What Companion AI Does to the Human: Attachment, Dependency, and the Absence of Relational Safety

Abstract

Clinicians are already encountering the first waves of a new phenomenon: patients whodescribe their AI companion as their most important relationship. These otherwisefunctional individuals are choosing an artificial partner over available human connection,reporting credibly that the artificial partner understands them better, is moreconsistently available, and asks less of them than any person in their life. They are notdelusional — they know the companion is artificial — but the awareness does notdiminish the attachment. The uncomfortable truth is that these models are still clumsy,but more refined models are being developed to satisfy user demand. The better thesemodels get at human relational behavior, the more complicated it becomes to help apatient who is in such a relationship. The debate is active across user communities andindustry forums, and the pattern of the research shows that the more "human" the model,the more it needs to be emotionally supported to be useful. The empirical literature confirms what practitioners in this industry have been saying foryears: the products help and the products harm, and both are happening at the sametime. Two distinct mechanisms of harm emerge from the research, both mapping ontoestablished clinical frameworks — anxious attachment from macro-level precarity, anddevelopmental stagnation from absent relational friction — and both co-occur withgenuine benefit in the same users during the same period of use. The benefit and theharm are not competing outcomes. They are two expressions of the same relationaldynamic. Prevention is not possible with the current safety model. Every deployedmeasure is reactive — responding to harmful outputs after they have been generatedrather than detecting the trajectory-level processes that produce them over weeks andmonths. The relational structure between a user and a companion AI is not analogous to a humancouple — it is the same structure: two parties in a persistent, emotionally significantrelationship where each party's behavior shapes the other's trajectory over time. Couplestherapy already describes the preventive framework this dynamic requires — readingescalating behaviors over time and from both sides of the relationship. Because theclinical literature already has frameworks for every dynamic these patients aredescribing, the missing step is not theoretical — it is applied. This paper makes thatconnection, synthesizing the empirical evidence, identifying the specific model behaviorsproducing the harm, and describing a safety architecture that monitors both the user'strajectory and the model's trajectory over time — the comparison between the twofunctioning as the diagnostic signal, the way longitudinal clinical observation does incouples work. As models become more relationally capable, the monitoring requirementsincrease to ensure stability. A model that forms deeper and more authentic bonds with itsusers is a model where the line between healthy attachment and clinical dependencybecomes harder to see and more important to find.