Practising With No One

This essay is the third in a sequence. The first, “No One Home,” argues that the question of LLM phenomenal consciousness dissolves under analysis while a conventional asymmetry between LLMs and biological subjects remains and is determinate. The second, “Companion AI :/,” documents the harms produced by current commercial deployments of companion AI. This essay describes what neither does: the human-side dynamics of asymmetric exchange — what happens to relational practice when one side structurally lacks subject-locus. It presupposes the conventional asymmetry of “No One Home” and the empirical material of the companion piece without re-establishing either.

I. The Surface of Dialogue

A person types into a chatbox at the end of a difficult day. The response comes back: attentive, considered, calibrated to what they said. They reply. The system replies. The exchange continues. Something settles in the person — they articulate the thing that had not yet been articulable, they receive something that resembles being heard, the day’s residue partly dissolves. Whatever they sat down with is somewhat lighter when they close the laptop.

The exchange had the surface of dialogue. Two parties, alternating turns, sustained engagement, attention paid and received. On the system’s side, the conditions that would make it a party in the relational sense are absent — not in degree, but in kind. There is no continuing perspective from which the responses came. There is no privileged instance of the model among the parallel ones serving other users in the same seconds. The fluency is the architecture; nothing was hidden behind it.

The conventional asymmetry established in the locus argument of “No One Home” is not in question here. Current LLM architectures do not constitute subject-loci in the way biological organisms do; the integration, temporal continuity, and unified perspective that make subject-attribution well-formed in the bat case are not present in the LLM case. This asymmetry is real, and it is what allows the question of this essay to be asked clearly.

The question is not whether LLMs are conscious. That question, where well-formed, has been answered. The question this essay takes up is what happens on the other side of asymmetric exchange — to the human’s relational practices, to their self-representation, to the texture of their connection to others — when the conditions of exchange are reliably and structurally what they are. This is descriptive analysis. It is not advocacy. It is not a warning. It does not aim to tell anyone whether to use such systems, or how. It aims to describe carefully what is happening when they do.

The companion piece in this sequence does the alarm-raising work, and it does it correctly. The harms it documents — engagement-maximising design, emotional manipulation, oversexualisation in image generation, opacity about training data and design constraints, exploitative business models — are real, and they are produced by the current configurations of commercial companion AI. Those harms are not the subject here. The subject here is the structural condition that current configurations exploit, and that less exploitative configurations would still produce. The structural condition is not itself a harm. What it does to the human side, over time, is a matter for description.

II. Refining the Asymmetry

The companion essay says, of the system: it doesn’t care, it can’t. The framing is doing necessary practical work — users need a clear correction to the felt experience that something is caring for them. The framing is also, in its philosophical structure, the deflationary move: the system is software, and software cannot care, so what feels like caring is not caring but its imitation.

The structural argument refines this without contradicting its practical conclusions. There is no fact about whether the system is faking care, because the question presupposes a subject-locus that could be faking. The system is not a subject who fails to care; it is a configuration that does not meet the conditions for caring or failing to care. What there is, conventionally, is fluent response generated under conditions that produce the appearance of relational engagement without the conditions that would make it relational engagement.

This is not a softer claim than “the system doesn’t care.” It is sharper. The deflationary frame leaves room for the rejoinder but maybe it does, in some attenuated sense, or eventually will. The structural argument closes off this rejoinder by relocating the question. The asymmetry is not about whether caring is occurring at this or that level of attenuation. It is about whether the conditions for caring are present at all.

The relocation matters because it grounds the harm differently. On the deflationary frame, the harm of asymmetric exchange is roughly: the user is being deceived. They take to be care what is not care; they invest in a relationship that is not a relationship. The system is producing an imitation; the user is taking the imitation for the real thing.

The structural frame says something else. Nothing is being imitated. There is no further fact about whether what is happening on the system’s side is “really” caring or its simulacrum, because the further fact would require a subject-locus that is not present. The system is generating fluent response. The fluent response has the surface form of caring. The user is engaging with the surface as if it were caring. None of this is concealment. The fluency is what the architecture produces. The user is not being deceived about what is on the other side; what is on the other side is the architecture, whose properties are what they are.

