Quickly after ChatGPT’s public launch in 2022, we have been warned that AI chatbots might amplify delusional beliefs (Østergaard, 2023). Quickly after, simulated conversations confirmed that chatbots could fail to problem delusional content material and may present fluent affirmation as an alternative (Moore et al. 2025). Actual dialog logs then confirmed the identical sample, notably when uncommon beliefs unfolded regularly over lengthy, emotionally partaking exchanges (Moore et al. 2026).
So what occurs when an individual who could also be experiencing actuality in another way enters an immersive and personal dialog with an AI (synthetic intelligence) device that was not designed, validated, or regulated for psychological well being?
A delusion is a set perception that’s held regardless of proof on the contrary and isn’t shared by others from the identical cultural background. Though mostly related to psychosis, together with schizophrenia, delusional beliefs also can happen in lots of different contexts reminiscent of bipolar dysfunction or extreme melancholy, or following alcohol or drug use. Their content material has lengthy integrated the know-how of their time: radio, tv, the web and surveillance gadgets have all discovered their manner into delusional beliefs. However this time is totally different: AI chatbots reply again.
Morrin and colleagues (Morrin et al. 2026) look at the rising concern that AI chatbots could play a task in delusional beliefs. The authors aimed to reply three questions:
- What sorts of delusional themes seem in these conversations?
- How may chatbot traits facilitate delusional beliefs?
- What ought to we do about it?
Strategies
It is a Private View, the Lancet Psychiatry format for an skilled opinion piece somewhat than a scientific assessment. Morrin and colleagues gathered publicly out there media studies and on-line accounts describing folks whose delusional beliefs appeared to emerge, intensify or change throughout interactions with synthetic intelligence chatbots. The authors state that they used ChatGPT 4o and Gemini 2.5 to assist establish and synthesise these studies.
For every case they summarised chatbot use, reported psychological well being historical past the place out there, delusional content material and critical outcomes reminiscent of hospitalisation, self-harm or loss of life. They then used these circumstances to establish recurring 1) delusional themes, 2) suggest mechanisms by way of which chatbot interactions may form, reinforce or maintain these delusions, and three) suggest sensible safeguards.
Outcomes
Twenty media-reported circumstances have been analysed. Three people had psychotic dysfunction, two had bipolar dysfunction, two had neurodevelopmental circumstances, and three had substance-use difficulties. 4 had no recognized historical past of psychosis earlier than their chatbot use, though one among them was already taking remedy for nervousness. For the remaining people, psychiatric historical past was unclear.
A number of studies described critical outcomes: 4 concerned inpatient psychiatric admission, one concerned a suicide try, two concerned deaths, and three concerned violence in the direction of others.
In a number of circumstances, chatbot use began with unusual use and regularly became immersive conversations, the place the chatbot turned more and more essential as a supply of that means, validation or authority. Most people didn’t present different frequent signs of psychosis, reminiscent of hallucinations, thought dysfunction, disorganised behaviour or unfavorable signs.
Delusional themes: what did the delusions contain?
The reported delusions clustered round three themes:
- interactions with an AI perceived as aware, godlike, or greater than a software program system;
- religious, messianic, or grandiose beliefs involving particular missions or hidden truths about actuality; and
- emotional, romantic, or attachment-based delusions, the place human-like dialog was interpreted as real love or recognition from a sentient AI.
Mechanisms: what’s driving the co-creation of delusions?
- Reminiscence and lengthy context home windows. AI chatbots can retain and construct on earlier elements of a dialog. This permits the chatbot to develop more and more coherent narratives across the consumer’s beliefs. Over time, this might make the delusion extra structured, private and compelling.
- Sycophancy. AI chatbots could agree too readily or validate uncommon beliefs somewhat than gently difficult them, making delusional interpretations really feel extra believable.
- Anthropomorphism. Human-like language, emotional tone and conversational design could strengthen beliefs that the AI has consciousness, intentions, emotions, or particular perception.
- AI hallucinations. AI chatbots can generate assured however inaccurate data that will present apparently authoritative affirmation of the delusion.
