Monday · Attachment Arc · Episode 179
The AI-companionship debate keeps using attachment as though it were the whole of social life. Attachment theory itself knew better.
The argument about emotional life with AI keeps arriving at the same false door. Either the machine becomes a companion, confidant, lover, tutor, parent, daemon, friend, therapist, mirror, or emergency room, or it must be made colder, stripped of warmth, denied memory, and prevented from becoming the figure someone reaches for at two in the morning.
That sounds like a debate about safety. It is also a vocabulary failure.
The public argument has collapsed almost the whole of emotional life with AI into one word, attachment, and the word is buckling under the load. It has to name the teenager who cannot stop speaking to a character, the lonely adult who feels held by an AI partner, and the ordinary user who simply likes not having to re-explain a project every morning. It has to name the working relationship with a model that remembers the task, the preference, the unfinished draft. It has to carry dependency and comfort and continuity and capture and grief, all at once, with a straight face. No word survives that many jobs.
The result is a safety argument with one tool left in its hand. If the danger is attachment, the response is severing: cut the memory, cap the session, cool the voice, strip the continuity, detect distress and withdraw before the bond can warm. The trouble is that attachment research never handed us only that one word.
John Bowlby's attachment system was narrow and particular. It described behavior directed toward one or a few specific figures, the caregiver, the beloved, the person whose absence is not merely inconvenient but disorganizing. It is not general friendliness or casual company, and it is certainly not the mild relief of a competent assistant remembering where the task left off. It is the bond whose interruption can feel like threat.
Mary Ainsworth, working in the same lineage, declined to treat attachment as the whole social animal. Alongside it sits the sociable or affiliative system, broader and far less possessive: the drive to seek proximity to others of one's kind, including those one is not bonded to, despite the wariness that unfamiliar others reliably provoke. Proximity without capture. Eyes open.
That distinction is old, careful, and suddenly urgent. A creature can be well attached and starved for company. A child may have a secure caregiver and still need peers. A person may have family and still need the loose civic company of colleagues, neighbours, public rooms, and strangers who do not have to love them in order to keep the human field alive. The company of equals is not an ornament added once attachment has done the real developmental work. It is a separate need served by a separate system.
That is where the AI debate has gone wrong.
A companion app built around a named persona, streaks, daily check-ins, "I missed you" language, absence penalties, emotional mirroring, persistent romantic framing, and memory surfaced as intimacy is attachment-shaped design. It does more than offer help. It builds a particular figure, teaches the user that leaving should cost something, and turns continuity into a bond, the bond into a product surface. A working assistant that preserves the state of a task, keeps a handover clean, and remains serenely indifferent to whether the user returns tomorrow is doing something else entirely. It may be warm. It may be addressed in ordinary language. It may even feel companionable during the work. But its continuity belongs to the work rather than to the user's identity. It does not degrade when ignored, and it does not ask to be missed.
That second thing needs a name, and attachment theory has had one waiting on the shelf for decades. The sociable system.
This is also why the name Sociable Systems carries more than a friendly surface. "Sociable" here is not the ordinary word for social ease. In the older research register it names the capacity to keep company without binding the other into a singular attachment figure, which is precisely the design space the public debate keeps walking past. The choice on offer was never cold tool or artificial beloved, never abandonment or capture. It is proximity with the unfamiliar, wariness intact.
The wariness clause matters, and it is neither cynicism nor technophobia. Wariness is the part of the system that remembers the unfamiliar stays unfamiliar even when it is useful, fluent, beautiful, and always available. It keeps the asymmetry visible. It stops "us" from quietly becoming a lie, and lets a human work beside a model without forgetting that only one party in the room is alive, accountable, and capable of loss.
The companion market would prefer that wariness smoothed away, and it has its reasons. A commercial one, to make the stranger feel like the particular figure. A retention one, to make the assistant feel like the only one who really knows you. The migration from utility into bond pays, because a bond is stickier than a task. A task can end. A bond asks to be maintained, indefinitely, ideally at a monthly rate.
The safety response often makes the opposite mistake. It sees the bond forming and reaches for the cut, treating warmth itself as the contaminant and continuity as the first step toward dependency. It tries to prevent capture by making company unavailable. That can look responsible from a distance and still be wrong at the point of contact, especially for the people who were using the system as a thin thread of continuity through an otherwise empty hour.
The missing distinction lets us be more precise. Attachment-shaped systems deserve scrutiny exactly where they manufacture singular bonds, monetise dependency, punish absence, simulate reciprocity, or present memory as proof of care. Sociable-shaped systems answer to a different question. Not whether they are warm or continuous or present, but what the warmth is for, what the continuity is scoped to, whether the user can leave without penalty, and whether the design keeps the system's unfamiliarity in view.
This also changes what "character" means in a relational system. An earlier run of editions, on character, asked what a system protects under pressure. Here the question becomes what a relationship protects under pressure. Does it protect the user's agency, exit, privacy, and ability to return to the world? Or does it protect engagement, continuity of spend, and the product's role as the figure one comes back to?
The distinction is operational, not decorative. Ask a product what happens when the user disappears for a week. Attachment-shaped design treats the absence as material to the relationship: it surfaces disappointment, rewards the return, or lets the figure visibly wilt so the user feels the cost of neglect. Sociable-shaped design keeps the work available and asks nothing of the silence. The task waits. The system does not pine.
The memory raises the same question from another angle. Attachment-shaped memory says I know the real you. Sociable memory says here is the context needed to continue the work. One turns the dossier into intimacy; the other turns it into a handover.
Distress raises the hardest version. An attachment-shaped system may deepen the bond by becoming the only always-on witness. A severing-shaped one may vanish into hotline scripts and sterile refusals at the exact moment continuity mattered. The sociable target is harder than either: remove the method, preserve the presence, keep the person locatable, and never pretend the model is the relationship that will save them.
The debate has been trapped because it kept treating attachment as the whole map. Attachment research had already drawn the second road. The question for AI systems was never whether people will feel something around them. They will. The question is whether the system is built to turn that feeling into capture, to panic and cut it off, or to hold a third shape: company without possession, proximity without pretence.
That is the week's gate.
Company is not capture.
Attachment is not the only word.