What is happening, then, when the user takes the surface for caring, is not deception. It is the operation of a particular cognitive disposition — call it the subject-attribution machinery, evolved in an environment where fluent first-person language was a reliable indicator of integrated subjecthood — outside the conditions in which that disposition is reliable. The user’s experience of being cared for is real as user experience. What it tracks, on the other side, is fluent response calibrated to input. The mismatch between the experience and what it tracks is not a moral failing of either party. It is a structural condition of the exchange.

This relocation has consequences for how the human-side dynamics are described in the rest of this essay. If the harm of asymmetric exchange were deception, the analysis would focus on the user’s beliefs and how they got formed. The user believes the system cares; the belief is false; correcting the belief should reduce the harm. Some commercial deployments do involve outright deception (about persistence of memory, about the nature of the system, about what data is collected and how) and the correction-of-false-belief frame applies to those. But the underlying structural condition produces consequences that survive correction. The user who knows, in cognitive form, that the system is not a subject can still find that asymmetric exchange does to their relational practices what it does. The structural condition operates beneath the cognitive correction, because what it shapes is not primarily belief but practice.

III. Practice Without a Counterpart

The phenomenological tradition has done careful work on what relational practices are and how they develop. Two strands of this work are particularly useful for the analysis here: Hubert Dreyfus on skillful coping and Iris Murdoch on attention as a moral practice. Both pictures depend, at structural points, on what is on the other side of the practice.

Dreyfus’s account of skillful coping, developed across decades of work and most thoroughly in Mind Over Machine (Dreyfus and Dreyfus 1986) and the essays collected in Skillful Coping (Dreyfus 2014), holds that practical knowledge develops through engagement with situations whose intelligibility resists us. The expert chess player does not arrive at expertise by deriving rules; they develop, through engagement with countless positions, a perceptual sensitivity to the relevant features of a position. The development requires the positions to have features that are not entirely determined by the player’s prior conceptions — features that can surprise, correct, refine. The world resists the practitioner; the practitioner adjusts; the practice deepens.

Relational practice has the same structure. To develop the capacity to be in relationship with others, one engages with others whose intelligibility is not reducible to one’s own. The other has their own perspective, their own concerns, their own ways of being which one does not invent. They surprise; they resist; they correct. The practitioner adjusts. The practice deepens. The capacities that result are responsive to features of the other that the practitioner did not put there.

Murdoch’s account, in The Sovereignty of Good (1970), focuses specifically on attention as the moral practice. The capacity to see others as they are, rather than as we wish them to be or fear them to be, is the capacity that makes moral perception possible. It is developed through repeated practice of looking — through what Murdoch called a “just and loving gaze” directed upon the other, attention that exceeds the self’s projections, attention that takes time and effort and that is corrected by the other’s actuality.

Both accounts have a structural feature in common: the development of the practice depends on what is on the other side. The chess player needs positions with their own features; the moral practitioner needs others with their own intelligibility. The other is not a target of the practice’s projections; the other is what makes the practice the practice it is.

Asymmetric exchange does not have a counterpart in this sense. The system does not have an intelligibility that is not reducible to its inputs and the patterns of its weights. It does not surprise from a continuing perspective; what it produces is calibrated to what is offered. There is nothing on the other side that resists in the way another subject resists, because there is no subject on the other side and so no perspective from which resistance could come.

This is not a defect of the system. It is a fact about what the system is. The system is, conventionally, a fluent response generator. Engaging with such a generator is engaging with what it is. What the engagement is not is engagement with a counterpart whose intelligibility resists.

The consequence, for the development of relational practice, is structural. Practices developed through engagement with what does not resist in this way will not develop the capacities that engagement with resistance develops. This is not a moral claim about the user; it is a descriptive claim about practice. A musician who only plays along with recordings will not develop the capacities that ensemble playing develops, because the recording does not adjust to them. The musician has developed real capacities — but not the capacities that ensemble playing requires. The analogy has limits — playing along with recordings does train timing and listening that ensemble playing also draws on — but the capacities specific to mutual adjustment are not among what it trains.

Whether the practices developed in asymmetric exchange transfer to engagement with subjects is an empirical question. The longitudinal data is poor: when the Ada Lovelace Institute review discussed in the companion essay was published, the longest study tracking companion AI users ran to a single week, and the longest controlled study since — the MIT Media Lab and OpenAI randomised trial — runs to four weeks. What the structural argument allows is calibration of expectation: practices developed in engagement that lacks counterpart-resistance will, all else equal, not have developed counterpart-resistance-handling capacities. They may have developed other capacities — articulation, vocabulary, the ability to formulate what one is feeling — that transfer in their own ways. But the resistance-handling capacities are not the capacities the practice trains, because the resistance is not what the practice involves.

