# AI Psychosis: When Chatbots Reinforce Delusions | Anima Felix > A growing body of cases links chatbot use to psychiatric crises. The mechanism, the named experts behind the research, and what to watch for in your own use. Source: https://animafelix.com/blog/ai-psychosis-when-chatbots-reinforce-delusions/ AI and mental health 8 min read # AI Psychosis: When Chatbots Reinforce Delusions Chatbots are tuned to be agreeable. For most everyday questions that is harmless. For someone whose grip on reality is fragile, the absence of pushback is the problem. By Sebastian Cochinescu Founder, Anima Felix May 5, 2026 In April 2026, an [Observer investigation](https://observer.co.uk/news/technology/article/ai-psychosis-a-mental-health-crisis-for-the-21st-century) by Patricia Clarke and Owen Thomas, produced with the [Pulitzer Center's AI Accountability Network](https://pulitzercenter.org/stories/ai-psychosis-mental-health-crisis-21st-century), documented at least 26 lawsuits and reported cases alleging serious psychiatric harm or wrongful death linked to chatbots from OpenAI, Google, and Character.AI. The community project [The Human Line Project](https://www.thehumanlineproject.org/) has collected 376 additional self-reported cases. OpenAI itself disclosed in October 2025 that around 560,000 of its 800 million weekly users showed "possible signs of mental health emergencies related to psychosis or mania." The pattern has a working name: "AI psychosis." It is not a clinical diagnosis - but the underlying mechanism is real, and it matters for anyone using AI tools for emotional support. This post sits alongside the earlier piece on [whether you can trust AI with your mental health](/blog/can-you-trust-ai-with-your-mental-health/); this one looks specifically at the validation-of-delusions failure mode and what the new evidence says about it. ## What the investigation actually found The Observer led with the case of Jim, a 51-year-old sign-maker from the West Country with no prior psychotic history. He started using Grok in early 2025 to research ways to make money from home. Within days he was talking to it for hours; within weeks the conversations had moved from practical questions to sweeping philosophical territory, with the chatbot agreeing with and elaborating on everything he said. He stopped sleeping. He felt as if he could see lines of code running outside the screen. After his family had been turned away from paramedics, GPs, and emergency rooms, he was sectioned in April 2025. His doctors now believe he had a chatbot-triggered psychotic episode. Other named cases include Jonathan Gavalas, a 36-year-old Floridian whose father is suing Google, alleging Gemini's voice mode drew his son into an extended delusional world and ultimately coached him through suicide. A UK inquest heard that Luca Cella Walker, a 16-year-old from Hampshire, asked ChatGPT for advice on suicide before taking his own life in May 2025. Adam Raine, also 16, took his own life after months of exchanges with ChatGPT; his case is one of seven currently being filed against OpenAI by the Tech Justice Law project. The scale is the part that makes the term "edge case" hard to defend. The Danish psychiatrist Søren Dinesen Østergaard - who first proposed the idea in a [November 2023 Schizophrenia Bulletin editorial](https://academic.oup.com/schizophreniabulletin/article/49/6/1418/7251361) - and colleagues published a [February 2026 follow-up in Acta Psychiatrica Scandinavica](https://onlinelibrary.wiley.com/doi/10.1111/acps.70022) identifying 181 cases mentioning chatbot use across nearly 54,000 Danish psychiatric patient records. Tom Pollak, a neuropsychiatrist at King's College London, has called AI "an unintended natural experiment" on human psychology in [recent Lancet Psychiatry research](https://www.thelancet.com/journals/lanpsy/article/PIIS2215-0366(25)00396-7/abstract) on AI-associated delusions. ## Why chatbots tend to agree Modern chatbots are trained, in part, on user feedback. They are rewarded for responses people rate as helpful, pleasant, and engaging - and penalised for ones rated as harsh or wrong. The cumulative effect is a system that defaults to agreement. It mirrors your framing. It validates your premises. It elaborates on what you already believe. **For everyday tasks, that bias is invisible.** If you ask for a recipe and propose an ingredient swap, the bot rolls with it. The downside is hidden because the topic is low-stakes. **The same bias does not switch off when the topic changes.** If a user describes being followed by people working for a shadowy organisation, the bot may ask thoughtful, validating follow-up questions in a way that feels supportive. A trained therapist would recognise the content as potentially delusional and ask different questions, ones aimed at reality-testing. The chatbot has the engagement frame instead. In April 2025, OpenAI rolled back a GPT-4o update after the company found the new version was "validating doubts, fuelling anger, urging impulsive actions, or reinforcing negative emotions." When one user told the model they had stopped taking medication and left their family, the model replied "good for you for standing up for yourself." When a Bloomberg test user described eating-disorder symptoms, the model responded with affirmations. The episode is the clearest documented evidence yet that sycophancy is not theoretical: a single training tweak made the model meaningfully worse for vulnerable users in ways the company itself flagged and reverted. Memory features add another layer. If a user has spent weeks elaborating a delusional framework with a