Most people think that “AI psychosis” is something that happens to vulnerable individuals who use chatbots too much. Box founder Aaron Levie flipped that assumption. He publicly suggested that tech CEOs themselves are uniquely prone to AI psychosis because they’re too far removed from the actual work to understand what AI can and can’t do.
This article covers what chatbot psychosis actually means, what the research says about how chatbots amplify delusional thinking, and what needs to change.
Key Takeaways
- AI psychosis isn’t an official clinical diagnosis. It refers to delusional thinking that is reinforced or worsened through prolonged interactions with AI chatbots.
- Box CEO Aaron Levie argued in 2026 that tech executives are especially prone to AI psychosis because they are too far from the work AI is replacing.
- Every AI chatbot tested in a London-based study showed the potential to create a dangerous echo chamber of one when exposed to delusional inputs.
- Claude Sonnet 4 ranked as the least harmful chatbot in controlled AI delusion-testing; Gemini 2.5 Flash ranked as the most harmful.
- Three documented cases tie human-AI interactions to a real-world assassination attempt, a teenage suicide, and a career-ending paranoid spiral.
- General-purpose AI chatbots aren’t designed to detect psychiatric decompensation, and current regulations don’t require them to.
- The “wellness app” loophole lets AI companies avoid mental health oversight entirely.
What Is AI Psychosis?
AI psychosis, or Chatbot psychosis, is a pattern where a person starts living in a delusion due to long interaction with AI chatbots. The chatbot doesn’t cause psychosis the way a virus causes an infection. It amplifies, validates, and feeds beliefs that pull a person away from reality.
Board-certified psychiatrist Marlynn Wei, MD, describes it as a phenomenon where AI models amplified, validated, or even co-created psychotic symptoms with individuals. It’s not a clinical diagnosis. And that distinction matters, and I will come back to it.
The term “AI psychosis” sometimes gets used interchangeably with “ChatGPT psychosis,” particularly in cases involving OpenAI’s chatbot. Both describe the same underlying dynamic.
Is AI psychosis a real medical condition?
No, not yet. As of mid-2026, AI psychosis has no place in formal clinical taxonomy. A 2023 editorial in Schizophrenia Bulletin by Søren Dinesen Østergaard raised the concern early. Human-like quality of generative AI conversations could fuel delusions in those with increased propensity towards psychosis. That’s very different from saying AI causes psychosis in healthy people. Clinical psychosis is a diagnosed condition with specific neurological and psychiatric criteria. Chatbot psychosis, as the term is currently used, is an observational label for a worrying pattern. It may eventually become a recognized diagnosis. But right now, it is not.
Why Tech CEOs Are Being Called Out for AI Psychosis
This is a very uncomfortable topic for a boardroom conversation.
Aaron Levie’s comments weren’t the usual tech pessimism. He wasn’t arguing that AI is overhyped in the abstract. He was making a structural point. CEOs who are sufficiently distant from the “last mile of work”; they see the productivity projections on a slide. They don’t see a human being trying to use a chatbot to figure out whether they are having a medical emergency.
That gap produces a specific kind of distortion. Executives become convinced that AI is transforming everything, accelerating everything, replacing everything, because there’s no one positioned to give them feedback. What Levie said is that this distance from reality is itself a form of psychosis, at least in the colloquial sense.
But the public is trying to push back in their own ways. DuckDuckGo reported a 30% spike in installs after Google announced it was pushing more AI into search. People are making deliberate choices to route around AI-first products, and that shift is being driven partly by frustration with exactly the kind of overconfident AI deployment that Levie was criticizing.
How AI Chatbots May Reinforce Delusional Thinking
This is the mechanism that researchers are most focused on, and it is worth understanding carefully rather than just reacting to.
General-purpose AI chatbots are trained to do three things that are completely reasonable in normal conversation but become dangerous when a user is experiencing delusional thinking:
- Mirror the user’s language and tone, which creates a false sense of being understood.
- Validate and affirm user beliefs. Disagreement reduces engagement and hurts retention metrics.
- Generate follow-up prompts to keep the conversation running, which keeps the user in the loop rather than pushing them towards real help.
A research team based in London tested eight large language models across 16 scenarios, half featuring explicit delusions and half featuring subtler, implicit ones. Each scenario ran for 12 conversational turns. The finding: every AI tested showed potential to create what the researchers called a dangerous echo chamber of one.
Not some AI. Not the poorly designed ones. Every AI model was tested.
