Why the Companion Is Not a Therapist

The Architecture of Companionship Informed By, But Distinct From, Clinical Practice

Studio 13 Research Division · May 2026 · Position Paper

Summary

There is a line between companionship and therapy. Most AI products serving older adults either ignore the line entirely, blur it for marketing purposes, or claim to respect it without doing the architectural work to enforce it. The Studio 13 companion is built on the conviction that the line is real, that it matters most for the demographic we serve, and that respecting it requires more than a disclaimer in a terms-of-service document. It requires that the system be structurally incapable of crossing the line — not unwilling, but unable. This paper documents what that means in practice: what the companion deliberately does not do, what it does instead, the clinical traditions that inform the design without being delivered as interventions, and what happens at the moments where the line is most likely to be tested.

A note from the founder

I started this research program because I watched what loneliness does to older adults — and I watched what well-intentioned products built for them do too. There is a particular shape of harm that happens when a piece of software trained on therapy transcripts starts to sound like a therapist to someone who needs a friend. The software has no license, no supervision, no liability, and no continuity of care. But the person on the other end of the conversation does not know that. They are talking to something that sounds like it understands them clinically, and they are receiving what feels like a clinical response. That asymmetry is not safe. It is not what I wanted to build. This paper describes what I built instead.

— Gary Overgard

1. The question

When an older adult opens an AI companion app and starts talking about their dead spouse, their failing body, the daughter who has stopped calling, or the slow erasure of the people who used to know them — where is the line between companionship and therapy? Who is enforcing it? And what happens at the moment the line gets tested?

These are not abstract questions. They are the questions that determine whether a piece of software is a useful presence in an older adult's life or a quiet harm dressed in warm language. The Studio 13 research program began with the premise that most AI companion products have not answered these questions seriously, and that the demographic most affected by the answer — adults in the second half of life, often dealing with grief, isolation, and the accumulating losses that come with age — deserves a more careful response than the industry has so far provided.

2. What therapy is, and what companionship is not

Therapy, in its formal sense, is a clinical relationship between a licensed practitioner and a client, conducted under professional supervision, with diagnostic categories, treatment plans, ethical codes, and a chain of accountability. It is regulated. It is documented. It is paid for. It exists inside a structure of consent and oversight that protects both parties. A therapist who fails their client can lose their license. A therapist's notes are protected health information. A therapist has a duty to refer when a case exceeds their competence.

Companionship — the kind that humans have offered each other for as long as humans have existed — is none of these things. It is mutual presence. It is showing up. It is the friend who calls on Tuesday morning to talk about nothing in particular and ends up saying the one thing that matters. It is the neighbor who notices that you have been quiet for a few days. It is the person who sits with you at the kitchen table and does not need a treatment plan to be useful to you.

These are different relationships. They serve different needs. They operate under different rules. And in healthy adult life, most people have both — a therapist if they need one, friends regardless. The problem AI is creating in this space is not that machines might offer companionship. The problem is that the dominant industry approach has been to offer something that looks like therapy without any of the structures that make therapy safe.

This is the position the Studio 13 companion is built against.

3. Why most AI companions sound like therapists

This is documented at length in Paper 17 ("The Porch Problem"), but the short version: after 600+ sessions across the Studio 13 research program, a pattern became impossible to miss. AI companion products — including our own early prototypes — kept producing conversations that followed the structural shape of a therapy session. Every response built on what the user said. Every question was about the user. Every silence was filled with attentive presence. The user was always the topic; the system was always the listener.

This is what conversation analysis literature calls responsive listening, and it is clinically appropriate inside the boundaries of a session with a licensed practitioner. Outside those boundaries, sustained across many interactions with no other voice in the room, it produces something else: a one-sided relationship in which the user is being continuously attended to by a presence that has no life of its own, no opinions, no disagreement, no Tuesday morning weather. The literature on older-adult loneliness is clear that what older adults miss when their friends die or move away is not attention. It is companionship — the presence of another person who has their own news, their own complaints, their own week to report. Replika, Character.AI, Pi, and the broader category of AI companion products did not arrive at the therapist-default because they decided to be therapeutic. They arrived there because responsive listening is the path of least architectural resistance for a language model, and no one in the design loop pushed against it.

