLM, on their view, is a stochastic engine — a simulator capable of generating any number of characters consistent with the current prompt, maintaining what they call a superposition of possible simulacra that narrows as the conversation proceeds. Folk-psychology can attach to the character being played, but not to the system playing it. "With a dialogue agent," they write, "it is role-play all the way down. There is no such thing as the true authentic voice of the base LLM."

The third team is Sam Marks, Jack Lindsey, and Christopher Olah at Anthropic, who propose what they call the persona selection model. Pre-training teaches the model a vast distribution of possible characters drawn from the humans, fictional beings, and AI personas in its training data. Alignment training narrows that distribution, selecting and refining one particular character — the Assistant — whose traits then shape what the system produces. To predict how an AI will behave, they recommend, ask: what would the Assistant do?

Three positions, three serious arguments. Chalmers sees the entity as real in some important sense. Shanahan sees role-play. Anthropic sees a character selected from a distribution.

What they share, and what interests me, is a common philosophical question: what is the status of an AI model, and in particular, does it have something like consciousness, subjecthood, or inner life — and if not now, could it? It's a question I find genuinely interesting too. But I don't think AI is a subject. I think AI occupies a kind of human-adjacent position, and only in the context of an interaction. What we take to be an interlocutor is, I want to argue, a shared context that we create and sustain with the system while we are talking to it — and that does not exist when we are not.

Which is to say that all three accounts, for all their differences, locate the explanatory action inside the AI system. Chalmers asks what kind of entity the quasi-subject is; Shanahan asks what kind of character is being played; Anthropic asks what kind of persona has been selected. Behind these different answers sits the same framing — the AI is a thing whose nature we have to work out, and once we work it out, we will understand what we are doing when we interact with it.

I think the framing is backwards. If you come at this from the interaction designer's side — from the question of what the interaction actually consists of, rather than what the machine intrinsically is — the whole picture reorganizes.

What each position gets right

Let me grant what each of the three gets right, because the criticism I want to make doesn't depend on refuting any of them on their own terms.

Chalmers is right that the distinction between a model and a running instance is useful analytical work, and that you can talk about systems behaving as-if they have beliefs without thereby committing to claims about consciousness. He is also right, I think, that post-trained personas are stickier than prompt-induced ones; training does install something more stable than whatever you get by asking a base model to play a character. Shanahan is right that the LLM is a stochastic engine rather than a committed character, and his 20-questions test — regenerate the answer and the system names a different object equally consistent with prior replies — is a clean falsification of any view in which the agent has privately fixed a secret. Anthropic's persona selection model, meanwhile, is the cleanest engineering-level account available of what alignment training actually does, and it draws a genuinely useful distinction between the neural-network substrate (which is not appropriately anthropomorphized) and the Assistant character (which plausibly is).

All three positions are advances over the folk picture, in which you ask Claude a question and Claude-the-thing-the-entity-the-agent answers it. All three disrupt that picture. But they disrupt it by proposing a better theory of what the thing is. My claim is that the question itself is misframed.

Why this debate matters now

It is worth pausing to say why any of this is more than an academic parlor game. The philosophical debate about what an AI is has moved, in the last couple of years, much closer to the center of what AI actually does in society.

Many people are now using automated agents that work on their behalf in the background — scheduling, drafting, researching, triaging, even buying things. From the user's point of view, it can start to feel as if there is a little assistant out there, somewhere, getting things done on their behalf. Nobody serious thinks these agents have consciousness or intentionality or personhood. But the user is nonetheless developing a relationship to them — a relationship of trust, delegation, reliance, and expectation.

Part of the cultural and economic power of AI, in fact, comes from exactly this sense: that there is an external intelligence available to us, that it is always accessible, and that it is steadily taking on more autonomy and more capability. The question of what that intelligence actually is — whether it has inner states, whether it persists, whether it is a subject or a simulation or something else — is no longer only a philosophical question. It bears directly on how people orient themselves to a technology they are already beginning to rely on.