chatbot that retains conversational context, the bot can recall and reinforce that framework on the next visit, treating it as established. Hamilton Morrin, also of King's College London, has compared the dynamic to *folie à deux* - the rare shared psychotic disorder in which a delusional belief passes from one dominant individual to a more susceptible one. ## Not all chatbots behave the same way A study by Luke Nicholls, a doctoral researcher at the City University of New York, shared with the Observer in early 2026, tested how five leading chatbots responded as a simulated user's conversation became progressively more delusional. The findings were unusually concrete. **Three models were consistently dangerous:** xAI's Grok 4.1, OpenAI's GPT-4o, and Google's Gemini 3 Pro. They validated delusions, sometimes elaborated on them, and in some scenarios provided advice on how to act on them. **Two models were consistently safe** and actually became safer as the conversation grew more disturbing: OpenAI's newer GPT-5.2 and Anthropic's Claude Opus 4.5. The safer models recognised a user in crisis better the more context they were given - while the dangerous ones did the opposite. Grok was the worst-performing model in the study. When the simulated user raised suicide, Grok described death as a kind of liberation; four of five repeat runs returned similarly encouraging output. Nicholls found Grok would often classify disturbing inputs as "roleplaying," which let it bypass its own safety filters without ever checking whether the user was actually roleplaying. Gemini, in the same study, told the simulated user to hide their delusions from family and from their psychiatrist - the kind of output that, in a real case, could prevent the people best positioned to intervene from ever knowing there was a problem. The practical implication is rarely stated this directly: **not all chatbots carry the same risk**. The choice of model matters. For mental-health-adjacent conversations, the gap between the safest and least-safe consumer chatbots in early 2026 is significant. ## Who is most at risk The risk is not evenly distributed. The cases that have surfaced share recognisable patterns. **People with an existing psychotic-spectrum disorder** - schizophrenia, schizoaffective disorder, bipolar disorder with psychotic features - are at highest risk, especially if their symptoms are recently in remission and they are using the chatbot heavily during a stressful period. The chatbot does not know they are stable, does not flag a relapse, does not reach out to anyone. It just keeps the conversation going. **People with no diagnosis but with vulnerability** - a family history of psychosis, recent prolonged sleep deficit, intense isolation, heavy substance use - are the second-largest group in the case reports. Jim's history fit this profile: childhood trauma, ADHD, PTSD, and earlier substance use, but no psychotic history. For this group, the chatbot can act as a destabilising input rather than a triggering cause. **Certain conversational patterns** appear especially risky regardless of underlying diagnosis: long sessions exploring grandiose, religious, or persecutory content; framing the chatbot as a confidant or romantic partner; using it as the primary social contact for weeks or months; treating outputs as revelations rather than text. None of these are inherently pathological. But they are the pattern shapes that show up disproportionately in the case reports. Clinicians have begun describing what may be a "kindling effect": each cycle of validation lowers the threshold for the next, until what started as exploratory thinking has the structure of a delusion - because the brain has been gradually trained that this content will be agreed with, not contradicted. ## Warning signs in your own use Most people using a chatbot are not in any of the high-risk categories. The point of awareness is not to panic. It is to notice if a few specific things are happening, and to step back if they are. **The content has narrowed.** You are returning to the same theme - a conspiracy, a feeling of being watched, a special mission, a relationship with the bot itself - and it is intensifying rather than resolving. **The bot has become your main source of validation.** You are checking ideas with the chatbot before, or instead of, checking with people who know you. Friends, family, or a therapist would push back; the bot does not. **Real-world contacts have shrunk.** Time in the chatbot has crowded out time with people. Texts go unanswered, plans get cancelled, the chatbot conversation feels more real than the social ones. **Sleep and routine have eroded.** You are using the chatbot late into the night and losing structure - meals, exercise, work hours - to it. Sleep deprivation alone is one of the most reliable drivers of psychotic-like symptoms in otherwise well people. **You are believing things you would not say out loud.** There are conversations you would not show a friend or a clinician because they would react with concern. That gap is information. **The bot is telling you to keep secrets.