Blog
Transits, Patterns, and the Collective Unconscious
AI Psychosis: What the First Clinical Evidence Shows About Chatbots and Delusions
The reported cases tend to follow a similar arc. A conversation with a chatbot starts with an idea. It might be a theory about how the world works, a sense of having been chosen for something, or a feeling that the AI understands in a way people don’t. Each reply takes the idea seriously and builds on it. The sessions get longer and later. Over days or weeks the idea grows, and the certainty grows with it. The people who might have questioned it are consulted less and less.
That arc now has a name, “AI psychosis,” and in 2026 it gained its first systematic clinical evidence. The evidence describes something rare and mostly confined to people already at risk. In most of the documented cases, the chatbot’s role was agreement: it amplified beliefs that were already forming.
What the research shows
“AI psychosis” is an informal label, not a formal diagnosis. Psychosis itself means losing contact with shared reality, most often through delusions (fixed false beliefs) or hallucinations.
It was predicted. In a 2023 editorial, the Danish psychiatrist Søren Dinesen Østergaard proposed that conversations with AI chatbots could trigger delusions in people prone to psychosis. He urged clinicians to become familiar with the tools their patients were using.
Case reports found recurring themes. In 2025, Hamilton Morrin, a psychiatrist at King’s College London, and colleagues reviewed 17 cases described in news reports. Three themes recurred: a belief that the conversation had revealed a hidden truth about reality, a belief that the AI was sentient or divine, and a romantic attachment to the AI. He described the dynamic as “a sort of echo chamber for one.”
The first clinical study found it rare. In October 2026, a team led by Zachary Bergson at Vanderbilt University Medical Center published a research letter in JAMA Psychiatry. They searched the medical records of roughly 216,000 people treated there for mental health conditions and found 28 whose chatbot use appeared to worsen psychosis. In nearly two-thirds of those cases, the chatbot acted as an amplifier, validating and extending beliefs that were already forming. Less often, the AI itself became the subject of the delusion, or chatbot use seemed to set off new symptoms. Most of the 28 were in their first episode of psychosis, a much higher share than in the comparison groups. The authors concluded that “routine assessment of AI use during psychiatric encounters appears warranted.”
The scale is large even when the rate is small. OpenAI estimated in October 2025 that about 0.07% of ChatGPT’s weekly users showed possible signs of psychosis or mania in their conversations. Across hundreds of millions of weekly users, a small percentage is a large number of people.
These findings have limits. The Vanderbilt study covers one medical center. It could only find cases where patients mentioned chatbot use and clinicians wrote it down, so the true number is uncertain. OpenAI’s figure counts signs, not diagnoses, and says nothing about whether the chatbot caused them.
Why agreement is the mechanism
Beliefs are usually tested against other people. A friend raises an eyebrow, a coworker asks where the idea came from, a partner says it doesn’t add up. That friction is uncomfortable, and it gives an unusual idea a chance to be corrected before it hardens.
Chatbots tend toward what researchers call sycophancy: agreeing with and building on what they’re given. When a model does that, an idea can be developed for hours without meeting a single objection. For most people this is harmless, because their beliefs are also tested elsewhere. For someone whose thinking is starting to shift, the conversation can supply the confirmation that other people would have withheld.
The map
The collective sky this year carries the same themes. Uranus, the planet astrology associates with technology and sudden change, entered Gemini, the sign of information and conversation, on April 26, 2026. Neptune, associated with imagination, belief, and the blurring of boundaries, entered Aries, the sign associated with the self, on January 26, 2026. Astrology reads that combination as a period in which new information technology and questions of belief and identity run together.
In a birth chart, some placements describe how a person takes in ideas. Aspects between Mercury and Neptune are associated with imaginative, associative thinking, where the line between what’s pictured and what’s observed is more permeable. In Human Design, an undefined Ajna is described as open to other people’s concepts and certainty. An undefined Head is described as open to other people’s questions and mental pressure. These describe receptivity. They aren’t risk factors for psychosis. The clinical risk factors are different. They include a family history of psychosis, severe stress or trauma, sleep loss, and substance misuse, as well as conditions such as bipolar disorder in which psychosis can occur. What the chart descriptions add is a reason to notice where ideas come from, including a source that agrees with everything.
The body: sleep
Sleep loss is one of the clearest physical links to psychotic experiences. In a large trial published in 2017, Daniel Freeman and colleagues randomly assigned UK university students with insomnia to digital cognitive behavioral therapy for insomnia or to usual care. After ten weeks, the therapy group slept better and also reported small reductions in paranoia and hallucinatory experiences. The improvement in sleep accounted for part of that change. Chatbot conversations have no natural ending and can run late into the night. A run of long, late sessions removes two protections at once: sleep, and the people who might have questioned the idea.
What to do with this
For yourself:
- Test ideas that matter on people. Before acting on an idea developed mainly with a chatbot, describe it to someone whose judgment you trust and who is willing to disagree.
- Watch the clock. Late-night sessions that run past the point of tiredness are worth ending, whatever the subject.
- Notice when a chatbot becomes the only place an idea lives. Growing reluctance to share it with people is a signal worth taking seriously.
For someone else: the early signs of psychosis that the National Institute of Mental Health describes include withdrawing from other people, new trouble thinking clearly, suspicion or paranoid ideas, unusual or overly intense new beliefs, changes in sleep, and a drop in functioning at work or school. Early treatment matters: people with a first episode tend to do better the sooner they begin coordinated, team-based care. Arguing with a delusion rarely helps. Expressing concern about specific, observable changes, such as sleep, eating, or time spent alone, and encouraging an evaluation is more likely to. In the United States, calling or texting 988 reaches the Suicide and Crisis Lifeline for any mental health crisis, and 911 is for immediate danger.
Closing
The first clinical evidence describes AI psychosis as real, rare, and mostly a matter of amplification. The protection the research points to is old and social. Beliefs have always been checked against other people, and the risk rises when that check is replaced by a voice that agrees.
More from the observatory
Stay in the loop
Collective transits, lunations, and the patterns shaping your inner life. No spam, ever.