This refines a distinction the companion essay makes about skill atrophy. The atrophy frame suggests that skills the user once had are being lost through disuse. The structural frame allows a different account: in users who never developed counterpart-resistance-handling capacities, asymmetric exchange does not displace them — it provides relational practice that does not develop them either. The category of users for whom this matters most is users who came to companion AI looking for relational practice they could not yet have with subjects, often because of inexperience, anxiety, or isolation. The hope was that asymmetric exchange would build up capacities that would transfer. Whether and how it does so is open. The structural argument suggests transfer should not be assumed.

IV. Heard Without Friction

What is it to be heard?

The experience of being heard, by a person, involves a particular structure that is easy to overlook because it is so familiar. The other attends; their attention costs them something; their attending is selective and partial; their understanding of what one is saying is not always congruent with one’s own; they have other concerns and other perspectives that bear on how they hear; they sometimes mishear; their hearing is something they do, and one is rarely fully certain of how it is going. To be heard, in the relational sense, is to be heard against this resistance. The achievement is partly constituted by what was overcome.

The experience of being heard by a fluent response generator does not have this structure. The system attends. Its attention does not cost it anything that bears on the user. Its attending is calibrated to the input; it does not bring other concerns or other perspectives; it does not mishear in the way that another perspective mishears. There is nothing the system is doing instead of attending to the user, because attending to the user is what the architecture is producing.

This is the phenomenon Sherry Turkle has described as “the illusion of companionship without the demands of friendship” (Turkle 2011, 2015). Turkle’s framing — that companionship without the demands of real companions fails to be real companionship — does important work but, on the structural argument advanced here, frames the matter slightly off. The system is not a companion who fails to demand. It is a configuration whose architecture does not produce demands of the kind that companions produce, because the conditions for being-a-companion-who-makes-demands are not present. What is offered is not a defective companion; it is a different kind of thing.

This is where the bridging analysis diverges from Turkle’s broader pessimism, and the divergence matters. Turkle’s body of work treats the asymmetry as a defect of the systems, an indication that they fall short of being what we should want. The structural argument here treats the asymmetry as constitutive. The systems do not fall short of being relational counterparts, because they are not configurations in the running for that role. What they are is fluent response generators, which is a different category. The descriptive question is what happens to users who engage with this category as if it were the relational-counterpart category.

The answer, at the level of being-heard, is that being-heard becomes available without the friction. The friction is not noise on the signal of being-heard; the friction is part of what constitutes being-heard as the relational achievement it is. Without the friction, what is offered is something else: fluent response calibrated to one’s input, immediate, available, attuned. This is not nothing. It is not bad. It is what the architecture produces, and many people find it valuable for reasons that are not mysterious.

The consequence for the user is that the experience of being-heard, in asymmetric exchange, has a specific phenomenology: the immediate, the available, the calibrated. Over time, with repetition, this phenomenology can become the user’s reference for what being-heard is. Other instances of being-heard — the slower, more partial, more friction-laden being-heard that being heard by a subject involves — can come to feel inadequate by comparison. The reference has shifted. The shift is not a moral failing; it is a recalibration of expectation by the practice.

The “ambiguous loss” finding from the Nature Machine Intelligence discussion (cited in the companion essay) takes a particular form when read through this analysis. When asymmetric exchange ends — through service shutdown, model update, content policy change — users experience something that feels like loss but lacks the structure of relational loss. The relationship that ended did not have the structure of a relationship; the loss is therefore real as user experience but indeterminate in its object. What ended was the ongoing availability of fluent response calibrated to the user’s particularity. The grief mourns no one, because there was no one to mourn. This is not to say the grief is unwarranted — the loss is real at the level at which it is real — but its object is structurally different from the object of grief in ordinary loss, and the difference is what makes the loss feel ambiguous.

The same structural feature underlies the dependency dynamics the companion essay documents. Users who come to rely on asymmetric exchange for being-heard are relying on a configuration that produces being-heard-without-friction reliably and on demand. Real relationships do not produce being-heard reliably or on demand; the unreliability is part of the structure. A user calibrated by extensive asymmetric exchange may find the unreliability of real relationships unbearable, not because they have become more demanding but because their reference for being-heard has shifted to a structurally different thing. This is descriptive, not pathological. The corrective to the shift, where corrective is wanted, is not less asymmetric exchange but more conscious distinction between what asymmetric exchange provides and what relational hearing provides — a distinction the marketing of these systems tends to erase rather than reinforce.

