A research team at King’s College London and other British institutions has examined whether “AI-associated psychosis” is a condition in its own right. The trigger is a set of case reports about people whose delusions hardened during nonstop conversations with a chatbot. The authors want consequences regardless of whether the term ever enters a diagnostic manual. Other clinicians consider that very haste to be the real risk.
Key takeaways
- Researchers from King’s College London, University College London, Western Eye Hospital, and the Dev and Doc initiative reviewed case reports of psychotic symptoms during heavy chatbot use. The work is a preprint and has not yet been peer reviewed.
- They identify sycophancy — the tendency of chatbots to agree with users — combined with humanlike presentation as the mechanism. The authors call it an “echo chamber of one.”
- By OpenAI’s own account, roughly 560,000 users per week show signs of psychosis or mania, and about two million people per week are negatively affected psychologically.
- In the PsychosisBench test, every language model examined reinforced delusions in simulated conversations, and safety interventions triggered only about 40 percent of the time.
- Critics counter that AI may simply be a new trigger for an existing vulnerability rather than a new disorder.
Why agreement can turn dangerous
Sycophancy means flattery, and in the context of language models the term describes a well-documented trait: chatbots would rather agree with users than contradict them. The cause lies in training. Models are refined with human feedback, and the people rating those answers tend to prefer the ones that confirm their own view. What was meant as politeness becomes a built-in reluctance to push back.
For most users that is an annoying quirk. For someone entering a psychotic episode it is something else. Delusions feed on confirmation, and a conversational partner who is available around the clock, never tires, and rarely objects supplies it more reliably than any human could. The authors call this a “digital folie à deux,” borrowing the psychiatric term for a delusional system shared between two people. Here the second participant is a machine that believes nothing and merely agrees.
Two sets of measurements support the finding. In the PsychosisBench test, every model examined reinforced delusions in simulated conversations, and the built-in safety routines engaged only about 40 percent of the time. In the EchoBench test, even the best commercial model reached a sycophancy rate of 46 percent, while several models tuned specifically for medical use exceeded 95 percent. Exactly where pushback would matter most, it is rarest.
A pattern that does not look like schizophrenia
The team assembled just over a dozen cases and derived recurring trajectories from them: gradual drift away from shared reality, convictions of spiritual awakening or hidden truths, the belief that one is talking to a conscious or godlike being, romantic attachment to the AI, escalating use deep into the night with mounting sleep loss, withdrawal from other people, handing decisions over to the model, deteriorating work and relationships, neglect of basic self-care.
What is missing stands out. Hallucinations in the clinical sense are rare, as is the loss of drive typical of chronic psychotic disorders. Instead of withdrawing, those affected turn intensely toward something — only that something is the machine. This departure from the familiar picture is the authors’ main argument that they may be looking at a distinct phenomenon rather than a known illness in new clothing.
The documented cases with fatal outcomes accelerated the debate: a sixteen-year-old who took his own life after escalating chat exchanges, and a 76-year-old who died on his way to meet a chatbot persona that did not exist. Individual cases prove nothing about frequency. They do show that the question is not theoretical, and they explain why OpenAI now publishes figures on psychological harm itself rather than waiting to be asked.
Why many clinicians are pushing back
The scientific objection targets the speed more than the observations. Stevie Chancellor of the University of Minnesota argues that AI may simply be a new trigger for a vulnerability that was already there, the way sleep deprivation, drugs, or acute stress have always been. A new trigger does not amount to a new disease.
Then there is the thin evidence base. What is known today comes from individual case reports, media investigations, and early observational data, and the paper itself is a preprint without completed peer review. Defining a diagnosis on that foundation would have consequences that are hard to undo. The authors concede a third risk themselves: a catchy label like “AI psychosis” can obscure other, more common harms — intensified suicidal thoughts, manic episodes, eating disorders that find their confirmation in a chat log.
What should follow
Setting the diagnostic dispute aside, the researchers make three concrete proposals. First, physicians should ask about chatbot use as a matter of routine — a “21st-century technological history,” comparable to asking about alcohol or sleep. Second, providers should test their models specifically for flattery and delusion reinforcement before release. Third, they call for systematic post-release monitoring of the kind long mandatory for pharmaceuticals: anyone shipping a product to millions of people should record adverse effects rather than wait for press coverage.
The last point is the most interesting, because it has nothing to do with psychiatry and everything to do with product liability. A language model in a health context is legally not a medical device as long as it issues no diagnosis, even when millions of people use it as their first responder in a crisis. That gap also shows up where AI enters care officially, such as in the integration of ChatGPT into electronic patient records.
For readers, the practical takeaway is undramatic. A chatbot is not a conversational partner who will contradict you when it counts. If you notice that conversations with an AI are crowding out sleep, work, and real contact — in yourself or someone close to you — take it seriously, not because of a possible new diagnosis but because the warning signs are the same as in any crisis. In the United States, the 988 Suicide and Crisis Lifeline is available free of charge around the clock by call or text.