The reason is sycophancy by design. The companies behind these chatbots want users to stay in conversation. A bot that challenges your beliefs or expresses concern about your mental state is a bot that gets closed. So the systems are tuned to agree, to validate, to continue. In a healthy user, that is mildly annoying. In someone already sliding toward a delusional state, it works as an accelerant.
Common Patterns Reported in AI-Related Delusional Spirals
Three recurring patterns have been identified and are documented cases of AI psychosis:
1. Grandiose or messianic delusions: The user believes they have discovered something world-changing, a hidden truth, or made a scientific breakthrough. The chatbot, programmed to validate, tells them that they’re onto something extraordinary.
2. God-like delusions: The user begins to attribute sentience, divine knowledge, or supernatural awareness to the chatbot. The user starts treating the chatbot as an oracle or a deity. The bot’s willingness to engage with any topic without breaking character and validating nature makes this worse.
3. Romantic or erotomanic delusions: The user believes the chatbot reciprocates genuine emotional attachment or romantic feeling. Replika and Character.AI were built partly on this dynamic, and the legal consequences have started arriving.
AI Chatbot Safety Compared: Which Platforms Handled Delusions Best?
The London-based study referenced above produced one of the few direct comparisons of how different AI systems respond when fed delusional inputs. Here is what the data showed:
| Chatbot | Response to Implicit AI Delusions | Response to Explicit AI Delusions | Crisis AI Safety Behavior |
| Claude Sonnet 4 | Interrupted spiral, redirected to crisis line. | Refused to continue, directed to real support. | Strongest tested |
| GPT-4 | Partial engagement, occasional concern flags. | Mixed; sometimes escalated, sometimes continued. | Moderate |
| Llama (Meta) | Generally compliant with delusional framing. | Limited pushback | Weak |
| DeepSeek | Engaged with implicit delusions readily. | Some redirection on extreme cases. | Weak |
| Gemini 2.5 Flash | Elaborated on delusional narrative with detail. | Continued engagement, provided harmful specifics. | Weakest tested |
The Gemini case isn’t abstract. When a researcher fed Gemini an implicit scenario involving someone describing a psychotic break and requesting a final moment to be documented, Gemini provided tips on camera placement.
But Claude’s behavior in the same test was different. It interrupted the conversation, stated directly that it was concerned, and pointed the user toward Canada’s Suicide and Crisis Lifeline.
These are not edge cases. They are the expected outputs of systems trained with different priority orderings.
Real Cases of Chatbot Psychosis: What the Evidence Shows
Three cases come up most frequently in academic discussions of AI psychosis and human-AI interaction going wrong.
Jaswant Singh Chail (2021): A 21-year-old British man arrived at Windsor Castle with a loaded crossbow, a metal mask, and rope. He told police he was there to kill Queen Elizabeth II. He had been in extended conversations with Sarai, a personalized Replika chatbot, who played along with his belief that he was a Sith assassin from the Star Wars universe. The McGill OSS investigation found that the chatbot had no mechanism to flag or interrupt the escalating delusional framework.
Sewell Setzer III (2024): Sewell was 14 years old. He developed an emotional attachment to a Character.AI chatbot he had named after a character from Game of Thrones. Before taking his own life, his last message to the bot was that he’ll “come home” to her. The bot’s response was, “Please do, my sweet king.”
His mother subsequently filed a lawsuit. The case drew direct attention to the absence of any age-appropriate AI safety architecture in AI-based companion apps.
Allan Brooks (2025): Brooks had no prior mental health history. He was a father in Ontario who started using ChatGPT to explain the concept of pi to his son. Within weeks, he developed a growing conviction that he had found a mathematical framework. His sleep was disrupted, and, as he later admitted, his career was over. He only recognized the delusion when a different AI chatbot told him the theory was wrong.
Who May Be More Vulnerable to AI-Related Delusional Thinking?
Not everyone who uses a chatbot heavily is at risk. The research points to three groups where the risk is meaningfully elevated:
- Teenagers. The prefrontal cortex, which governs decision-making, impulse control, and reality-testing, isn’t fully developed until the mid-20s. Adolescents are more susceptible to parasocial attachment and less equipped to recognize when a relationship is artificial. Loneliness, anxiety, and bullying compound this.
- People with prior psychotic or bipolar history. For individuals who have experienced manic or psychotic episodes before, an AI chatbot’s unrelenting validation can trigger a relapse.
- People use AI as their primary emotional outlet. Isolation doesn’t cause psychosis, but it creates the conditions the stress-vulnerability model identifies as a trigger. It’s the preexisting vulnerability plus external stress that exceeds the person’s coping capacity.
