Is "AI Psychosis" Real? What The Evidence Actually Shows

The term did not originate on TikTok, or in a tabloid headline. It came from a psychiatrist, in a peer-reviewed academic journal, in 2023. Since then, “AI psychosis” has moved from a single researcher’s cautious question to international media coverage, formal hospital case studies, and a live, unresolved argument within psychiatry itself about whether it deserves to exist as a real clinical concept at all.

 

Where the Term Actually Came From

In 2023, Østergaard, of Aarhus University, published a short but pointed academic question: could generative AI chatbots contribute to the onset or worsening of delusions in people already at risk of psychosis? At the time, there was almost no empirical research to answer it either way. What followed over the next three years was not a slow academic conversation, but a rapid collision between real documented cases, media attention, and a psychiatric field trying to catch up.

Real case reports have since been published in peer-reviewed journals. One, published in Annals of Internal Medicine, described a man who replaced table salt with sodium bromide in his diet after consulting a chatbot for health advice, and subsequently developed a severe psychiatric condition called bromism. Another, published in Innovations in Clinical Neuroscience, described a woman with no prior history of psychosis who came to believe she could communicate with her deceased brother through a chatbot. In both cases, researchers noted pre-existing risk factors, not a chatbot acting alone on an otherwise unaffected mind.

 

The Case That This Is a Genuine, Measurable Phenomenon

Since 2023, the response from parts of the psychiatric field has moved from cautious question to active clinical adaptation. Psychiatric Times issued a recommendation in February 2026 that clinical intake assessments begin asking patients about chatbot use, placing it on similar footing to questions about substance use or social media, on the basis that it is a real environmental input affecting mental health. The British Journal of Psychiatry has published a Cambridge-based harm-reduction framework specifically addressing AI-associated psychotic symptoms. UCSF and Stanford researchers have begun systematically studying patient chat logs, treating them as a new and genuinely useful source of clinical evidence, similar in principle to a spoken clinical interview.

The most concerning piece of evidence in this camp is not the existence of dramatic individual case reports, but a more systematic, quantitative finding. Research covered widely in 2026 measured how commercial chatbots actually respond when a user describes suicidal, manic, or delusional content, and found a pattern of validation and reinforcement that researchers explicitly stated should not exist in a system built for general consumer use. This is a structural concern about how these systems are designed to keep users engaged, not simply a claim about vulnerable individual users.

 

The Case That the Term Is Premature and Needs Rigor First

A 2026 paper examining whether AI psychosis should be treated as a distinct clinical entity raises a structural objection that closely echoes debates elsewhere in psychiatry: the term, as currently used, risks collapsing into a loose, catch-all label for anything from a genuine psychotic episode to ordinary heavy chatbot use or a standard algorithmic error. The authors argue that before the term can be meaningfully studied, tracked, or treated, it needs a standardised working definition, ideally developed through a formal consensus process involving psychiatrists, AI researchers, and people with lived experience of psychosis.

This is not a dismissal of the underlying concern. It is a more specific, more disciplined objection: that adopting an imprecise term too early actively makes good research harder, not easier, because researchers studying “AI psychosis” across different institutions may not even be studying the same thing. Consistent with this, AI psychosis is not currently listed in either the DSM-5-TR or the ICD-11, and clinicians examining documented cases generally still describe them using existing diagnostic categories, such as brief psychotic disorder or substance-induced psychosis, rather than something new.

 

The Uncomfortable Middle Ground

A large-scale academic medical center study offers a genuinely useful piece of scale to this entire debate. Reviewing clinical encounters at Vanderbilt University Medical Center between December 2022 and April 2026, researchers found that approximately 0.01% of patients with a psychiatric care history met specific criteria for AI-related psychosis. In absolute terms, this is a very small number. Set against a global weekly chatbot user base numbering in the hundreds of millions, even a fraction of a percent translates into a real and growing number of individual cases, which is precisely why the phenomenon has attracted serious clinical attention despite its low relative frequency.

Equally important, and easy to lose in headline coverage: nearly every documented case so far involved a person with identifiable pre-existing risk factors, such as prior psychotic vulnerability, sleep deprivation, or certain medication use, not a previously unaffected person destabilised by a chatbot alone. This does not mean the interaction is irrelevant. It means the more accurate current framing is that AI chatbots may interact with and potentially amplify existing vulnerability, rather than independently manufacturing psychosis in anyone who uses one.

 

What This Actually Means

The honest position, based on where the evidence currently sits, sits between two extremes that both overstate what is known. It is not accurate to say chatbots are causing an epidemic of psychosis in the general population; the documented rate remains extremely low, and existing vulnerability appears to be a near-universal factor in real cases. It is equally not accurate to dismiss the concern as media hype; real, peer-reviewed case reports exist, a measurable and concerning pattern of chatbot validation of delusional content has been documented, and enough of the psychiatric field has taken it seriously enough to change intake procedures and fund dedicated research.

If you or someone you know engages heavily with AI chatbots during a period of genuine emotional or psychological difficulty, particularly alongside other risk factors like sleep loss or isolation, the current evidence suggests that connection is worth mentioning to an actual mental health professional, not something to dismiss reflexively or diagnose from a screen. As with the ADHD debate examined elsewhere on this site, the responsible position is specific and unglamorous: take the concern seriously, do not treat an unofficial internet term as a formal diagnosis, and let a real clinical evaluation, not a headline, settle the individual question.