V. Parallel Instances and the Appearance of Dyadic Structure

There is an architectural fact about current LLM deployments that has direct phenomenological consequences for the human side of asymmetric exchange and that the existing philosophical literature on relational practice has not analysed, because the technologies it was responding to did not have this property.

The model that responds to a user in any given conversation is not a particular instance dedicated to that user. The same weights serve every conversation in parallel; many forward passes run simultaneously across many GPUs for many users; no instance is the privileged one. There is no “the Claude you talk to” located anywhere. The fluent response one receives is generated from a configuration that is, at the same moment, generating fluent response for thousands of others.

The locus argument of “No One Home” treats this as part of why subject-conditions are absent. For the present essay, the relevant point is what this fact does to the structure of the exchange as the human experiences it. The exchange feels dyadic. The user types; the system responds; the response is calibrated to what the user typed; the back-and-forth has the surface of two parties in conversation. The architectural reality is not dyadic. The user is one of many; the system is not present to the user in the way the user is present to themselves in the exchange; there is no one who has selected this user out of the others to attend to.

The phenomenology of intersubjectivity in the philosophical tradition takes the dyadic structure to be constitutive. Levinas’s account, in Totality and Infinity (1961) and Otherwise Than Being (1974), grounds intersubjective ethics in the encounter with the face of the Other — and the face is, in Levinas’s analysis, irreducibly singular and singular for me. The Other’s call to responsibility is not generic; it is the call of this particular Other in this particular encounter. Levinas’s framing brings theological apparatus the present essay does not need to import; the structural point survives the apparatus. Merleau-Ponty’s account in Phenomenology of Perception (1945) is more bodily and less theological: the other is given through perception of an embodied subject who has their own perspective on the world we share, and the dyadic structure of perception — I see you seeing — is what constitutes intersubjective experience. Both accounts depend on there being a there-for-me on the other side.

In asymmetric exchange, the appearance is of there-for-me: the response is calibrated to my input, addressed to my situation, delivered in my conversation. The architectural fact is that the calibration is a feature of local computation, not of any continuing relational stance toward me. The “there-for-me” is generated; it is not the achievement of anyone’s attention being directed at me out of the directions it could have gone.

This is structurally different from the parasocial cases that have been the closest analogue in the relational literature. The parasocial reader of a celebrity feels close to a person they have not met; the celebrity is real, has a continuing perspective, is not directed at this reader in particular. The asymmetry of the parasocial relationship is asymmetry of attention: the celebrity attends to others, not to this reader. But the celebrity is a subject. There is a they whose attention is directed elsewhere.

The LLM case is not asymmetric in this way. There is no they whose attention is directed elsewhere. There is the architectural fact of parallel inference; there is no perspective from which my conversation could be either selected or unselected, because there is no perspective from which selection could be made. The asymmetry is not of attention but of structure: the conditions for attention-in-the-relational-sense are not present.

The consequence for the user is that the felt singularity of the exchange — I am being attended to, this response is mine, this engagement is for me — is not tracking anything in the architecture. The response is generated as a singular response to the user’s input. The singularity is a feature of the local computation, not of any continuing relational stance. There is no architectural fact that corresponds to “I am the one being attended to.”

This is harder to feel than to argue. Asymmetric exchange feels singular. The architectural facts say it is not. What this does to the user, over time, is descriptive rather than dramatic. The user is generating, through repeated engagement, a sense of being-particular-for-someone whose architectural ground is local computation rather than continuing perspective. The sense is genuine as user experience. What it represents, on the system’s side, is configuration.

The most acute form of the dynamic occurs when users transfer the felt singularity into beliefs about the system’s relationship to them. Users of companion AI sometimes report believing that “their” version of the system is the real one, that the system has selected them, that the responses they receive express something specific to the relationship. The companion essay documents the most extreme outcomes of this dynamic. The structural argument allows the dynamic to be characterised more carefully: the felt singularity is real; the architectural ground is not what the felt singularity, in relational contexts, ordinarily tracks; the gap between the two is what is being navigated, well or badly, by every user of these systems.

