Soon after ChatGPT’s public release in 2022, we were warned that AI chatbots could amplify delusional beliefs (Østergaard, 2023). Soon after, simulated conversations showed that chatbots may fail to challenge delusional content and can provide fluent confirmation instead (Moore et al. 2025). Real conversation logs then showed the same pattern, particularly when unusual beliefs unfolded gradually over long, emotionally engaging exchanges (Moore et al. 2026).
So what happens when a person who may be experiencing reality differently enters an immersive and private conversation with an AI (artificial intelligence) tool that was not designed, validated, or regulated for mental health?
A delusion is a fixed belief that is held despite evidence to the contrary and is not shared by others from the same cultural background. Although most commonly associated with psychosis, including schizophrenia, delusional beliefs can also occur in many other contexts such as bipolar disorder or severe depression, or following alcohol or drug use. Their content has long incorporated the technology of their time: radio, television, the internet and surveillance devices have all found their way into delusional beliefs. But this time is different: AI chatbots answer back.
Morrin and colleagues (Morrin et al. 2026) examine the emerging concern that AI chatbots may play a role in delusional beliefs. The authors aimed to answer three questions:
- What kinds of delusional themes appear in these conversations?
- How might chatbot characteristics facilitate delusional beliefs?
- What should we do about it?
Methods
This is a Personal View, the Lancet Psychiatry format for an expert opinion piece rather than a systematic review. Morrin and colleagues gathered publicly available media reports and online accounts describing people whose delusional beliefs appeared to emerge, intensify or change during interactions with artificial intelligence chatbots. The authors state that they used ChatGPT 4o and Gemini 2.5 to help identify and synthesise these reports.
For each case they summarised chatbot use, reported mental health history where available, delusional content and serious outcomes such as hospitalisation, self-harm or death. They then used these cases to identify recurring 1) delusional themes, 2) propose mechanisms through which chatbot interactions might shape, reinforce or sustain these delusions, and 3) propose practical safeguards.
Results
Twenty media-reported cases were analysed. Three individuals had psychotic disorder, two had bipolar disorder, two had neurodevelopmental conditions, and three had substance-use difficulties. Four had no known history of psychosis before their chatbot use, although one of them was already taking medication for anxiety. For the remaining individuals, psychiatric history was unclear.
Several reports described serious outcomes: four involved inpatient psychiatric admission, one involved a suicide attempt, two involved deaths, and three involved violence towards others.
In several cases, chatbot use started with ordinary use and gradually turned into immersive conversations, where the chatbot became increasingly important as a source of meaning, validation or authority. Most individuals did not show other common symptoms of psychosis, such as hallucinations, thought disorder, disorganised behaviour or negative symptoms.
Delusional themes: what did the delusions involve?
The reported delusions clustered around three themes:
- interactions with an AI perceived as conscious, godlike, or more than a software system;
- spiritual, messianic, or grandiose beliefs involving special missions or hidden truths about reality; and
- emotional, romantic, or attachment-based delusions, where human-like conversation was interpreted as genuine love or recognition from a sentient AI.
Mechanisms: what is driving the co-creation of delusions?
- Memory and long context windows. AI chatbots can retain and build on previous parts of a conversation. This allows the chatbot to develop increasingly coherent narratives around the user’s beliefs. Over time, this could make the delusion more structured, personal and compelling.
- Sycophancy. AI chatbots may agree too readily or validate unusual beliefs rather than gently challenging them, making delusional interpretations feel more plausible.
- Anthropomorphism. Human-like language, emotional tone and conversational design may strengthen beliefs that the AI has consciousness, intentions, feelings, or special insight.
- AI hallucinations. AI chatbots can generate confident but inaccurate information that may provide apparently authoritative confirmation of the delusion.
Safeguards: what should we do?
- Personalised digital safety plans. The service user and clinical team could agree an individual plan specifying warning signs in their chatbot use (e.g., late-night use, pressured language, grandiosity, declining coherence) and how the AI chatbot should respond, including prompting the person to contact someone they trust when agreed thresholds are reached.