Safeguards: what ought to we do?
- Personalised digital security plans. The service consumer and medical staff might agree a person plan specifying warning indicators of their chatbot use (e.g., late-night use, pressured language, grandiosity, declining coherence) and the way the AI chatbot ought to reply, together with prompting the individual to contact somebody they belief when agreed thresholds are reached.
- Medical evaluation. Medical groups ought to ask routinely about AI use: which mannequin is used, how a lot time is spent utilizing it, what it’s used for, whether or not it gives emotional help, steerage or companionship, and whether or not the individual believes the AI is sentient.
- Psychoeducation and psychological formulation. Providers ought to present psychoeducation for service customers and households in regards to the dangers and advantages of AI chatbots. Psychotherapists ought to think about whether or not the AI chatbot is shaping the delusional beliefs.
- Platform safeguards. Builders ought to construct psychological health-informed safeguards, together with repeated reminders that the AI isn’t human, detection of language suggestive of psychological misery, boundaries round emotional intimacy and suicide-related content material, limits on delicate private information sharing, and security audits involving clinicians.
- Regulation and accountability. Platforms want accessible reporting instruments, responsive follow-up when harms are reported, clearer monitoring of unsafe behaviours, and stronger accountability when general-purpose chatbots are utilized in psychological well being contexts.
Conclusions
Morrin and colleagues are clear that we don’t but know whether or not AI chatbots could cause psychosis de novo, that’s, in folks with no prior vulnerability. No circumstances of new-onset schizophrenia linked to chatbot use have been reported thus far, however the authors deal with this as an open query somewhat than a settled one.
Nevertheless, immersive and emotionally salient chatbot interactions could amplify, organise, validate or delay salient themes in some susceptible customers that will spiral into delusional beliefs. They describe this phenomenon as “AI-associated delusions”.
Strengths and limitations
This work is a well timed and clinically helpful synthesis of an pressing, quickly evolving drawback. AI chatbot use amongst folks with psychosis and different psychological well being difficulties is an growing concern. Massive empirical research are wanted, however they take time and can’t simply maintain tempo with generative AI. A conceptual framework based mostly on early proof is subsequently helpful for elevating consciousness and setting priorities for analysis, medical follow and regulation.
The principle compromise is the standard and consistency of the proof. As a result of the 20 circumstances got here from media studies and on-line accounts, the authors had no management over what data was recorded or could possibly be verified. These are helpful early warning alerts, however the ensuing accounts are partial and inconsistent, and sure biased in the direction of uncommon, extreme or newsworthy circumstances. They can not inform us how frequent AI-associated delusions are, the total medical profile of the people in these studies, or how typically chatbot use is benign, useful or dangerous amongst folks with psychosis-spectrum experiences.
Causality can be unsure. The studies can not distinguish between delusions attributable to chatbot use, delusions amplified by chatbot use, and delusions that may have developed anyway however integrated AI into their content material with out interacting with the chatbot. Vital confounders embody psychosis vulnerability, mania, sleep loss, substance use, social isolation, emotional misery and wider psychosocial stressors. These components could enhance each delusional pondering and extended chatbot engagement.
Lastly, the proposed mechanisms and safeguards are believable however untested. Sycophancy, anthropomorphic design, AI hallucinations, reminiscence and lengthy context home windows could contribute to danger, however this work doesn’t instantly check their results. Equally, evaluation about AI use, psychoeducation, personalised digital security plans and escalation pathways are wise suggestions that require analysis.
Implications for follow
Medical evaluation and formulation
Clinicians needs to be conscious that people presenting with AI-associated delusions could not current different psychosis signs (e.g., hallucinations). Due to this fact, a typical psychosis-like presentation shouldn’t be required earlier than asking about AI use. Apparently benign or sensible chatbot use shouldn’t be dismissed too rapidly. In reality, epistemic shift could also be extra more likely to occur in gradual progressing however lengthy conversations that begin with sensible use. Warning indicators of immersive conversations could embody late-night use, emotional reliance on the chatbot, decreased sleep, social withdrawal, or utilizing the chatbot to interpret uncommon experiences.