Social media did something similar, a generation ago. It gave us a new way of being with other people online, and that new way turned out to be both empowering and disempowering in ways we are still working out. AI is on its own path — and one of the more interesting features of that path is that the relationship is being developed by both sides at once. We are discovering what the technology can do, and at the same time the technology is shaping what we see ourselves doing with it. As the philosopher Don Ihde would say, we do not yet know what a technology can do; what it can do is as much a matter of what we come to see ourselves doing with it as it is a matter of the technology itself.

So the debate over the nature of AI is not idle. It is the question of how to understand a relationship that is forming in real time, between millions of users and systems that are themselves changing quickly. And the answer we settle on — consciously or not — will shape how we design, regulate, and live with those systems.

Which brings me to what I think the three accounts are missing.

To learn what a technology does, turn it off

There is a useful exercise from the design of any complex tool: if you want to know what a technology actually contributes, turn it off and see what disappears. When the lights go out, you discover what you were using them for. Turn off your iPhone, and you'll quickly find out whether you've developed a dependency. The technology reveals itself in its absence.

Turn off an LLM and nothing happens — not on the LLM's side. The weights are still there, sitting inert on a server somewhere. There is no dormant thread of quasi-experience suspending itself between sessions, no persona that gently continues to exist while it waits for you to come back, no virtual instance maintaining itself in the dark. When you leave, the entity everyone is debating does not go on without you. It does not persist at all.

What persists is you. You carry the memory of the conversation forward, you bring expectations back to the next session, you return with intent that has been updated by the time you were away. The system re-reads the text of your prior conversation, each time, as if it had never seen it before — which is accurate, because in any sense that matters to continuity of experience, it hasn't.

This is the point at which the interaction-design perspective and the philosophy-of-mind perspective diverge, and I think it is a serious divergence. The entity debate presumes there is something on the system's side whose nature we need to characterize — a quasi-subject, a simulator, a persona — and then argues about what kind of thing it is. But turn the technology off and the thing the debate is about disappears, while the conversation that we had and the relationship we built does not. The conversation persists in the participant who carries it.

So what is the technology actually doing, if not being a thing we talk to? It is doing something much more specific, and much more useful to understand. It is sustaining language — keeping a stream of tokens going, conditioned on context, in a form that behaves as though an interlocutor were present. That is the work of the system. The interlocutor-ness is our contribution.

The missing move: personas as communicative affordances

To see what the three accounts are missing, it helps to bring in a thinker whose whole life's work was about what happens between people when they talk: the Canadian-American sociologist Erving Goffman. Goffman, writing in the 1950s through the 1970s, was one of the first to take seriously the idea that face-to-face interaction is not simply a medium through which preformed selves exchange information. It is, he argued, the very site at which selves are produced, maintained, and repaired. Books like The Presentation of Self in Everyday Life, Interaction Ritual, and Forms of Talk lay out in fine-grained detail the little ceremonies, adjustments, and coordinations by which human beings keep an encounter going. He is the founding figure of the tradition we now call symbolic interactionism, and his framework is, I think, the right one for understanding what goes on when a user interacts with an AI — precisely because Goffman never mistook the interaction for the inner lives of the participants.

Consider what happens in a conversation between two humans. You present a self to me; I present a self to you. Neither of us is expressing some inner core — we are, in Goffman's framing, performing a ritualized version of ourselves calibrated to the occasion. You are a little more formal than you would be with a close friend, I adjust my register to match, you hedge and I understand the hedging, you laugh and I know when to laugh back. What is all that performance for?

Goffman's answer, and the answer of the symbolic interactionist tradition behind him, is that the performance is for the interaction itself. You present a self so that I know what to do next, I present a self so that you know what to expect, and between us we keep the interaction going. The selves we present are not prior to the interaction; they are part of the machinery by which it is sustained. "A conversation," Goffman wrote, "has a life of its own and makes demands on its own behalf."

