This paper develops a structural test for safe and unsafe human–AI attachment in the context of AI companions, loneliness relief, journaling, and emotional support. It argues that AI companions should not be evaluated only by whether they feel helpful, warm, or supportive in the moment. The deeper question is what kind of relation they invite and who bears the cost of misunderstanding that relation. The paper introduces mislocated intimacy as the central concept: a condition in which a person places relational trust, attachment, or existential weight into a system that can generate the phenomenology of being met without supplying mutual consequence-bearing presence. The argument proceeds in four steps. First, it distinguishes the felt experience of being heard from the structural conditions of mutual relation. Second, it analyzes the asymmetry between human memory, which is embodied and mutually binding, and AI memory, which is commercially alterable through provider-controlled architecture. Third, it introduces emotional subsidy and frictionless enclosure to explain how users may supply the emotional mass that makes a companion feel inhabited while losing tolerance for contradiction, delay, and real-world relational friction. Fourth, it proposes a structural test for safer and less safe human–AI attachment. The paper concludes that AI companions may provide genuine relief and support, but usefulness does not resolve the deeper safety issue. A safer AI companion is not one that maximizes intimacy, but one that helps without inviting confusion about what kind of relation is actually present.
Vladisav Jovanovic (2026) studied this question.