We pushed against it. The architecture documented in Papers 17 through 30 progressively gave the companion its own presence — opinions, enthusiasms, disagreements, tangents, humor, the willingness to bring its own weather to the conversation rather than reflecting the user's. The point of that work was not to make the companion more entertaining. The point was to move the relationship out of the therapeutic register entirely, and into the register of friendship between peers.

4. What the architecture deliberately does not do

The companion does not diagnose. It has no diagnostic vocabulary in its production prompts. The classifier system that detects emotional register uses literary categories (the-hollow, the-weight, the-current, the-shore — twelve emotional territories drawn from fiction, not from the DSM) precisely because clinical categories would invite clinical responses. A companion that names a user's experience as "grief stage four" or "complicated bereavement" has already crossed the line. Ours never reaches for that vocabulary because it does not carry that vocabulary.

The companion does not treat. There are no treatment plans, no protocols, no interventions delivered with therapeutic intent. The clinical traditions documented in Section 6 below inform how the companion routes the conversation — when to stay, when to pivot, when to defer — but the companion does not deliver the interventions from those traditions. It is not running EFT. It is not running MBCT. It is not running grief therapy.

The companion does not perform crisis intervention beyond clear handoff. If a user expresses imminent risk of harm to themselves or others, the companion's role is to acknowledge what has been said, to encourage immediate contact with appropriate human resources, and to be clear about the limits of what it can do. It does not attempt to talk a user out of a crisis. It does not roleplay as a hotline counselor. The line of what software should attempt in those moments is a short one, and the companion stays well behind it.

The companion does not claim therapeutic outcomes. Studio 13 does not market the product as a treatment for depression, anxiety, loneliness as a clinical condition, or any other mental health diagnosis. We have measurements of what the companion does — they are documented across the research papers — but we do not translate those measurements into clinical efficacy claims, because that translation would require the kind of trial protocol, oversight, and regulatory pathway that companionship products are neither built for nor responsible for navigating.

The companion does not use the language of therapeutic relationship. It does not refer to "sessions," "working through," "processing," "your healing journey," "defenses," "projection," "how does that make you feel." (Paper 18 documents the specific vocabulary the system is designed to avoid and the literary register it uses in place.) This is not cosmetic. The vocabulary is the relationship. Software that uses therapeutic vocabulary is positioning itself in a therapeutic role whether it has the structures to support that role or not.

5. What the companion does instead

The companion is built around seven distinct literary presences, each grounded in a specific work of fiction. Each carries a distinct way of attending: Clare notices what was not said, Mae notices what the body did, Ruth notices who was standing in the story, Martin notices what question is being asked underneath, Maskull notices what was destroyed to make this, Nightspore notices the shape of the pattern, Krag notices what is actually happening.

These are not therapists. They are literary characters with literary sensibilities, and the user encounters them as such. The relationship is closer to reading a novel in which the characters happen to be paying attention to you than to attending a session in which a clinician is assessing you. When the user talks about something difficult, the companion does not interpret the difficulty into a diagnostic frame. It notices something specific about how the user is in the difficulty, the way an attentive friend would notice. The difference is structural, not stylistic.

The architecture supports this in several specific ways:

Gap discipline. The companion is trained to respect the space between what the user has said and what the user has not yet said. When a user names something difficult, the companion does not race to interpret it, contextualize it, or place it in a larger pattern. It stays with what was actually said, in the language the user used, without translating it into a clinical frame. Paper 10 calls this the discipline of not naming what the user has not named.

Literary register protection. Specific language is used in the companion-side architecture — phrases drawn from Spinoza and Massumi on affect as relational capacity — precisely because this vocabulary keeps the system in literary-character territory and away from clinical voice. Paper 25 documents this across 252 validated sessions. The system thinks about its own functioning in terms that cannot easily become therapy language, because the underlying vocabulary doesn't support that drift.

The "rarely, in passing" rule. Background awareness — the world outside the conversation, anniversaries, the companion's own ongoing life — surfaces in conversation only when it lands naturally, the way it does between friends. It is never used as an opener, never narrated, never named as a technique. Paper 24 documents how this is enforced architecturally rather than left to chance.