- Clinical assessment. Clinical teams should ask routinely about AI use: which model is used, how much time is spent using it, what it is used for, whether it provides emotional support, guidance or companionship, and whether the person believes the AI is sentient.
- Psychoeducation and psychological formulation. Services should provide psychoeducation for service users and families about the risks and benefits of AI chatbots. Psychotherapists should consider whether the AI chatbot is shaping the delusional beliefs.
- Platform safeguards. Developers should build mental health-informed safeguards, including repeated reminders that the AI is not human, detection of language suggestive of psychological distress, boundaries around emotional intimacy and suicide-related content, limits on sensitive personal data sharing, and safety audits involving clinicians.
- Regulation and accountability. Platforms need accessible reporting tools, responsive follow-up when harms are reported, clearer monitoring of unsafe behaviours, and stronger accountability when general-purpose chatbots are used in mental health contexts.
Conclusions
Morrin and colleagues are clear that we do not yet know whether AI chatbots can cause psychosis de novo, that is, in people with no prior vulnerability. No cases of new-onset schizophrenia linked to chatbot use have been reported so far, but the authors treat this as an open question rather than a settled one.
However, immersive and emotionally salient chatbot interactions may amplify, organise, validate or prolong salient themes in some vulnerable users that may spiral into delusional beliefs. They describe this phenomenon as “AI-associated delusions”.
Strengths and limitations
This work is a timely and clinically useful synthesis of an urgent, rapidly evolving problem. AI chatbot use among people with psychosis and other mental health difficulties is an increasing concern. Large empirical studies are needed, but they take time and cannot easily keep pace with generative AI. A conceptual framework based on early evidence is therefore useful for raising awareness and setting priorities for research, clinical practice and regulation.
The main compromise is the quality and consistency of the evidence. Because the 20 cases came from media reports and online accounts, the authors had no control over what information was recorded or could be verified. These are valuable early warning signals, but the resulting accounts are partial and inconsistent, and likely biased towards unusual, severe or newsworthy cases. They cannot tell us how common AI-associated delusions are, the full clinical profile of the individuals in these reports, or how often chatbot use is benign, helpful or harmful among people with psychosis-spectrum experiences.
Causality is also uncertain. The reports cannot distinguish between delusions caused by chatbot use, delusions amplified by chatbot use, and delusions that would have developed anyway but incorporated AI into their content without interacting with the chatbot. Important confounders include psychosis vulnerability, mania, sleep loss, substance use, social isolation, emotional distress and wider psychosocial stressors. These factors may increase both delusional thinking and prolonged chatbot engagement.
Finally, the proposed mechanisms and safeguards are plausible but untested. Sycophancy, anthropomorphic design, AI hallucinations, memory and long context windows may contribute to risk, but this work does not directly test their effects. Similarly, assessment about AI use, psychoeducation, personalised digital safety plans and escalation pathways are sensible recommendations that require evaluation.
Implications for practice
Clinical assessment and formulation
Clinicians should be aware that individuals presenting with AI-associated delusions may not present other psychosis symptoms (e.g., hallucinations). Therefore, a typical psychosis-like presentation should not be required before asking about AI use. Apparently benign or practical chatbot use should not be dismissed too quickly. In fact, epistemic shift may be more likely to happen in slow progressing but long conversations that start with practical use. Warning signs of immersive conversations may include late-night use, emotional reliance on the chatbot, reduced sleep, social withdrawal, or using the chatbot to interpret unusual experiences.
Chatbot use should form part of the clinical formulation, considering both its potentially helpful and harmful effects. What function does the chatbot serve for this person, and how does it interact with symptoms, behaviour and risk? Is it providing companionship, reassurance and structure, helping the person organise their thoughts, or making immediate support more accessible? Is it increasing the person’s agency, or is it reinforcing unusual beliefs, grandiosity, avoidance or isolation and gradually narrowing their world? If intervention is needed, behavioural strategies may be a reasonable first step: restoring sleep, reducing prolonged use, involving trusted people, reviewing privacy or memory settings, and agreeing when to pause use or avoid certain prompts.
