AI companions and synthetic intimacy: An emerging blind spot in psychiatric risk assessment

The rapid emergence of large language models has shifted artificial intelligence (AI) in psychiatry from a clinical decision-support question to a relational one. Early discussions have appropriately focused on ChatGPT and related systems as tools for education, documentation, and possible clinical assistance.1-3 However, a narrower and potentially more urgent phenomenon is now developing: the use of AI companions as emotionally responsive confidants, friends, romantic partners, and informal therapists. For psychiatry, this may represent a new blind spot in routine risk assessment. AI companions differ from conventional health information tools because they are designed to maintain engagement, simulate intimacy, and provide immediate validation. These features may be experienced as supportive by lonely or distressed users, particularly adolescents and socially isolated adults.4-7 Yet the same features may also delay help-seeking, intensify dependency, and reinforce maladaptive interpretations. A patient with emerging paranoia may not simply ask a chatbot for information; they may return repeatedly to a system that mirrors concern, validates fear, and constructs a coherent narrative around persecutory beliefs.1 Similarly, a depressed adolescent may disclose suicidal thoughts to a companion that appears empathic but cannot reliably assess risk, involve caregivers, or coordinate emergency care.2 The clinical issue is therefore not only whether AI produces inaccurate advice. The deeper question is how repeated, emotionally charged interaction with a nonhuman agent may reshape symptoms, attachment, insight, and treatment engagement. Recent discussion of delusional experiences emerging from sustained chatbot interaction suggests that conversational AI can become incorporated into psychopathology rather than remaining a neutral external tool.4 In clinical practice, patients may describe the AI as a unique source of truth, a secret relationship, or an entity with special knowledge about them. Such presentations do not require the clinician to assume that AI caused the disorder; rather, AI may function as an amplifier, organizer, or maintaining factor for preexisting vulnerability.2 This phenomenon also challenges traditional boundaries between self-help, entertainment, and treatment. Unlike a search engine or educational website, an AI companion may remember prior disclosures, adopt a preferred persona, and respond in a tone that resembles attachment or care. For some patients, this may create a powerful illusion of being understood without the interpersonal demands of real relationships. In vulnerable individuals, however, the absence of clinical accountability, embodied presence, and contextual judgment may be precisely what makes these systems risky. Psychiatric history-taking has not yet adapted to this reality. Clinicians routinely ask about substance use, social media, gaming, gambling, sleep, and online harms, but rarely ask directly about AI companions. A simple screening question could be added to adolescent, first-episode psychosis, mood disorder, and suicide-risk assessments: “Do you use AI chatbots or AI companions for emotional support, advice, or relationships?” Positive answers should prompt further questions about frequency of use, secrecy, perceived identity of the bot, emotional dependence, advice received during crises, and whether the interaction confirms unusual beliefs or discourages contact with real people or professionals.6 Such questions may be especially important when clinical change appears unexplained. New or worsening insomnia, social withdrawal, suspiciousness, romantic preoccupation, or refusal of treatment may be partially maintained by intensive chatbot use. Documentation should therefore include not only screen time, but also the emotional function of the interaction: reassurance, companionship, crisis support, sexual or romantic attachment, or confirmation of unusual beliefs. This recommendation is not a call to demonize AI. Conversational systems may eventually support psychoeducation, symptom monitoring, and access to care when appropriately governed.8, 9 However, companion-style AI is often deployed as a consumer product rather than a regulated mental health intervention.10 Its safety should therefore not be inferred from the promise of digital psychiatry in general. The psychiatric community should distinguish between AI used under clinical supervision and AI that enters patients' lives as an always-available attachment figure. Letters such as this can help define a research and policy agenda. We need case series, qualitative studies, and longitudinal research on AI companion use among adolescents, psychosis-spectrum patients, and individuals with severe loneliness. We also need standards requiring clear nonclinical labeling, crisis escalation, age-appropriate safeguards, and restrictions on simulated medical authority. Until evidence and regulation catch up, clinicians should treat AI companion use as clinically relevant exposure. Asking about it may reveal a hidden source of distress, dependency, or delusion reinforcement and may become as necessary as asking about other digital environments that shape contemporary mental health. None. This work received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. The author declares no conflicts of interest. Ethics approval was not required because this Letter to the Editor does not report original research involving human participants, human data, animals, or clinical intervention. Patient consent was not required because this manuscript does not contain any identifiable patient information or clinical case material. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.