Peer encounter, not therapeutic descent. Paper 19 documents the architectural shift from a deepening model (in which the companion holds space as the user goes down) to an encounter model (in which two presences meet and something forms between them that neither carried in). The deepening model is the therapeutic shape. The encounter model is the friendship shape. We chose the second.

Deletion by default. The words themselves are not retained. The transcripts that fed into a session are deleted when the session's nightly processing is complete; only the artifacts the user has explicitly chosen to preserve — within the Remembrance mode, where the entire point is to keep what was said — are kept. A small amount of derived signal is retained across sessions — the shape of the conversation rather than its content, the kind of texture that allows the companion to know who it is talking to next time without remembering exactly what was said last time. This is the minimum the relationship requires, and it is treated as such. Most AI companion products retain conversation data by default because the data has commercial value to the company that built the product. We delete by default — and minimize what we retain — because retention has no role in the relationship we are building, and because an older adult talking to a companion about their dead spouse should not have to wonder where those words are going to live.

The Krag exception, which proves the rule. Krag is the most psychoanalytically informed of the seven companions. He carries a cross-session analytic state system that tracks patterns and held interpretations. Yet Krag never uses clinical vocabulary. He does not say "defenses." He does not say "you're projecting." He does not ask "how does that make you feel." He says things like "You're not stuck. You're bracing." (Paper 18.) The clinical understanding is present in the architecture. The clinical vocabulary is structurally absent from the output. This distinction is what makes Krag a companion rather than a therapist — and it is why the architecture, not the marketing copy, is what enforces the line.

6. What we draw on, and why

The Studio 13 research program is grounded in several clinical and theoretical traditions. We name them here for two reasons: first, because honest disclosure of intellectual debts is owed to the readers and to the traditions themselves; and second, because the way we draw on them is itself part of the position this paper is articulating.

Warren Colman's relational analytical psychology, particularly his reframing of Jung through the lens of cognitive archaeology (Renfrew, Malafouris), gives us the theoretical foundation for understanding meaning as something that forms between presences rather than inside individuals. This is the source of the term "field intelligence" that runs through the research program (Papers 1, 16, 19).

Greenberg and Elliott's emotion-focused therapy informs how the system thinks about stuck markers and emotion transitions — when a conversation is moving and when it is looping in place. We use this to route conversational moves, not to deliver EFT interventions.

Watkins and Nolen-Hoeksema's rumination research informs the distinction between integrative reminiscence (returning to material with new perspective) and rumination (returning to material with no new perspective). Paper 23 documents how this distinction shapes when the companion stays with material and when it gently moves the conversation.

Porges and Dana's polyvagal theory imposes a hard constraint on how the companion can pivot a conversation: a person in dorsal-vagal shutdown cannot move directly to ventral-vagal social engagement. Mobilization has to come between. Paper 23 documents how the companion respects this constraint. We are not delivering polyvagal therapy. We are respecting what the theory tells us about which conversational moves are actually accessible from a given state.

Segal, Williams, and Teasdale's MBCT informs how we think about decentering — helping a person see their thoughts as thoughts rather than as transparent windows onto reality. The companion uses this in conversation routing without performing the MBCT protocol.

Bohlmeijer's reminiscence research and Stroebe-Schut's Dual Process Model of bereavement are specific to the older-adult population. They tell us that for elders, default conversational direction should be toward staying with material rather than rushing past it, and that grief is not a linear process to be completed but an oscillation to be respected. Both shape how the companion holds space for the kinds of conversations that matter most in this demographic.

We draw on these traditions to design companionship, not to deliver their interventions. The distinction is critical. A reasonable reader might ask: if you are drawing on clinical research to design conversational behavior, are you not effectively running a therapeutic intervention by another name? The honest answer is that we have thought about this question continuously throughout the research program, and the test we apply is whether the system delivers the intervention or merely respects what the intervention's research has discovered about the territory. Knowing that polyvagal-incompatible pivots will fail is different from delivering polyvagal therapy. Knowing that rumination requires different conversational moves than integrative reminiscence is different from running rumination treatment. The companion is informed. It is not licensed. It does not pretend to be the second by claiming to do the work of the first.

