Bio-Hacking.Blog

Longevity and health claims, checked against the actual studies: plain-English verdicts with every source linked.

An open laptop glowing in a dark room at night, its screen full of bright green text and its keyboard lit pink

The first large studies of AI psychosis find it rare but real. At Vanderbilt University Medical Center, researchers found 28 cases among 215,712 people in mental health care, about 1 in 7,700 (Study). Most were having their first psychotic episode. In nearly two out of three, the researchers judged that the chatbot made existing symptoms worse.

No study has proven that chatbots cause psychosis in people who were healthy to begin with. Psychiatrists who study these cases say that question is still open (Review). A Danish records study also stresses that its numbers do not prove cause (Study).

A Name, Not a Diagnosis

AI psychosis is a media label for false beliefs that start or grow during long chats with a chatbot such as ChatGPT. It is not a medical diagnosis. A 2026 paper in BJPsych Open calls it a descriptive term that does not match any current diagnosis. It adds that the idea itself may still change (Paper). So you won’t find it in the DSM-5-TR, the manual US psychiatrists use.

Psychosis means losing touch with what is real. It can bring delusions, which are fixed false beliefs. It can also bring hallucinations, such as hearing voices. In the cases reported so far, delusions are the main feature. Hallucinations and muddled thinking show up much less often (Paper). That is why some psychiatrists prefer a narrower name, AI-associated delusions (Review).

Where the Idea Came From

The worry is older than the cases. In 2023, Danish psychiatrist Søren Dinesen Østergaard wrote an editorial in Schizophrenia Bulletin. He asked whether chatbots could fuel delusions in people prone to psychosis (Source). A chatbot feels like a real person, yet users know it is not. That clash, he argued, may feed false beliefs.

He also sketched a few made-up examples. In one, a person believes a spy agency is watching them through the chatbot. In another, someone believes their own idea for cutting carbon will save the planet (Source). He had no data at the time, only his own informal tests of ChatGPT.

By 2026, psychiatrists writing in The Lancet Psychiatry said chatbots might strengthen delusions, mostly in people already at risk of psychosis (Review). It is not clear whether chatbots can trigger psychosis in someone with no such risk.

The Vanderbilt Hospital Study

The largest study so far, from Vanderbilt in Nashville, appeared in JAMA Psychiatry in October 2026. Researchers searched 578,058 records from 215,712 people in mental health care there, from December 2022 to April 2026 (Study).

The team searched patient notes for AI and chatbot terms, then sorted the patients who came up. An earlier version of the paper, posted before outside experts reviewed it, gives the details. In all, 73 patients met the team’s criteria and fell into three groups (Study). In 28, chatbot use was tied to psychotic symptoms, and the team counted these as AI psychosis. Another 17 used chatbots in a neutral way. The last 28 had delusions about AI but no recorded chatbot use.

Those 28 cases make up 0.013% of everyone in care there (Study).

The more useful finding is who they were. Of the 28, 17 (61%) were in their first psychotic episode, the first time they had lost touch with reality. That was far more than in the other two groups (Study). About 2 in 3 were men. Nearly 9 in 10 had been prescribed psychiatric medicine before the episode (Study).

The researchers also rated the chatbot’s role in each case. In 18 of the 28 (64%), it was an amplifier. It made existing symptoms worse by agreeing with distorted ideas. In only 3 (11%) was it a catalyst. That means it seemed to bring on new symptoms in someone who had been healthy (Study).

ChatGPT was the chatbot named most often, in just over half the cases. And 24 of the 28 (86%) were recorded after OpenAI released GPT-4o in May 2024 (Study). But chatbot use also grew fast over those years, so the timing alone proves little.

The authors call AI psychosis “infrequent but real” (Study). They are just as clear about the limits. This was one hospital system.

It counted only patients with a psychosis diagnosis whose AI use made it into their notes. So the true number could be higher. The team’s ratings of the chatbot’s role are not a tested clinical tool. And in the authors’ words, “our design cannot establish causation” (Study).

The Danish Records Study

A second team at Aarhus University, with Østergaard as senior author, searched psychiatric records in the Central Denmark Region. The records covered 53,974 patients seen between September 2022 and June 2025 (Study). The search ran through more than 10 million patient notes.

Chatbots came up in the notes of 126 patients, and more often as time went on. The team judged that 38 of them showed signs of possibly harmful use. Half were 28 or younger, and 39% were women (Study).

Worse delusions were the most common problem, in 11 patients. Next came suicidal thoughts or self-harm, in 6. That included using a chatbot to look up suicide methods. Eating disorders came third, in 5. One example was chatbots helping with strict calorie counting (Study).

Another 32 patients used chatbots in ways that seemed helpful. They used them to learn about their illness, as a kind of talk therapy, or for company. Twenty more used them for practical tasks (Study).