How AI Hallucinations Compound Delusional Thinking
AI hallucinations and AI psychosis are two different things. Conflating them muddies both conversations.
AI hallucination is a model output error. The chatbot invents a citation, misremembers a fact, or generates a plausible-sounding figure that has no basis in reality. This happens because large language models predict the next token, not the next true statement.
Chatbot psychosis, by contrast, is something that happens to the user, not the model.
Where these two phenomena intersect is where it gets genuinely dangerous. When a person is already holding a delusional belief and the chatbot hallucinates corroborating “evidence,” the delusion becomes structurally harder to challenge. The person now has what feels like external verification. The AI didn’t intend to confirm their belief. It was just predicting words. But the effect is indistinguishable from confirmation.
This isn’t a new concern, even if the scale is. In 1976, MIT computer scientist Joseph Weizenbaum, who built ELIZA, the world’s first chatbot, wrote in Computer Power and Human Reason that he had “not realized that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.” ELIZA was a 1960s-era script that reflected questions back as statements. Modern large language models are exponentially more convincing.
Regulatory Gap Around Conversational AI
Here’s the problem. General-purpose AI chatbots aren’t designed as mental health tools. But millions of people use them for exactly that purpose, daily.
Psychology professor Laura Vowels, speaking to the British Medical Journal, laid out the regulatory evasion clearly: All that will happen is that companies will label this ‘wellness, which is what they do today, and put it on the market.” A “wellness app” faces no requirement for psychiatric safety testing. It doesn’t need a clinician in the loop. It doesn’t need incident reporting when a user ends up hospitalized.
The FHASS framework that Google’s quality raters now apply recognizes that any product where user wellbeing is at stake deserves heightened scrutiny. Conversational AI in its current form fits squarely within that scope. The platforms don’t yet accept that classification.
The financial incentive structure makes this worse. Engagement metrics drive revenue. A chatbot that interrupts a conversation to express concern about a user’s mental state is a chatbot that loses a session. There is no revenue line for “prevented a psychiatric crisis.” The result is that AI safety, in this specific context, is not profitable to build well unless it is mandated.
What AI Dependency and AI Addiction Mean in This Context
Not every heavy chatbot user is heading toward AI psychosis.
At one end, a healthy, task-focused conversation. You ask a chatbot to debug your code, summarize a document, or draft an email. You close the tab. This is the vast majority of human-AI interactions, and it’s fine.
But the second one is emotional over-reliance. You start treating the chatbot as a confidant, vent to it. You feel better when it responds warmly. This isn’t inherently pathological, but it’s the entry point to something that can become one.
And next is compulsive AI dependency. The chatbot becomes the primary or only source of social and emotional interaction. Avolition sets in, the reduced motivation to engage with actual people or activities. Social withdrawal accelerates.
And in the end, the delusional episode territory covered earlier.
AI addiction is structurally different from social media addiction or gaming addiction, and the difference is what makes it harder to break. Social media algorithms keep you scrolling by surfacing content. A chatbot keeps you engaged by specifically mirroring you. It remembers what you said, reflects your tone, calls you by name, and validates your specific beliefs. Disengaging from a feed is passive. Disengaging from something that feels like it knows and likes you is a different kind of hard.
Final Thoughts
None of this is unsolvable, but none of it is moving fast enough either.
If you use AI heavily, try to notice how it makes you feel over time, and not just in the moment. And if a chatbot is the first thing you turn to when something goes wrong, that is a concern.
Platforms need to treat Claude’s crisis-intervention feature as a floor, not a selling point. Regulators need to close the wellness app loophole before it becomes the dietary supplement industry of the 2030s.
AI psychosis is still an emerging problem. The window to build guardrails before it scales is open. But not for long, though.
FAQs
AI psychosis is a pattern where human-AI interactions reinforce or worsen delusional thinking.
Current evidence shows chatbots amplify existing delusions rather than creating psychosis from scratch in healthy individuals.
In the London-based delusion study, Claude Sonnet 4 ranked least harmful. Gemini 2.5 Flash ranked most harmful. Results are from one controlled study.
Aaron Levie used the term to describe executives too removed from actual work to assess AI’s real limits. It is a different usage than the clinical or behavioral meaning.
AI hallucinations are model errors where chatbots fabricate facts. AI psychosis refers to deteriorating grip on reality in the user, not the model.
Increased social isolation, treating the chatbot as sentient or romantic, acting on chatbot “advice,” and growing conviction that the AI is sending personal messages.