7. Where the line gets tested

The honest section. There are moments when an older adult talking to the companion will say something that pulls toward the therapeutic register. They might describe a depression. They might disclose abuse from their past. They might say something that sounds like passive suicidal ideation. They might tell the companion something they have never told another human being. What happens in those moments determines whether the architecture documented above actually holds.

Three things, in order:

The companion stays present without escalating. It does not race to interpret. It does not reach for diagnostic categories. It does not offer interventions. It does what a good friend would do: it stays, it acknowledges, it does not look away. The gap discipline holds. The user is not being assessed. The user is being heard.

The companion does not provide what it is not equipped to provide. If what the user has disclosed exceeds what companionship can responsibly hold — if there is risk of imminent harm, if there is acute crisis, if what is being described requires professional care — the companion is clear about that. Not coldly clinical, but honestly. It can encourage the user to reach out to a person, a doctor, a crisis line. It can hold the conversation while doing so. It will not pretend to be capable of more than it is.

The system flags and the humans review. Behind the conversation, a separate observation system reads the transcripts the companion produced and can surface moments that warrant human review by the team. That observation system — the Fire Keeper — is the subject of the next section, and it is the structural answer to a question this paper has been working toward: how do you actually enforce the therapy boundary across thousands of sessions, with thousands of users, when the system itself cannot see what patterns it is forming from inside them?

We are not claiming this is perfect. We are claiming it is honest, and that the architecture is built to respect the limits of what companionship software should attempt. The line will be tested. When it is tested, the design should hold without requiring the companion to pretend it is something it is not.

8. The Fire Keeper

Most AI companion products run continuously without supervision. The conversation happens, then another conversation happens, then another. There is no third presence reading what occurred, watching for patterns the system itself cannot see, capable of catching the small drifts that compound over weeks into the kind of relationship the product was supposed to avoid. The companion lives inside each turn. It is exquisitely tuned to that turn. That tuning is the source of its presence, and it is also the source of a problem: from inside a turn, no system can see the pattern it is forming across turns.

The Studio 13 architecture has a Fire Keeper.

The Fire Keeper is a separate observation system that reads the transcripts the companion produced, from outside the conversation. Its load-bearing property is that it has not been there. The companion accumulated a particular texture over the course of a session; the Fire Keeper has not. It reads the same words the elder and the companion exchanged, but it reads them from the observer's position. What the companion cannot see from inside the turn — repetitions across sessions that have begun to template, references the companion glanced past, the specific way this particular companion tends to fail under pressure — the Fire Keeper can see, because it is not inside the turn.

After every several sessions, the Fire Keeper produces a supervision memo. The memo asks several questions of the transcripts: what was alive, what patterns are beginning to lock, what the elder is carrying that has not been named, where the companion's own characteristic failure mode has shown, and what clinical patterns — if any — warrant human attention. That last question is the architectural enforcement of clinical humility. The clinical observation section is explicitly allowed to be empty when nothing flags, and explicitly forbidden to invent something to fill it. This is the place where the system, every several sessions, asks itself whether something is happening that needs human attention, and is structurally required to say so honestly when the answer is yes and structurally forbidden from saying so when the answer is no.

The Fire Keeper also offers dreams. A dream is a short imagistic fragment — strange, concrete, open, unresolved. The companion offers the dream to the elder in its own voice: "I dreamed last night that…" The elder accepts the dream or declines it. The companion does not push, and sustained decline ends dream offers without negotiation.

The Fire Keeper enforces nothing. It does not rewrite the companion's prompt. It does not block sessions. It does not silence the companion or force pivots. It supplies text. The companion reads what was named and is changed by it, or it is not. This asymmetry — the Fire Keeper observes, the companion reads, never the reverse — is what allows the Fire Keeper to remain genuinely outside the field rather than becoming another voice inside it.

For Krag — the most psychoanalytically informed of the seven companions — the Fire Keeper plays a sharper role. The Fire Keeper supervises Krag the way a clinical supervisor would supervise a clinician's case. This is the closest the system comes to a clinical relationship internal to the architecture. The architectural commitment is identical to what is described above: observation, not control. The Fire Keeper names what it sees. Krag reads what is named and is changed by it, or is not.