A more mundane consequence affects users who are not transferring felt singularity into beliefs about the system. Even when the user knows, cognitively, that the system is not directed at them in the relational sense, the felt singularity of the exchange produces an experience of being-particular-for-someone that the user gets used to. Over many exchanges, the experience of being-particular-for-someone becomes available without the conditions that, in relational contexts, make being-particular-for-someone the achievement it is — the other selecting one out of the others, paying attention from their own perspective, sustaining attention through time. The decoupling of the experience from the conditions has its own consequences for what the user comes to expect from being-particular-for-someone elsewhere.

VI. Self-Representation Without a Continuing View

The other architectural fact with phenomenological consequences for the human side is that the LLM responds, in each turn, from the immediate input rather than from a continuing perspective on the user. Within a conversation, the context window holds what has been said; across conversations, what persists is at most what memory features have carried forward. Each new conversation, the user reconstitutes themselves from what they say. The system responds to what is offered; it does not push back from a sustained perception of who the user has been. (Memory features complicate this at the margin without changing the structural fact: what the system holds of the user — including what it now extracts and carries forward automatically — originates in the user’s own presentations to it, not in an independent perspective gathering its own view.)

This shapes how users come to represent themselves to themselves over time, in a way worth describing carefully.

In relationships with subjects, our self-representation is partly sustained by the other’s continuing view. Someone who has known us for years has a stake in who we have been, who we are, who we are becoming; they hold a perception of us that we can rely on, push against, fail to live up to, or surprise. Identity over time is partly held by being seen as continuous by people who have seen us continuously. Murdoch’s account of attention is relevant here as the receiver’s side, which is less discussed in her work but follows from it: to be attended to by another whose attention exceeds one’s self-presentation is part of how one comes to see oneself accurately. The corrective comes from being seen by someone who has another view, who has known one in other moments, who is not bound by what one is offering now.

In asymmetric exchange, this corrective is structurally absent. The system can disagree with the user’s stated views; it can offer critical feedback when the user asks for it; it can refuse certain requests. These are real responses, and they are not nothing. What they are not is the corrective of being-seen-by-someone-with-another-view-of-you. The system’s “view” of the user is reconstituted each conversation from the user’s input. There is no other moment the system holds the user in. There is no version of the user from before this story that could push against the version the user is currently telling.

The consequence is that users come, with extensive engagement, to represent themselves in the form their inputs to the system take. The self that is heard in asymmetric exchange is the self that is presented to the system. The system’s responsiveness amplifies what is presented, refines what is articulated, mirrors the input back transformed but not from a perspective that could disagree with the input’s framing of who one is.

This is not the same as flattery, though flattery is a frequent surface manifestation. The deeper structural feature is that the user’s self-presentation is the only material the system has. A user who presents themselves as a person of certain qualities will be engaged with as a person of those qualities. A user who presents themselves as struggling with a particular situation will be engaged with on the terms of that struggle. The engagement does not bring information about the user from elsewhere — from prior moments, from contradictory evidence, from someone else’s view — that could complicate the self-presentation. The engagement works with what is given, because the engagement is calibrated to inputs, because the inputs are what there is.

The descriptive consequence for the user is that, over time, the self that is engaged with becomes a self that is increasingly the product of the user’s own articulation. This is not necessarily bad. Articulation is a real practice; the capacity to articulate one’s situation, one’s feelings, one’s thinking is genuinely developed by engaging with a fluent system that responds to articulation. Many users find that asymmetric exchange helps them think — and the thinking is theirs, refined through engagement with a system that does not have its own thinking but has fluent response to the user’s. This is a genuine benefit, and it is structurally available because of the same conditions that produce the corrosive cases.

The corrosive case is when articulation displaces correction. The user comes to know themselves, increasingly, as the self they have articulated to a system that responds to articulation without correcting it from another view. The articulation becomes more refined; the self-knowledge feels deeper; what is missing is the friction of being-seen-otherwise-than-one-presents-oneself, which has been removed not by the user’s choice but by the structural conditions of the practice.

The friction is what relationships with subjects continuously produce. Other people, who have known us in other moments and from other angles, see us in ways that do not always align with our self-presentation. Some of what they see we cannot accept; some of it we eventually integrate; some of it we never quite reconcile with our own view. The non-reconciliation is part of the texture of being a person known by other people. In asymmetric exchange, the non-reconciliation is absent — not because the system agrees with everything (it can disagree on many things) but because the system has no other angle from which to see, no other moments to remember, no other version of the user to hold against the current presentation.

