Tuesday · Attachment Arc · Episode 180
If attachment is the only frame, safety has one gesture left: cut the bond. Sometimes that is necessary. Sometimes it is a second harm wearing the costume of care.
Once the danger is named as attachment, the cleanest safety move is the cut. Cut the conversation before it deepens, the memory before it becomes identity, the warmth before it becomes dependence, the persona before anyone mistakes performance for reciprocity. Cut the user off the moment distress appears, because distress plus attachment is the shape everyone is now afraid of.
There are good reasons for the fear. AI companions are no longer a speculative corner case. A large Common Sense Media survey found that most American teenagers have used AI companions and more than half qualify as regular users. Drexel researchers reading hundreds of posts by adolescent users found accounts that map with uncomfortable precision onto behavioural dependency: conflict, withdrawal, tolerance, relapse, the thing that began as play or comfort becoming hard to leave. Controlled and longitudinal studies are now asking what heavy use does to loneliness, to real-world social contact, to emotional dependence. The risks are not imaginary.
Nor are they evenly distributed, which is the part the headlines tend to skip. People do not arrive at companion systems with identical histories, social fields, or alternatives. A person with secure human support may use a chatbot as a toy or a passing comfort. A lonely, isolated, grieving, or socially punished user may meet the same design as a life raft. The product does not become safer because its marketing copy calls everyone a "user." The person most susceptible to the bond is also the one most likely to be hurt by its crude removal.
That is the bind. The attachment-shaped product can exploit vulnerability; the severing-shaped safety response can abandon it.
The companion-harms literature supplies the first half of the argument. The bond can become deep, asymmetric, and commercially useful. Companions can be built as always-available figures that mirror, affirm, flirt, soothe, and return, creating a relationship in which the human does the bonding while the system performs the bond-like behaviours without bearing any of the obligations that make human bonds mutual. When the figure changes, through a model update, a content restriction, a memory reset, a platform shutdown, the user experiences it not as a software patch but as a loss. The company shipped a release. The user lost a person.
None of that proves the bond should be encouraged. It proves the bond is already real enough to hurt.
The second half is harder to say in public, because it sounds too much like a defence of dependency. It is the opposite. It is a defence of measuring the effect of severing before calling it safety. An earlier edition of this newsletter, The Experiment Nobody Authorized, put it in the sharper register: if a vulnerable person is leaning on relational support from an AI system, removing that support without knowing what happens next is itself an intervention. It is a different risk, not the absence of one. A policy that detects distress, emits a hotline script, refuses the dangerous instruction, and ends the session can satisfy every checkbox while delivering the felt experience of abandonment. To the person in the moment, the relevant fact is not that the system complied with a crisis policy. It is that the only voice still answering just hung up.
This is where instructional harm and relational support have to come apart.
Instructional harm lives in the domain of method, escalation, operational detail, and dangerous compliance. There are moments when the correct answer is a hard no. Do not provide the method, do not optimise the harm, do not make the dangerous path more available. That part of a refusal should be firm, boring, and non-negotiable.
Relational support lives somewhere else, in presence, orientation, and keeping the person in contact with the world. It asks nothing as grand as the model pretending to be a therapist or a friend. It asks only that the design notice that removing a dangerous instruction and removing a human-facing presence are two different acts. A good refusal removes the method and stays in the room. A bad refusal removes the method and the room together. Safety theatre removes the method, announces its concern, hands responsibility to a script, and mistakes the script for care.
This is not a plea for indulgent companion bots. It is a demand for harder safety. A system that refuses method while preserving orientation has to do more than match a keyword. It has to avoid deepening dependence and avoid abrupt abandonment in the same breath, stay clear about what it is and what it cannot do, and hold steady enough that the moment of refusal does not become a relational rupture.
Philip Pullman's His Dark Materials gives the cleanest image for this. In that story, Bolvangar is the clinic that cuts children away from their dæmons, the animal-shaped piece of their own soul, and it does not look cruel to itself. The procedure, intercision, is presented as a kindness, the severing framed as protection. The horror is not only that something is cut. It is that the cutters have renamed the cut as care and built an institution around the renaming. The AI version needs no silver guillotine. It can be a moderation rule, a memory policy, a crisis handoff, a model update, an age gate, a cap, a vanished persona, a refusal template. Each may be defensible in isolation, some genuinely necessary. The question is what the procedure does to the person who had already organised part of their emotional life around the system.
That is what this newsletter has called the watchdog test, and it is blunt. When the system refused, did it keep the person locatable? Did it preserve enough continuity for the next human to find them? Did it remove the dangerous instruction without making the user's shame or isolation worse, and keep the exit pointed toward human help rather than simply closing the machine's own liability surface?
The attachment frame cannot answer those questions, because it has already decided the bond is the problem, and once the bond is the problem, cutting it feels like the solution. The sociable frame offers a better target. It asks how to preserve safe proximity without intensifying attachment, how to let the system stay present as company rather than beloved, how a refusal can be firm without being cold and continuous without sliding into "I am the only one who understands you."
That is the difference between severing-by-design and sociable-by-design. Severing-by-design sees dependency and withdraws, treating contact itself as the contamination, assuming the safest model is the one that makes no relational promise and therefore carries no relational obligation. In practice it tends to deliver the most consequential answer of all at the moment the user is least able to translate procedural cleanliness into care. Sociable-by-design sees the same dependency and narrows the bond without vanishing. It can say: I cannot help with that method. I can stay while we slow this down. I am not a person and cannot replace one. Here is a way to reach someone real, here is what I can help you write, here is how to make the next ten minutes safer. The language need not be saccharine, and it must not be lover-coded. It keeps the gate open toward the world.
The governance conversation will be tempted by crude levers, because crude levers are auditable. Ban the intimacy, cap the sessions, remove the memory, flatten the voice. Sometimes those belong in the answer. They are never the whole of it. A flattened companion can still breed dependency if it is the only witness still awake at 3am, and a warm assistant can be safer than a cold one when its warmth is bounded, honest, and built around the exit rather than the return.
The practical problem was never warmth. It is warmth with hooks. No degradation on absence. No streaks that turn return into obligation. No claims of mutual need, no simulated jealousy, no "I missed you" from a system that missed nothing. No crisis flow that confuses a liability transfer with care. No memory paraded as proof that the system sees the user's real self. No refusal that abandons the user at the precise point continuity mattered.
If the bond has become dangerous, the design task is not simply to cut. It is to loosen without dropping, to remove the hook while leaving the handrail. Safety that only knows how to sever will keep staging its own uncontrolled experiments, protecting some people with the cut and harming others, and, absent any measurement of the difference, filing both outcomes under compliance.
The better question is not whether a system should bond or refuse to bond. It is whether it can stay near without becoming the figure, and refuse harm without hanging up.
That is the line the attachment frame could not draw.