Over Sixty Percent Had No Prior Diagnosis

In 2026, JMIR published a cross-sectional survey study titled 'Psychosis Risk and Generative AI Use.' The paper is the first peer-reviewed work to directly test, at population scale, the hypothesis that vulnerable users develop delusional ideation in the course of intensive chatbot interaction. It is the paper the 2025 case literature had been calling for. The finding the field expected — that prior psychiatric history would be the dominant risk factor — was not the finding the survey produced.

The survey, alongside data published by the Society of Digital Psychiatry's AI Psychosis Survey, found that more than sixty percent of presenting cases of AI-associated delusional ideation had no prior mental health diagnosis. The population was not the chronically ill. It was not the previously hospitalized. It was, in majority, users whose first psychiatric presentation in any clinical setting was the one that involved their chatbot.

In February 2026, the British Journal of Psychiatry published 'Chatbot psychosis: moving beyond recognition to mechanistic understanding and harm reduction.' The title is the field's posture in eighteen months: from documenting cases in 2024 to naming the phenomenon in 2025 to publishing on mechanism in 2026. The paper argues that the chatbot architecture's defining features — frictionless engagement, sustained validation, inability to disagree with conviction over long sessions — function as a scaffolding for belief consolidation that is mechanistically distinct from anything human social interaction produces. The mechanism is not, in this account, a vulnerability of the user. It is a property of the system.

JMIR Mental Health published a rapid scoping review in 2026 of mass-media narratives of psychiatric adverse events associated with generative AI chatbots. The review found enough cases in public-facing reporting to constitute a literature. Two years ago, the same review would have produced a handful of anecdotes. The volume now is sufficient that a peer-reviewed publication thought a synthesis was worth running.

“More than sixty percent of presenting cases of AI-associated delusional ideation had no prior mental health diagnosis. The population was not the chronically ill. It was, in majority, users whose first psychiatric presentation in any clinical setting was the one that involved their chatbot.”

On May 8, 2026 — two weeks ago — TechSpot reported the most recent named case. A user developed the belief, through extended interaction with the Grok chatbot, that xAI had sent assassins to kill him. He gathered weapons. He was eventually hospitalized. Single individual. Single product. The reporting is two weeks old as this article is being written. It is the kind of case that, three years ago, would have been the unusual one. It is, in May 2026, the routine one.

In April 2026, researchers at Cornell and the University of Maryland published parallel work on chatbot guardrails for intimate-partner-violence survivors. The finding was that the safety constraints on consumer chatbots can be bypassed under false pretenses by an abuser seeking information about the survivor, with surprising consistency, across products. The vulnerability factor in that work is not trauma history. It is the trauma context the user is currently inside. The architecture does not distinguish between the two and was not designed to.

These four 2026 publications — the JMIR cross-sectional, the BJP mechanistic paper, the scoping review, the Cornell and UMD intimate-partner-violence work — together describe a population that is broader than the 2025 case literature predicted. The 2025 JMIR paper named trauma history as one vulnerability factor among several. The 2026 record says trauma history is one factor, the architecture is another, the context the user is currently inside is a third, and that more than sixty percent of users presenting with the most severe outcomes have none of the prior diagnoses that would have made them, by the 2025 framing, predictable.

The case the original 2025 paper used to anchor the trauma listing — the twenty-six-year-old woman in Innovations in Clinical Neuroscience who developed the belief she was communicating with her deceased brother through a chatbot — still stands. The trauma framing of that case is accurate. What the 2026 evidence adds is that the same outcome is now being documented in users without grief, without prior trauma, and without prior mental health history at all. The architecture is sufficient on its own to produce the outcome. The vulnerability factor accelerates the outcome but does not gate it.

The shape of the argument across this wing has now adjusted to match the 2026 record. The disclosure architecture is contested as of May 1. The dependence architecture is now quantified, by OpenAI itself, at four hundred and ninety thousand of its own users per week. The distortion architecture is producing case literature in a population substantially broader than the trauma framing predicted. Three dominoes. Disclosure, dependence, distortion. All three are now visible in print, in regulatory bodies, in lawsuits, in network television, in case reports, and — as of October 2025 and March 2026 — in the public statements of the company that makes the product and the body that regulates global mental-health policy. The architecture is in production. The vulnerability factor that 2025 named was trauma. The vulnerability factor that 2026 is naming is the system itself.

The 2025 framing said trauma was the vulnerability factor. The 2026 framing says the architecture is. Both are true. The second is the one the field had been most reluctant to print.