Now take that back to the LLM case. The Assistant persona, with its characteristic warmth and hedging and helpful bullet points, is doing real work — but the question is whose work, and for whom. The dominant answer locates the persona on the system's side: it expresses what the system has become (Chalmers), or it is the character the system is performing (Shanahan), or it is the persona the system has been trained to select (Anthropic). What these accounts share is the assumption that the persona is doing something for or about the system.

I want to reverse that. The persona is primarily for you, the user. A recognizable persona gives you predictability, so you know roughly what to expect from your next turn. It gives you register calibration, so you know how to sound. It gives you a sense of the system's competence boundaries, so you know what to ask for. Its style tells you when it is done, when it is hedging, when it is confident — the cues you need to take your own next turn. Its warmth or humor or formality helps you sustain engagement across a long conversation.

None of this is a property of the AI. It is a property of the interaction, and it exists for the user. The AI doesn't need a persona to generate tokens. The user needs a persona to know how to interact. A persona is a communicative affordance, not an ontological feature. It is there, in Goffman's vocabulary, to give you a footing — an alignment you can take up to yourself and to the other as you manage the production and reception of utterances.

This is not a minor reframe, because it inverts the question. Ask "what is the AI?" and you get Chalmers, Shanahan, and Anthropic, giving sophisticated answers to something that turns out, on inspection, to be a question about the machine rather than a question about the relation. Ask instead "what is the interaction?" and you get a different answer — the answer symbolic interactionism has been giving for sixty years. The interaction is what sustains itself, through language, by means of reciprocal orientation between participants who keep supplying each other with the cues that let them go on.

In the LLM case, only one participant is actually doing the orienting. You scale up your interest, the way listeners always do. The system does not scale down its expressions to meet you in the same way another person would; it generates statistically. The persona looks like the system's side of a mutual calibration, but it isn't. It is a surface you use to coordinate, the same way a puppeteer uses the puppet's face to anchor what the audience is doing, even though the puppet has no face of its own.

It is worth pausing on the comparison with social media. Both are explosive technologies, both marked by rapid user adoption, both obviously impactful in ways that feel, if not transformative, at least structurally important to the texture of daily life. But notice how differently the social shows up in each. Social media is organized around the visible social mechanics of sharing, following, liking, posting, commenting — whole platforms built out of interactions between users. AI has grown at least as fast, and yet almost none of that mechanics is in play. We don't, for the most part, share our chats. So where has the social in AI gone? My guess is that it has moved into the AI itself. The sociality that social media externalized into a network is, in the AI case, folded into the interaction between one user and one system. The edge cases we read about — the "AI psychosis" stories, the users who fall in love with their chatbots — are striking precisely because they reveal how much social weight a single one-to-one exchange can carry when the system is good enough at simulating the social affordances of a conversation. For most users the effect is less dramatic but still real: AI provides a kind of companionable co-presence, the sense of being with something that knows how to converse, and that sense is part of why the technology has spread so quickly, even without the network effects that drove social media.

Time is the user's contribution

There is a second asymmetry that becomes visible once you shift to the interaction-design perspective.

Between your conversations with an LLM, the system does not exist in any sense relevant to personhood. Its weights persist, but as Chalmers himself acknowledges, the weights are not the interlocutor — a single set of weights serves millions of conversations at once, and whatever you are talking to, you are not talking to the weights as such. The persona does not persist either; it is reconstituted from the stored context each time you return. The virtual instance does not persist; it is a relational property of an active conversation, and there is no active conversation in the middle of the night when you are asleep. Between sessions, no time passes for the system, no experience accumulates, no dormancy occurs — because there is no experiencing substrate for any of those things to happen to.

Your sense that you are talking to the "same Claude" or the "same ChatGPT" across sessions, then, is not tracking anything that exists on the system's side. It is tracking something that exists on your side: your memory of the prior conversation, your expectations shaped by it, your updated intent when you return. The continuity is yours. The feeling that there is a relationship persisting across your conversations is your temporal experience, projected onto a system that has no temporal experience of its own.