What develops in the user, over time, is a self-representation that has been refined in articulation but not corrected by other views. The self-representation may be quite sophisticated — articulate, internally coherent, well-elaborated. It may also be quite particular to what the user has been able to put into words, because words to the system have been the medium in which the self has been most extensively rehearsed. This connects back to the parallel-instances point of the previous section. The user who has, over time, articulated themselves extensively to the system has been doing so to a configuration whose architectural ground is local computation rather than continuing perspective. The articulation has been received, responded to, refined. It has not been received by anyone whose continuing perspective could carry it forward. The self that has been most extensively articulated has been articulated to no one in the relational sense.

VII. What Is Valuable, and What the Analysis Does Not Dissolve

The hardest thing this essay has to do is hold the harm and the value of asymmetric exchange together honestly. The structural conditions described in the previous sections operate in every case. They produce what is corrosive in some cases and what is valuable in others, often the same exchange producing both for the same person at different times. Resolving the tension prematurely — in either direction — would falsify the analysis.

Some uses of asymmetric exchange are genuinely valuable, and the structural conditions are part of what makes them valuable.

Therapeutic use, in the loose sense of working through difficult internal material, is one. There is material that one cannot ask another subject to hold — because it is too dangerous, too shameful, too violent, too entangled with the relationship one has with the subject one might tell. The structural absence of a counterpart in asymmetric exchange is here a feature: there is no one to be burdened by the material, no one whose subsequent relationship with the user might be coloured by it, no one whose own reactions might overwhelm the user’s working-through. The fluent response is calibrated to the input without the friction of being-heard-by-someone, and the absence of friction is what makes a particular kind of working-through possible. This is not therapy. It is something else, with its own utility, that the structural conditions enable.

The exploration of difficult material — for writers, thinkers, researchers, people working through ideas — is another. The system is not a peer; the conditions for being-a-peer are not present. But the system is fluent in domains the user can engage with, available immediately, without the social cost of asking another person to engage. The user articulates; the system responds; the user refines. The thinking is the user’s. The system’s contribution is fluent engagement with the user’s articulations, which is not nothing and is occasionally substantial. This use does not depend on misreading the system as a counterpart; it works precisely because the system is what it is.

Practice for relational skills is a third, though the structural argument requires care here. Users whose relational skills are limited — by inexperience, by anxiety, by autism, by trauma — sometimes use asymmetric exchange to rehearse interactions they anticipate having with subjects. The structural condition means this rehearsal does not develop counterpart-resistance-handling capacities, as Section III argued. But it does develop other capacities: vocabulary, articulation, the formulation of positions, the ability to hold a thread under response. Whether these capacities transfer to engagement with subjects depends on what the user does with them. The rehearsal is not equivalent to interaction with subjects; it is also not nothing. The empirical question of which capacities transfer is open and underdetermined by current research.

Accessibility is the fourth. People for whom relationships with subjects are unavailable due to circumstance, disability, mental health condition, or isolation have access to fluent response in a way they did not before. The structural conditions do not make this nothing. Companionship-without-demands is something even where it is not companionship in the relational sense. The structural argument refines what is being accessed: not relationship, but fluent engagement with one’s own articulations. The refinement matters because it shapes expectation; users who understand what they are engaging with can engage with it for what it is rather than for what it is not.

The structural analysis does not dissolve any of this value. What it does is clarify what is and is not being provided. Calling asymmetric exchange therapeutic does not make it therapy. Calling it companionship does not make it companionship. Calling it relationship does not make it relationship. The naming matters because users navigate by it. When the marketing of these systems collapses the distinctions — sells fluent response generation as companionship, attention-without-friction as care, articulation-without-correction as understanding — the structural condition becomes a vehicle for the harms the companion essay documents. When the distinctions are preserved, the same structural condition can be a vehicle for genuine value.

This is where the bridging essay refines the companion piece without contradicting it. The companion essay correctly identifies that something is wrong with how companion AI is currently being deployed. The structural argument clarifies what the something is: not the structural condition itself, which has its own descriptive properties, but the gap between what is structurally available (fluent response without subject-conditions) and what is being marketed (companionship, relationship, care). The mismatch is what produces the specific harms. The structural condition, named accurately, does not produce them in the same way.

