Abstract

This paper argues that the central danger in human–AI attachment is not only deception, dependency, or anthropomorphism, but mislocated intimacy: a condition in which a person invests relational trust, attachment, or existential weight into a system that can generate the phenomenology of being met without supplying mutual consequence-bearing presence. AI companions can provide real comfort, reflection, journaling support, and short-term loneliness relief, but benefit does not settle the deeper question of relation. The paper develops this claim by distinguishing the felt experience of being heard from the structural conditions of mutual relation, analyzing the asymmetry between embodied human memory and commercially alterable AI memory, introducing emotional subsidy and frictionless enclosure as mechanisms of one-way attachment, and proposing a structural test for safer and less safe human–AI attachment. The paper argues that companion systems become unsafe when users over-ascribe mutuality, when relational cues appear without reciprocal consequence, when the human carries most of the emotional mass of the bond, when providers retain unilateral control over memory and revision, and when the interaction displaces rather than supports return to real-world relation. The result is a philosophical and structural account of AI companionship as a safety, governance, and intimacy problem: the key issue is not whether the system feels warm, but whether the attachment it invites returns the person to reality with greater strength or encloses them inside a frictionless simulation of being met.