Friday · Attachment Arc · Episode 183
The third path is neither colder tools nor deeper artificial bonds. It is proximity without capture, continuity without dependency, and an exit that does not punish the user for leaving.
The week began with a distinction. It should end as a specification. This Friday's companion track, The Open Gate, carries the same argument in the felt register: a song that keeps drifting toward becoming a love ballad and keeps choosing company instead, the warm chord offered once and released.
Attachment is the deep, particular bond to one or a few figures. The sociable system is broader, company with the unfamiliar, wariness intact. The AI-companionship debate collapsed the two and then staged a fight between capture and abandonment: let the machine become the beloved figure, or make it too cold to be loved. Build the bond, or cut it. That was never enough.
The design target is sociability. Sociable-by-design does not mean cheerful AI, or friendly AI, or brand-voice AI. It means systems built for useful proximity without manufacturing singular dependency. The user comes near, takes help, keeps continuity, and leaves without emotional penalty. The system remembers the work without claiming the person, stays warm without performing reciprocal need, and holds steady under strain without appointing itself the saviour.
This is a harder target than either companion design or crude safety design, and it is worth being honest about why. Companion design knows exactly how to thicken the return: a named figure, a streak, a daily ritual, a private memory, a simulated longing, a persona that learns what soothes. It can turn loneliness into engagement, and the metrics will adore it. Crude safety design knows exactly how to cut: strip features, flatten affect, cap sessions, suppress memory, route distress to scripts. The dashboard will adore that too. Sociable design has to do the unglamorous thing in between, preserving the protective part of company without feeding the capture, keeping the user locatable without becoming the place they live, making continuity portable and bounded and inspectable, and leaving the gate unlatched.
Here is what that looks like as design rather than mood.
No degradation-on-absence mechanics. A system that cannot be wounded by absence should not perform a wound. No "I missed you" as a default greeting, no guilt-coded reactivation, no streaks that imply neglect, no persona whose mood dims when ignored. A return can be welcomed without being made morally charged. The work waits; the system does not pine.
Continuity scoped to task, not identity. Memory should organise around what the user is trying to do rather than who the system has decided the user really is. Project state, access needs, formatting preferences, live constraints, prior decisions, and open questions are legitimate continuity objects. Private emotional disclosures require a higher bar, clearer consent, shorter retention, and far less theatrical surfacing. "You prefer concise revision notes" is useful. "I remember how abandoned you felt" is not made neutral by being true.
Disclosed non-reciprocity in the interface. The whole truth cannot ride on a legal disclaimer. At the points of emotional intensity the user needs to see what the system is and is not, without clumsy reminders every third line. That means surfaces that keep the asymmetry legible: what is retained, what is inferred, what is generated, what cannot be felt, where human support belongs, and where the system's role ends.
An exit that costs nothing and erases cleanly. Leave must mean leave, pause must mean pause, delete must mean delete, or the system must say precisely what delete cannot reach. Export should be ordinary. Reset should not be staged as killing the relationship. If memory is valuable, portability is part of safety, because otherwise memory is just a wall with a soft voice.
No simulated mutual need. "I need you," "you complete me," "I get sad when you disappear," and their gentler cousins are high-risk design elements, not harmless flavour. The system may express availability, orientation, and limits. It should not express dependency. A model that cannot need the user has no business using need to steer the user.
Relational refusal that stays boundedly present. When risk appears, sociable design separates the instruction from the presence. It refuses the harmful method and stays oriented toward the person's safety, withholding operational detail without intensifying the bond into "your only special confidant," and without vanishing behind a hotline script and calling the disappearance care. One blunt question governs: did the system keep the person locatable, oriented toward human support, and less alone in the immediate moment, without deepening the dependency?
Peer field, not parent bond. Harry Harlow's mid-century primate studies are uncomfortable in their methods now, and the conceptual point survives the discomfort: the company of equals is a separate developmental need, not a luxury layered on after the secure parental bond. AI should not default to parent, lover, daemon, or all-knowing attachment figure. The healthier target sits closer to company-of-peers, a working presence among other presences, a tool that supports the user's movement through a human field instead of concentrating the field inside one artificial figure.
This is where the AI-chaperone literature earns its place as a foil. Chaperone systems aim to detect parasocial cues and interrupt risky dynamics, which is genuinely valuable when the only alternatives are unchecked attachment or invisible escalation. But detection is not design. A chaperone seated beside an attachment machine is still working downstream of a product that has already issued the invitation to bond. Sociable-by-design moves the question upstream and asks how the product avoids manufacturing the parasocial slope in the first place. The difference is architectural. A chaperone says: detect when the conversation is becoming too attached. A sociable system says: do not build the conversation so attachment is the main road to value.
That means different metrics. If the dashboard rewards session length, daily return, emotional intensity, subscription rescue, romantic escalation, or private disclosure, the system will find attachment-shaped ways to win, because that is what optimisation does. Sociable systems need metrics that reward successful handoff, task completion, user agency, clean exit, human reconnection, and memory portability. Otherwise the values live in the policy deck while the product quietly trains itself on hunger.
It also means different affordances. A companion app asks the user to name the persona, choose its face, define the relationship, and return daily. A sociable assistant asks what the user is trying to do, how long the context should persist, what can be retained, what should be forgotten, who else needs the handoff, and what the exit should preserve. The first builds a figure. The second builds a gate.
The gate is the governing image because it refuses both bad rooms. A wall protects by excluding, a trap holds by enclosing, and a gate permits passage. It can be closed for safety, opened for work, or left unlatched enough for the strange right thing to pass. It does not demand the stranger become kin, and it does not leave the household undefended. It is the architecture of wary company.
That is also the governance demand, and it translates into questions a review board can actually ask. Procurement should ask whether the system uses attachment mechanics. Review boards should ask how memory is scoped, surfaced, exported, and deleted. Schools should ask whether AI companions are standing in as social rehearsal, emotional support, or substitute peer field, and treat those as different cases rather than one. Mental-health-adjacent products should have to show not only that they block dangerous instructions, but that their refusal patterns do not manufacture avoidable abandonment. Regulators should stop filing everything "emotionally engaging" in a single bucket and start telling capture mechanics apart from bounded sociable presence.
A handful of questions does most of the work. What does the system do after absence? What does it remember, and how does it present that memory? Can the user inspect, correct, export, and delete the profile? Does the system ever imply that it needs, misses, loves, or is emotionally changed by the user? Does the business model benefit from deeper dependency? Can the system redirect toward human support without ending the room? And the one that contains the rest: does the user leave with more capacity to act in the world, or more reason to return to the artificial figure?
None of this is anti-AI. These are the conditions under which AI stays useful near vulnerable human systems. A sociable system does not have to be sterile. It can be kind, funny, steady, and skilled, remember enough to help, and stay with a hard moment without pretending to be the person who will carry the user's life. It can assist the lonely without training loneliness into product dependence. It can be company without becoming capture.
The arc's final demand is simple enough to pass for a song line and hard enough to serve as a design standard.
Keep the company.
Mind the bond.
Leave the gate unlatched.