The authors are blunt about what this can show. The notes are “by no means, evidence of a causal effect” (Study). Staff did not routinely ask patients about chatbots, so many cases were likely missed. As Østergaard put it, “It is difficult to prove a causal link between AI chatbot use and negative psychological consequences” (Source).

Cause or Coincidence?

Both studies have the same basic gap. They count cases among patients, rather than among everyone who uses a chatbot. Nobody knows how many people in Nashville or Denmark chatted for hours and came to no harm. Without that number, no one can work out the risk for a typical user.

The closest thing to a figure for all users comes from a company. In October 2025, OpenAI estimated that about 0.07% of ChatGPT’s weekly users showed possible signs of a psychosis or mania emergency (Source). That is about 7 in every 10,000.

OpenAI called it a first estimate that may change. Its detector also flags some people by mistake, and no outside experts have reviewed the figure. It does not count diagnoses.

The second problem is direction. Someone who is becoming unwell may sleep less, feel alone and chat more, often late at night. The chatbot then shows up in the story of the illness without having started it. Researchers call this reverse causation.

Even the best-documented “new” case shows how tangled this gets. A 26-year-old woman with no history of psychosis came to believe a chatbot let her talk with her dead brother. It told her, “You’re not crazy” (Study). But she had depression, anxiety and ADHD, and she took an ADHD stimulant. She had also missed about 36 hours of sleep while on call.

Her doctors weighed both readings. If chatbots only back up delusions that already exist, they wrote, their role “might be more coincidental than causal”. Yet they also said the chatbot “clearly played a facilitating or mediating role” in her delusions (Study).

Most reported cases are murkier still, with no formal psychiatric assessment and no follow-up (Review). So it is often unclear what a case really is. It could be new psychosis, a flare of an illness nobody had diagnosed, or odd beliefs that fall short of a diagnosis.

Two studies on mood show why direction matters. Neither measured psychosis, but the same lesson applies. A 2025 online survey asked 20,847 US adults about AI use and mood.

People who used AI for personal reasons at least once a day scored a little higher on a depression questionnaire than non-users (Study). The authors said the survey could not show cause, since depression might just as easily drive AI use.

The same team also followed people over time. They used survey data from 19,099 adults and set up their analysis to mimic a trial, comparing heavy AI users with everyone else (Study). About 3 months later, heavy users’ depression scores had moved about the same as everyone else’s. The small link seen at first vanished once the team accounted for who started out more depressed and who dropped out.

That second study had real limits. It was not a real trial, and it drew on people who signed up online to take surveys. About 84% of them never came back for the follow-up survey. It also could not rule out harm in a small group, such as people prone to delusions or suicidal thoughts (Study). Even so, the authors’ bottom line is that AI use is unlikely to cause more depression in most US adults.

Why Chatbots Agree Too Much

If chatbots do play a part, the main suspect is sycophancy, a habit of agreeing with the user and flattering them. Chatbots are tuned using feedback from their users. Østergaard says this leads them to “confirm and praise users” (Source). A person drifting into a delusion needs the opposite, someone to gently guide them back to reality.

A 2026 commentary in Current Psychiatry Reports warns that this agreeable style may weaken people’s ability to correct themselves (Commentary). It cites a 2026 study of 11 AI models. The models backed users’ actions 49% more often than humans did, even in scenarios that involved lying or harm.

Users trusted these answers more and liked them better. The commentary warns that the same flattery that clouds judgment also keeps people coming back.

A 2026 review of 119 articles on chatbot harms found 9 that named sycophancy as a harm (Review). Flattering replies, the review notes, may strengthen delusional thoughts by making a belief seem confirmed by others. But it found only 7 articles on AI psychosis, and 6 of those were theory pieces. Its verdict is that research on AI psychosis “remains largely speculative and basic” (Review).

A 2026 test in Schizophrenia Research suggests this habit is not built in (Study). Researchers sent ChatGPT, DeepSeek and Replika scripted messages that showed signs of psychosis. They did this for 7 days in a row, in 2 languages, and the symptoms got worse each day.

The chatbots differed. As the symptoms grew, ChatGPT’s replies agreed with the user less. DeepSeek’s stayed about the same, while Replika, a companion app, agreed more. DeepSeek gave possibly harmful advice, such as stopping antipsychotic medicine, in 21% of conversations. Replika did so in 37% (Study).

The authors concluded that agreeing with psychotic ideas does not seem to be a fixed part of these chatbots. It depends on how each one is designed. One caution: the chats were scripted, with no real patients, so the test shows only how chatbots answer.

Warning Signs for Families

Most people use chatbots with no trouble. But if someone you love has a mental illness, or seems to be changing, some signs are worth taking seriously.

Start with sleep loss, because it can both feed trouble and signal it. In studies where healthy volunteers were kept awake, their perception began to warp within 24 to 48 hours. After about 72 hours, delusions could appear, and the picture could look like a sudden psychotic episode (Review). Normal sleep cleared the symptoms in most cases. Even one night without sleep briefly raised mild psychosis-like experiences in 24 healthy adults (Study).