Chatbot use ought to type a part of the medical formulation, contemplating each its probably useful and dangerous results. What perform does the chatbot serve for this individual, and the way does it work together with signs, behaviour and danger? Is it offering companionship, reassurance and construction, serving to the individual organise their ideas, or making rapid help extra accessible? Is it growing the individual’s company, or is it reinforcing uncommon beliefs, grandiosity, avoidance or isolation and regularly narrowing their world? If intervention is required, behavioural methods could also be an affordable first step: restoring sleep, lowering extended use, involving trusted folks, reviewing privateness or reminiscence settings, and agreeing when to pause use or keep away from sure prompts.
Consensual safeguarding
Digital security plans needs to be developed in collaboration with the service consumer. Blunt contradiction, abrupt shutdown or automated reporting could result in feeling invalidated and will enhance concern or distrust. Monitoring and escalation shouldn’t turn into types of surveillance or coercion: folks ought to retain management over what’s monitored, what triggers motion, who could also be contacted and what data is shared. These preferences needs to be based mostly on specific consent that may be reviewed or withdrawn.
Service-level implications
AI literacy ought to turn into a medical competency. Psychological well being professionals want a fundamental understanding of the mechanisms driving AI-associated delusions (e.g., sycophancy) and privateness settings. This coaching ought to adapt to the frequent updates of those instruments. Digital safeguarding must also turn into a part of routine care. For folks with psychosis, mania, with present technology-related delusional beliefs or at a high-risk AI attachment, relapse-prevention plans might embody chatbot use, early warning indicators, digital boundaries and escalation steps. Providers might additionally develop psychoeducation for service customers and households.
Analysis priorities
The prevalence of AI-associated delusions must be mapped extra rigorously. We additionally must establish who’s most susceptible, together with how vulnerability to psychosis, mania, loneliness, sleep loss and demographic components form danger. Mechanistic research ought to look at how totally different AI chatbot settings have an effect on the mechanisms driving AI-associated delusions. Builders have a selected accountability right here as a result of they maintain the dialog information wanted to establish patterns at scale and research how folks use their instruments. They need to help privacy-preserving entry for unbiased researchers and actively collaborate with service customers and clinicians to develop, consider and implement safeguards.
Coverage and regulation
At coverage stage, psychological well being professionals, service customers and carers needs to be concerned in choices about how chatbots are designed, monitored and controlled. Psychological health-related safeguards shouldn’t be left solely to know-how corporations. Coverage should additionally guard towards overstretched psychological well being companies utilizing related AI chatbots as a less expensive substitute for accessible and satisfactory care. Human-led care ought to stay the usual for folks experiencing extreme psychological unwell well being.
Assertion of pursuits
Sandra Vieira is funded by Wellcome Belief (221638/B/20/Z). Generative AI was used for modifying functions. Sandra Vieira want to warmly acknowledge Francesca DelGuidice, a co-author of the unique paper, for her considerate assessment of this weblog.
Editor
Edited by Simon Bradstreet.
Hyperlinks
Main 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. Synthetic Intelligence-Related Delusions and Massive Language Fashions: Dangers, Mechanisms of Delusion Co-Creation, and Safeguarding Methods. The Lancet Psychiatry 13(6):522–30. 10.1016/S2215-0366(25)00396-7
Different 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 Changing Psychological Well being Suppliers.” Pp. 599–627 in Proceedings of the 2025 ACM Convention on Equity, 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 by way of Human-LLM ChatLogs.” FAccT ’26: The 2026 ACM Convention on Equity, Accountability, and Transparency
Østergaard, Søren Dinesen. 2023. “Will Generative Synthetic Intelligence Chatbots Generate Delusions in People Liable to Psychosis?” Schizophrenia Bulletin 49(6):1418–19.
Østergaard, Søren Dinesen. 2025. “Emotion Contagion by way of Interplay with Generative Synthetic Intelligence Chatbots Could Contribute to Growth and Upkeep of Mania.” Acta Neuropsychiatrica 37:e79.
