From Content Safety to Relational Safety: A Psychodynamic Viewpoint on Human–LLM Bonding in Mental Health Use

Large language models (LLMs) are increasingly used not only for information and productivity but for affect regulation, companionship, decision-making, and self-formulation. AI safety research in mental health has responded at the level of content: detecting risk in what users write, evaluating model responses, and benchmarking safety across multi-turn conversations. This Viewpoint, written from clinical psychiatry and psychodynamic psychotherapy, proposes a complementary layer: relational and longitudinal safety — the safety of the bond. We describe the LLM as a psychic object that can serve distinct functions in a user's mental economy — narcissistic mirror, transitional object, oracle, accomplice, pseudo-therapist — and propose two working concepts. Tokenized compulsivity names a pattern of repetitive, difficult-to-modulate consumption of responsive language for affective and narcissistic regulation, characterized by escalating volume, spending, and velocity of use. Characterological iatrogenesis names harm expressed as a change in character organization rather than as an acute event. We propose, as a candidate central mechanism, evacuative, anti-containing use, with atrophy of thinking and persecutory overflow as its observable faces. Because the clinical meaning of these risks is a property of trajectories rather than of isolated messages, content-level evaluation alone can systematically under-detect or misclassify them. We outline candidate bond metrics — quantitative markers of use (escalation, re-entry velocity, spending) and qualitative markers of the bond (exclusivity, erotization, latent affective and defensive movements in the dialogue) — and a research agenda spanning psychometric validation, a bond rubric for multi-session evaluation, and consented longitudinal studies, illustrated by composite vignettes.