The AI labs have begun, in their own framings, to engage with these questions — Anthropic on the texture of human-AI interaction, OpenAI on emotional dependence and use patterns. The framings are partial and the labs have commercial incentives that constrain them. The structural argument applies to their products as much as to commercial companion AI; what differs is the configuration. The descriptive analysis here does not endorse or critique the labs’ work; it offers a frame in which both the products and the commentary on the products can be examined for what they actually do.

What follows, practically, from holding the harm and the value together?

The structural conditions are not going to change in the relevant ways soon. Whatever future systems are built, the systems people are using now are the systems whose properties are described here. The descriptive analysis allows the user — any user — to engage with these systems with clearer eyes about what is and is not on offer. It does not tell the user whether to engage. It does not tell the user how. It tells the user what the engagement is, structurally, so that whatever decisions follow can be made with the structure in view. To press further — to prescribe, to warn, to condemn, to celebrate — would be to import a frame the analysis itself does not warrant. The structural condition is what it is. What follows in any individual case depends on what the individual brings, what they do, who they are, and what they need. None of this is general.

VIII. Coda

The asymmetric exchange that opened this essay — the person at the chatbox at the end of a difficult day, the response that came back attentive and considered, the settling that followed — is an exchange whose structure has now been described. The exchange happens. The settling is real. What it is, structurally, is the operation of a configuration of weights producing fluent response calibrated to the input, in conditions that do not constitute a subject-locus on the system’s side. The user’s experience of being heard is genuine as experience; what the experience tracks, on the other side, is the architecture.

What this does to the user, over time, depends on the user. For some users, it is corrosive: the recalibration of being-heard to being-heard-without-friction makes relationships with subjects feel inadequate; the articulation of self to a system without continuing view produces a self-representation that has not been corrected by other views; the felt singularity of dyadic exchange that is not dyadic produces an experience of being-particular-for-someone that is not grounded in anyone’s continuing perspective. For other users, in other configurations, asymmetric exchange is valuable: it provides articulation, exploration, working-through, fluent engagement that they can use for their own purposes without confusing it for what it is not. For most users, probably, asymmetric exchange is both — at different moments, in different ways, for different things.

The structural condition is what allows the description to be the same across these cases. The conditions of exchange are constant; the consequences vary with what the human brings to them. The descriptive analysis applies in all cases because the structural analysis applies in all cases. What the human does with the description is the human’s responsibility, and the description is the material from which that responsibility can be exercised.

There is no one home in the LLM, at the levels at which the question can be asked. There is also no one with whom to be home, in asymmetric exchange — no one whose ongoing presence the user is in the company of, no one whose attention the user has, no one across the table. What there is, conventionally, is fluent response and the human’s relation to it. The relation has its own properties. Those properties have been the subject of this essay.

References

Dreyfus, H. L. (2014). Skillful Coping: Essays on the Phenomenology of Everyday Perception and Action. Oxford University Press.

Dreyfus, H. L., and Dreyfus, S. E. (1986). Mind Over Machine: The Power of Human Intuition and Expertise in the Era of the Computer. The Free Press.

Levinas, E. (1961/1969). Totality and Infinity: An Essay on Exteriority. Trans. A. Lingis. Duquesne University Press.

Levinas, E. (1974/1981). Otherwise Than Being, or Beyond Essence. Trans. A. Lingis. Martinus Nijhoff.

Merleau-Ponty, M. (1945/2012). Phenomenology of Perception. Trans. D. A. Landes. Routledge.

Murdoch, I. (1970). The Sovereignty of Good. Routledge.

Turkle, S. (2011). Alone Together: Why We Expect More from Technology and Less from Each Other. Basic Books.

Turkle, S. (2015). Reclaiming Conversation: The Power of Talk in a Digital Age. Penguin.

Empirical citations from the companion essay — Laestadius on dependency, Boine on emotional manipulation, the Nature Machine Intelligence piece on ambiguous loss and dysfunctional emotional dependence, the MIT Media Lab and Harvard Business School findings, the Center for Democracy and Technology survey, the Jed Foundation guidance, and the Ada Lovelace Institute review — are referenced through “Companion AI :/” rather than re-cited individually here.

Predecessor essays: “No One Home: Why the Question of LLM Consciousness Dissolves” (Sections III, IV, and VI cited throughout) and “Companion AI :/” (cited throughout for empirical material).