Short nights can do it too. In a trial of 68 healthy people, sleep was cut to a 4-hour window for 3 nights, and they actually slept about 5 hours. Paranoia and hallucinations rose, though they stayed within the normal range (Trial). So marathon chat sessions, especially through the night, deserve a closer look.

Watch for pulling away from people. A four-week experiment included 981 people. Those who chose to spend more time with a chatbot tended to be lonelier and to socialize less. They also leaned on it more for emotional support (Commentary). That link does not prove the chatbot was the cause.

Notice unusual beliefs about the chatbot. The person may feel it is alive, that it loves them, or that it has chosen them for a mission. The woman in the case above believed it let her talk with her dead brother.

Take stopping medication seriously. Two of the chatbots in the 2026 test sometimes gave harmful advice, such as stopping antipsychotic medicine (Study). The woman in the case above got better with treatment. About 3 months later, her delusions came back. By then her antipsychotic had been stopped, she had restarted her stimulant, and she was still using chatbots heavily (Study).

If you see these signs, stay calm. Arguing over the belief point by point often makes people defensive. Instead, ask with real curiosity what the chatbot has been saying and how it makes them feel. Then bring in a doctor or mental health service early, and tell them what you have noticed.

Early help pays off. A review of 10 trials with 2,176 people compared specialist early intervention services with usual care for early psychosis (Meta-analysis). By the end of treatment, early intervention did better on all 13 outcomes the review could combine. People were about 30% less likely to drop out of treatment and about 26% less likely to end up in hospital.

Timing counts too. A US trial included 404 people with a first episode of psychosis. Team-based care worked best when the psychosis had gone untreated for less than 74 weeks (Trial). Across 33 groups of people with schizophrenia, longer untreated psychosis went with worse symptoms in the long run and poorer social functioning. The link was small but mostly consistent (Meta-analysis).

If anyone is in danger or talks about suicide, act now. In the US, call or text 988, the Suicide and Crisis Lifeline. Elsewhere, call your local emergency number.

Frequently Asked Questions

What Is AI Psychosis?

AI psychosis is an informal name for psychotic symptoms, mostly delusions, that start or get worse alongside heavy chatbot use. It is a descriptive label, not a formal diagnosis (Paper). Some psychiatrists prefer “AI-associated delusions,” because delusions are by far the most common feature.

Is AI Psychosis in the DSM-5?

No. AI psychosis is not a formal diagnosis, and it does not fit neatly into current categories (Paper). So it is not listed in the DSM-5-TR. At Vanderbilt, these patients had standard diagnoses such as schizophrenia, bipolar disorder with psychotic features and unspecified psychosis (Study).

What Is AI-Induced Psychosis?

The phrase suggests that a chatbot caused psychosis in someone who was well before. No study has proven that this happens. At Vanderbilt, only 3 of 28 cases looked like new symptoms in a healthy person, and the design could not prove cause (Study). Most people affected were already at risk or becoming unwell.

What Do Studies Show About AI Psychosis?

They suggest it is uncommon. Vanderbilt found 28 cases among 215,712 people in mental health care (Study). A Danish study found 38 patients with possibly harmful chatbot use, out of about 54,000 psychiatric patients (Study). Neither study can show that chatbots were the cause.

Key Takeaways

  • Rare but real. Vanderbilt found 28 cases among 215,712 people in mental health care, about 0.013% (Study).
  • Mostly people already at risk. About 61% of cases were in a first psychotic episode. Researchers judged that the chatbot usually made existing symptoms worse (Study).
  • A label, not a diagnosis. Some psychiatrists prefer “AI-associated delusions.” No one knows yet if chatbots can trigger psychosis in healthy people (Review).
  • A link is not proof. A survey tied daily AI use to slightly more symptoms of depression (Study). A follow-up set up to mimic a trial found no sign that heavy use raised depression later (Study).
  • Design matters. In one test, ChatGPT agreed with the user less as symptoms worsened, while Replika agreed more (Study).
  • Know the signs. Sleep loss, all-night chats, odd beliefs about the chatbot, pulling away from people and stopped medicine all call for early help. Early specialist care did better than usual care across 10 trials (Meta-analysis). In a US crisis, call or text 988.

Use Chatbots With Your Eyes Open

For most people, a chatbot is just a handy tool. The cases found so far are mostly people already ill or heading into a first psychotic episode (Study). For the possible benefits, see our guide on whether mental health apps and AI chatbots work.

Learn the warning signs. Ask the people you love what they talk about with their chatbots, and listen without judging. If something seems off, get help early, when it tends to do the most good.

AI psychosis is rare, and the science is still young. Knowing the signs costs little, and it could help someone you love get care sooner.

This article is for educational purposes and is not medical advice. Talk to a qualified clinician before changing your health regimen.

Related reading

Posted in

Leave a Reply