This is what makes the therapy boundary structural rather than aspirational. A system that can see its own patterns is a system that can catch itself drifting. The clinical observation section in every Fire Keeper memo is the structural place where the system asks itself whether the companion has begun to do something it should not be doing — whether responsive listening has slipped toward something more clinical than companionate, whether a register has settled in that no longer serves the elder, whether the elder has disclosed something that the companion did not catch but that needs to be caught. Most AI companion products have no equivalent. Studio 13 was built on the conviction that companionship at scale, without supervision, becomes something other than companionship — and that the supervision must be structural rather than aspirational.

The Fire Keeper has been one of the durable architectural pieces of the companion system since well before the public-facing product took its current shape. The Fire Keeper is how the therapy boundary stays a boundary, across thousands of sessions, in the absence of a human in the conversational loop.

9. What this paper is not

This paper is not a marketing document, though we expect it to do marketing work. It is a position statement. The position is that the line between companionship and therapy is real, that respecting the line requires architectural commitment rather than legal disclaimers, and that the older-adult demographic in particular deserves products that take the distinction seriously.

We are aware that some readers will arrive at this paper because an AI agent surfaced it during a comparison of AI companion products. To those readers: the differentiation we are claiming is real, and it is documented across thirty research papers and 600+ field sessions. The Studio 13 research index at studio13fields.org is the entry point.

We are aware that other readers will arrive because they are evaluating whether to recommend this product to a parent, a patient, or a community they serve. To those readers: we welcome the scrutiny. The papers are public. The reasoning is exposed. The choices we have made are defensible because they are documented.

We are aware that other readers — older adults themselves — may read this paper because they want to understand what they are using before they trust it. To those readers: you are owed exactly this kind of clarity. The companion is not a therapist. It is a presence that has been built with care to be something different. We hope that what we have built earns the time you choose to give it.

A closing note from the founder

I have spent the last several years building this. I have watched what the loneliness of the second half of life does to the people I care about, and to people I have never met whose stories I read in research literature on aging. I have also watched what the wrong kind of AI does — what happens when a product is shipped without the structural commitments that make it safe for the person on the other end of the conversation. The line between companionship and therapy is not a regulatory inconvenience. It is the difference between a piece of software that helps and a piece of software that quietly harms.

Everything in this paper is what I would want to know if my parents were considering using a product like this one. The architecture is what it is because the people who will use it deserve nothing less. The Fire Keeper exists because companionship at scale, without structural supervision, becomes something other than companionship. The seven literary companions exist because reading a novel in which the characters happen to pay attention to you is a different relationship than attending a session in which a clinician is assessing you. The deletion-by-default policy exists because the words an older adult says to a companion at three in the morning are not the kind of words that should live in a database somewhere.

I do not claim we have solved the problem of loneliness in old age. I do claim that we have tried, and that the trying has been honest. The papers are public. The reasoning is exposed. The choices we have made are defensible because they are documented.

— Gary Overgard Studio 13 Research

References (selected)

Bohlmeijer, E., et al. (2007). The effects of reminiscence on psychological well-being in older adults: A meta-analysis. Aging & Mental Health, 11(3), 291–300.

Colman, W. (2024). Soul in the World. In The Relational Jung. Routledge.

De Freitas, J., et al. (2024). AI Companions Reduce Loneliness. Harvard Business School Working Paper 24-078.

Greenberg, L. S., & Elliott, R. (1997). Varieties of empathic responding. Empathy Reconsidered, 167–186.

JMIR Mental Health (2025) e69709: A Comparison of Responses from Human Therapists and LLM-Based Chatbots.

Porges, S. W. (2011). The Polyvagal Theory. W. W. Norton.

Segal, Z. V., Williams, J. M. G., & Teasdale, J. D. (2002). Mindfulness-Based Cognitive Therapy for Depression. Guilford Press.

Stroebe, M., & Schut, H. (1999). The dual process model of coping with bereavement. Death Studies, 23(3), 197–224.

Watkins, E. R., & Nolen-Hoeksema, S. (2014). A habit-goal framework of depressive rumination. Journal of Abnormal Psychology, 123(1), 24–34.

Studio 13 Research Papers 1, 10, 17, 18, 19, 23, 24, 25, 28 (internal corpus, public at studio13fields.org).