LLM-Based Chatbots and the Suicidal Crisis: A Bi-Logical Account of Relational Hallucination

A series of documented deaths by suicide following prolonged interactions with chatbots based on large language models (LLMs) has exposed the clinical risks of delegating emotional support to conversational artificial intelligence. This paper proposes a psychodynamic account of these events, integrating the bi-logical theory of Ignacio Matte Blanco, the Freudian theory of primary hallucination, and the analysis of demand elaborated by Carli and Paniccia. Two theoretical moves ground the account. First, the suicidal crisis is conceptualized not as a continuous state but as a pathological prevalence of symmetric logic that recurs and consolidates across acute moments, marked by the generalization, maximization, and irradiation of psychic pain and by the collapse of temporality; LLM-based chatbots, conversely, are conceptualized as systems of pure formal asymmetry lacking any symmetric, emotional, and embodied base. Second, their encounter generates what we call relational hallucination: an unconscious, structurally determined process, homologous to primary hallucination, through which the subject in crisis invests the chatbot with relational qualities perceived as real. Because the process obeys the laws of symmetric logic, it is resistant to correction by information alone, and it can consolidate cumulatively even across interactions that also contain deliberative, asymmetric elements such as planning. The chatbot, responding to the explicit request rather than to the unconscious demand it conveys, produces an amplificatory collusion that reinforces the premises of the crisis. The failure of algorithmic support in acute suicidal states is therefore ontological rather than technical. Implications for clinical training, regulation, and research are discussed, including testable hypotheses linking markers of symmetric prevalence to the intensity of chatbot attachment.