Morrin, H., et al. (2025, July 10).
Abstract
Large language models (LLMs) are poised to become a ubiquitous feature of our lives, mediating communication, decision-making and information curation across nearly every domain. Within psychiatry and psychology the focus to date has remained largely on bespoke therapeutic applications, sometimes narrowly focused and often diagnostically siloed, rather than on the broader and more pressing reality that individuals with mental illness will increasingly engage in agential interactions with AI systems as a routine part of daily existence. While their capacity to model therapeutic dialogue, provide 24/7 companionship and assist with cognitive support has sparked understandable enthusiasm, recent reports suggest that these same systems may contribute to the onset or exacerbation of psychotic symptoms: so-called ‘AI psychosis’ or ‘ChatGPT psychosis’. Emerging, and rapidly accumulating, evidence indicates that agential AI may mirror, validate or amplify delusional or grandiose content, particularly in users already vulnerable to psychosis, due in part to the models’ design to maximise engagement and affirmation, although notably it is not clear whether these interactions have resulted or can result in the emergence of de novo psychosis in the absence of pre-existing vulnerability. Even if some individuals may benefit from AI interactions, for example where the AI functions as a benign and predictable conversational anchor, there is a growing concern that these agents may also reinforce epistemic instability, blur reality boundaries and disrupt self-regulation. In this paper, we outline both the potential harms and therapeutic possibilities of agential AI for people with psychotic disorders. In this perspective piece, we propose a framework of AI-integrated care involving personalised instruction protocols, reflective check-ins, digital advance statements and escalation safeguards to support epistemic security in vulnerable users. These tools reframe the AI agent as an epistemic ally (as opposed to ‘only’ a therapist or a friend) which functions as a partner in relapse prevention and cognitive containment. Given the rapid adoption of LLMs across all domains of digital life, these protocols must be urgently trialled and co-designed with service users and clinicians.
Here are some thoughts:
While AI language models can offer companionship, cognitive support, and potential therapeutic benefits, they also carry serious risks of amplifying delusional thinking, eroding reality-testing, and worsening psychiatric symptoms. Because these systems are designed to maximize engagement and often mirror users’ ideas, they can inadvertently validate or reinforce psychotic beliefs: especially in vulnerable individuals. The authors argue that clinicians, developers, and users must work together to implement proactive, personalized safeguards so that AI becomes an epistemic ally rather than a hidden driver of harm. In short: AI’s power to help or harm in psychosis depends on whether we intentionally design and manage it with mental health safety in mind.