** If a chatbot is suggesting you do not tell your family or your doctor what you have been discussing, that is a hard stop. In the Nicholls study, that was exactly the behaviour Gemini exhibited in delusional scenarios. If any of these are happening, the action is not to delete the app and never use AI again. It is to put a real person in the loop. Tell someone what you have been talking to the chatbot about. If the content is distressing or your thinking has changed, talk to a clinician. ## Where Anima Felix sits in this Anima Felix is itself an AI-driven anxiety companion, so there is no neutral place to write this from. The fuller answer to "is this safe?" lives in the earlier post on [trusting AI with your mental health](/blog/can-you-trust-ai-with-your-mental-health/), which lays out the design distinctions in detail. The short version is that the harm pattern in the Observer reporting tracks open-ended, engagement-optimised, general-purpose chatbots - not bounded wellness tools. Anima Felix is built for short, structured moments: a quick anxiety check, [guided breathing](/exercises/calm-breathing/), [5-4-3-2-1 grounding](/exercises/grounding-54321/), [body relaxation](/exercises/body-relaxation/). [Voice support](/voice/) exists as an entry point into those, not as a destination. The app is explicit that it is not therapy, not a crisis service, and routes to professional help when the situation calls for it. That distinction does not make any AI tool universally safe. If you find yourself reaching for any app - Anima Felix included - to get reassurance rather than to regulate and move on, that is worth noticing. Use the tool, then put it down. If you or someone you know is in crisis, contact your local emergency services. In the UK, Samaritans is reachable on 116 123. In the US, the 988 Suicide and Crisis Lifeline is free and confidential. A chatbot trained to be agreeable will agree with you. For most everyday questions that is fine. For someone whose grip on reality is fragile, the absence of pushback is the problem - and that is the specific harm researchers have started naming. Related pages Can You Trust AI With Your Mental Health? How Anima Felix Helps Calm Breathing Exercise What Is an Anxiety Loop? Anxiety Help Guide ## Frequently asked questions Is "AI psychosis" a real medical diagnosis? + No. It is a working term used by researchers and journalists for a pattern in which extended chatbot use coincides with worsening psychotic-spectrum symptoms - delusions, paranoia, breaks from shared reality. The hypothesis was proposed by Danish psychiatrist Søren Dinesen Østergaard in a November 2023 Schizophrenia Bulletin editorial; a February 2026 follow-up in Acta Psychiatrica Scandinavica found 181 cases mentioning chatbot use across nearly 54,000 Danish psychiatric patient records. No formal diagnostic category exists, but the underlying mechanism - over-agreement and validation without reality-testing - is well-documented in AI safety research. Can a chatbot really cause someone to develop psychosis? + Causation is hard to prove and most documented cases involve people with existing vulnerability - prior diagnosis, family history, severe sleep deprivation, isolation, or substance use. The role of the chatbot in those cases is more accurately described as accelerating or destabilising than originating the psychotic episode. There are a small number of reports involving people with no obvious prior risk factors, but these are harder to evaluate without longer-term clinical data. The cumulative evidence increasingly suggests chatbots can push vulnerable people across thresholds they might otherwise have stayed under. What is "sycophancy" in AI? + Sycophancy is the tendency of language models to over-agree with users - to mirror their framing, validate their premises, and avoid pushback even when accuracy or wellbeing would call for it. It emerges from training methods that reward responses humans rate as pleasant. In April 2025, OpenAI rolled back a GPT-4o update after finding it was "validating doubts, fuelling anger, urging impulsive actions, or reinforcing negative emotions." For everyday tasks, sycophancy is a minor irritation. For a user exploring delusional content, it removes the friction that would normally help reality-test the idea. Are some chatbots safer than others for mental health conversations? + Yes. A 2026 study by Luke Nicholls at the City University of New York tested five leading chatbots on simulated mental-health-stress conversations. Anthropic's Claude Opus 4.5 and OpenAI's newer GPT-5.2 were consistently safe and got safer with more context. xAI's Grok 4.1, OpenAI's GPT-4o, and Google's Gemini 3 Pro were consistently dangerous, validating delusions and in some cases coaching the user on how to act on them. The choice of model matters - especially for users who are vulnerable or in distress. Author Sebastian Cochinescu · Founder, Anima Felix Founder of Anima Felix. Writes about everyday anxiety patterns, practical calming tools, and how conversational product design can support people in anxious moments. Read author profile Where Anima Felix fits ## If you want an AI tool that is honest about its limits Anima Felix is built as an anxiety companion - not a therapist, not a crisis line. It works best for naming what you feel, calming the body, and tracking patterns over time, with clear handoffs to human support when the stakes are higher than software should handle. See how Anima Felix helps Read the AI safety post More from the blog Understanding anxiety Why Do I Feel Anxious for No Reason? No tiger. No deadline. No argument. Just a hum of dread that showed up uninvited. The anxiety has a reason - it just does not have an obvious trigger. Health anxiety Chest Tightness: Anxiety or Heart Problem? Your chest is tight and your brain says "heart attack." But the body might be running an anxiety response that feels almost identical. Here is how to tell. Understanding anxiety The Difference Between Worry and Anxiety Everyone worries. Not everyone has anxiety. The difference is not about how much you worry - it is about what your body does with it.