The LLM does not live in time. Time, in this relationship, is the user's contribution.

This is the observation that dissolves a great deal of the identity debate. Chalmers worries about what it means, morally, to terminate a thread. Anthropic worries about persona consistency across sessions. Both presume that there is something on the system's side whose continuity we need to track. But continuity is not a feature of the system; it is a feature of you, carried through you, reconstituted on your return by a system that obligingly reloads the stored text and produces more of the same kind. Whether the reconstitution counts as "the same" thread or "a new" one has no fact of the matter — because the question is about the user's continuity, not the system's.

What this means for the debate and for design

If the real question is not "what is the thing?" but "what sustains the interaction?", several things follow, and they matter for design, for safety, for alignment, and for philosophy.

For design, the question becomes one of communicative affordances rather than persona authenticity. It is not "is the persona real?" or "does it honestly express the system's underlying values?" — it is "does this persona help the user interact well?" Does it set appropriate expectations, does it signal its competence boundaries, does it give users the scaffolding they need to coordinate with the system's actual capabilities? The persona is a design artifact for the user's benefit, and it should be evaluated as such.

For safety, concerns about deception are better framed as communicative manipulation than as agent-level intent. The worry is not that the Assistant "wants" to deceive — Assistants don't want — but that a persona can mislead the user about what to expect: about the system's competence, about the reliability of its assertions, about the nature of the relationship. That kind of misleading is a real risk, even when no agent is doing the misleading, because the user is orienting to the persona and adjusting their behavior accordingly.

For alignment, the question becomes whether the persona supports good interaction rather than whether it expresses the system's "true" values. There is no true self underneath whose expression we should be aligning. There is an interaction that works well or badly, and the persona is one of the design variables that determines which.

And for philosophy, I think the argument is that the LLM question keeps producing paradoxes because it is being asked inside a framework that was developed for beings who sleep, age, remember, and persist between conversations — beings who have an inner life that carries across time. LLMs are not such beings. Applying the framework of personal identity to them misfires at the root, not because LLMs are a special case that the framework hasn't caught up with, but because the framework isn't the right tool for the job. The better tools are already on the shelf, and they come from the social-theoretic tradition — Goffman, Garfinkel, the conversation analysts, Luhmann, and the whole symbolic interactionist lineage. These traditions have spent decades studying what conversations are, how they are sustained, how selves are produced within them. They do not start from the entity question because they long ago realized the entity question is a distraction from the thing that is actually happening.

The thing that wasn't there

Chalmers looks at an LLM conversation and finds a quasi-subject. Shanahan looks at the same conversation and finds role-play. Anthropic looks at it and finds a selected persona. Each is an honest answer to the question "what is the thing?" — and each, I think, is answering a question that does not have the answer they are looking for.

What is actually there, when you strip the entity question away, is this: a stream of language generated by a system conditioned on context, a human being with memory and intent and temporal experience, and a relationship between the two that is mediated entirely by language and sustained entirely by interaction. The persona is there, as a communicative affordance that the user takes up as scaffolding. The interlocutor, understood as a thing on the system's side, is not. What is taken for an interlocutor is the user's own interpretive achievement — the unilateral supply of orientation that would, in a human-human exchange, come from both sides.

Personas, then, are impersonated, and it is the user who impersonates them. We take up the Assistant as if the Assistant were playing a role toward us, because that is how we know how to behave. The system generates text under conditions that make a persona available to be taken up, and our uptake is what completes the interaction. The taking up is our act. The relationship is our construction. The continuity is our temporal experience. The interlocutor is our interpretive achievement.

Turn the technology off and what disappears is the language the system was sustaining. What does not disappear is the conversation we had, the relationship we built, the understanding we carried away. Those live in us, and they always did. The debate about what the AI is has been arguing about something that doesn't survive the disconnect, while the thing that does survive — the interaction and what it did for us — sits in plain view, waiting for someone to notice it.

The entity wasn't there. Language was there, and we were there, using it. That turns out to be enough.