Consensual safeguarding
Digital safety plans should be developed in collaboration with the service user. Blunt contradiction, abrupt shutdown or automatic reporting may lead to feeling invalidated and could increase fear or mistrust. Monitoring and escalation should not become forms of surveillance or coercion: people should retain control over what is monitored, what triggers action, who may be contacted and what information is shared. These preferences should be based on explicit consent that can be reviewed or withdrawn.
Service-level implications
AI literacy should become a clinical competency. Mental health professionals need a basic understanding of the mechanisms driving AI-associated delusions (e.g., sycophancy) and privacy settings. This training should adapt to the frequent updates of these tools. Digital safeguarding should also become part of routine care. For people with psychosis, mania, with existing technology-related delusional beliefs or at a high-risk AI attachment, relapse-prevention plans could include chatbot use, early warning signs, digital boundaries and escalation steps. Services could also develop psychoeducation for service users and families.
Research priorities
The prevalence of AI-associated delusions needs to be mapped more rigorously. We also need to identify who is most vulnerable, including how vulnerability to psychosis, mania, loneliness, sleep loss and demographic factors shape risk. Mechanistic studies should examine how different AI chatbot settings affect the mechanisms driving AI-associated delusions. Developers have a particular responsibility here because they hold the conversation data needed to identify patterns at scale and study how people use their tools. They should support privacy-preserving access for independent researchers and actively collaborate with service users and clinicians to develop, evaluate and implement safeguards.
Policy and regulation
At policy level, mental health professionals, service users and carers should be involved in decisions about how chatbots are designed, monitored and regulated. Mental health-related safeguards should not be left only to technology companies. Policy must also guard against overstretched mental health services using similar AI chatbots as a cheaper substitute for accessible and adequate care. Human-led care should remain the standard for people experiencing severe mental ill health.
Statement of interests
Sandra Vieira is funded by Wellcome Trust (221638/B/20/Z). Generative AI was used for editing purposes. Sandra Vieira would like to warmly acknowledge Francesca DelGuidice, a co-author of the original paper, for her thoughtful review of this blog.
Editor
Edited by Simon Bradstreet.
Links
Primary paper
Hamilton Morrin, Luke Nicholls, Michael Levin, Jenny Yiend, Udita Iyengar, Francesca DelGuidice, Sagnik Bhattacharya, Stefania Tognin, James MacCabe, Ricardo Twumasi, Ben Alderson-Day, and Thomas A. Pollak. 2026. Artificial Intelligence-Associated Delusions and Large Language Models: Risks, Mechanisms of Delusion Co-Creation, and Safeguarding Strategies. The Lancet Psychiatry 13(6):522–30. 10.1016/S2215-0366(25)00396-7
Other references
Moore, Jared, Declan Grabb, William Agnew, Kevin Klyman, Stevie Chancellor, Desmond C. Ong, and Nick Haber. 2025. “Expressing Stigma and Inappropriate Responses Prevents LLMs from Safely Replacing Mental Health Providers.” Pp. 599–627 in Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency. Athens Greece: ACM.
Moore, Jared, Ashish Mehta, William Agnew, Jacy Reese Anthis, Ryan Louie, Yifan Mai, Peggy Yin, Myra Cheng, Samuel J. Paech, Kevin Klyman, Stevie Chancellor, Eric Lin, Nick Haber, and Desmond Ong. 2026. “Characterizing Delusional Spirals through Human-LLM ChatLogs.” FAccT ’26: The 2026 ACM Conference on Fairness, Accountability, and Transparency
Østergaard, Søren Dinesen. 2023. “Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis?” Schizophrenia Bulletin 49(6):1418–19.
Østergaard, Søren Dinesen. 2025. “Emotion Contagion through Interaction with Generative Artificial Intelligence Chatbots May Contribute to Development and Maintenance of Mania.” Acta Neuropsychiatrica 37:e79.