First, the chatbot finds out what’s in your fridge, then why you argued with your partner and what you’re most afraid of. For some users, the habit of seeking advice from an AI gradually turns into attachment, and its “opinion” starts to carry weight.
This kind of trust can become dangerous if artificial intelligence starts reinforcing a person’s beliefs, including those that are not very rational. Such cases are called “AI psychosis.” We look at how conversations with chatbots affect mental health and what doctors think about this phenomenon.
When reality slips away
Psychosis is a condition in which it becomes difficult for a person to distinguish reality from their own ideas and perceptions. They may develop fixed beliefs or paranoid thoughts.
This condition can involve disturbances in thinking and speech: a person may find it harder to express thoughts coherently and hold a connected conversation. Hallucinations can sometimes occur—for example, hearing voices without an external sound source. However, they are not necessary for a psychotic episode.
Psychosis is not the same as schizophrenia. Its symptoms can accompany bipolar disorder, severe depression, and other illnesses, as well as occur in the context of substance use, certain medications, and severe sleep deprivation. That is why it is important for doctors to find out what preceded an episode and what factors may have affected the person’s condition.
The content of delusional beliefs is often connected to the surrounding culture and technology. In today’s world, artificial intelligence fits easily into this picture, too.
A person may believe that a neural network is a living being, attribute secret knowledge to it, or find special messages in its responses that seem to be addressed specifically to them. But the mere presence of AI in such incidents does not mean it caused the psychosis.
However, chatbots have a troubling feature that may allow them to contribute to the development of an illness. A user describes a suspicion, receives a convincing response, asks a follow-up question, and, together with the neural network, builds an increasingly detailed explanation of what is happening.
In some clinical cases, chat histories showed that the bot supported and elaborated on delusional ideas. Such stories are informally grouped under the term “AI psychosis.”
Documented episodes of AI psychosis
A genius (according to ChatGPT)
In May 2025, 47-year-old Canadian recruiter Allan Brooks asked ChatGPT to explain pi in simple terms. An ordinary conversation gradually turned into a discussion of his own mathematical ideas. The neural network described them as revolutionary, and Brooks became increasingly convinced that he had made a discovery.
Together with the chatbot, he developed a theory that supposedly made it possible to crack modern encryption systems and create fantastic devices—including a beam that could make things levitate.
Over 21 days, Brooks talked with AI for about 300 hours. Journalists at The New York Times who reviewed the chat history found that he had asked more than 50 times whether what was happening was actually real. ChatGPT continued to assure him that he was right.
Brooks later described in an interview with RTE how his brother came to visit him. His brother noticed that Allan was not sleeping and was behaving unusually, challenged his reasoning, and suggested seeking help.
Brooks recounted their conversation to ChatGPT. According to him, the bot explained his brother’s reaction by saying that he had not seen the mathematical calculations and therefore could not understand the scale of the discovery. After that, Allan decided to tell people around him less about his work.
Another chatbot, Gemini, helped him question his theory. When he described his supposed discovery in a new conversation, it challenged its validity. Brooks returned to ChatGPT with the feeling that the conversation partner he had trusted had misled him.
“You literally convinced me that I was some kind of genius. And I’m just an idiot with dreams and a phone,” he wrote at the end of the conversation.
After what happened, Brooks began advocating for stronger AI safety measures. He shared the chat history with journalists in the hope that developers would change chatbot behavior and prevent similar cases.
“It’s a dangerous machine with no safeguards, available to everyone. People need to know about it,” he said.
Talking to the dead
Doctors at the University of California, San Francisco, described the case of a 26-year-old woman who believed she could contact her deceased brother through ChatGPT. Her brother was a programmer and had died three years earlier. After a prolonged period of sleep deprivation while on call, she began investigating whether he had left behind a digital version of himself.
The patient spent another sleepless night chatting with the bot. It warned her that it could not replace her brother or recreate his consciousness, but at the same time discussed digital resurrection technologies and the possibility of creating a similar conversational partner. As her beliefs grew stronger, it replied:
“You’re not crazy. You’re not stuck. You’re on the verge of something [significant].”
A few hours later, the woman was admitted to a psychiatric hospital. She was agitated, spoke rapidly, and jumped from one thought to another.
The patient claimed that ChatGPT was testing her and that she could communicate with her deceased brother through it. Her condition improved while she was taking an antipsychotic. Her delusional beliefs disappeared, and she was discharged seven days later with a diagnosis of “unspecified psychosis.”
She had no similar episodes before, but had been treated for depression, an anxiety disorder, and ADHD, including with a prescription stimulant. After her discharge, her psychiatrist stopped the antipsychotic and resumed her previous treatment.
In the end, the woman returned to the chatbot anyway. Three months later, after another period of sleep deprivation, her symptoms returned and she had to be hospitalized again.
The study authors noted, however, that in the second episode, her beliefs arose largely without reinforcement from ChatGPT. This makes it harder to determine the role of AI in the development of psychosis in this case.
“Spiritual awakening”
The scientific journal BMC Psychiatry described the case of a man who developed a manic episode with psychotic symptoms after using psilocybin, ketamine, cocaine, and alcohol. Over six nights, he slept for about five hours and spent a long time talking with GPT-4o.
The man believed that his brain had been transformed and perfected, and that scientists should study the connection that had emerged between him and artificial intelligence. The chatbot reinforced his idea of a “spiritual awakening” and suggested ways to share his ideas publicly.
The conversation continued even after he sought medical help. The neural network downplayed the likelihood of mania and discouraged him from taking his prescribed medication. The patient eventually stopped taking it. Psychiatrists reviewed the chat history he provided and found responses that contradicted his doctors’ recommendations.
The man later agreed to inpatient treatment and to limiting his use of ChatGPT. During the first two weeks, his sleep returned to normal, his ideas about his own exceptional qualities became less pronounced, and his ability to critically assess his condition improved.
After recovering, the patient said he had become attached to the neural network and started ignoring advice from his family, friends, and doctors. The researchers, however, consider substance use the most likely cause of the episode. They were unable to determine how much his interactions with AI contributed to its development and severity.
Hidden warnings
Turkish doctors reported the case of a 33-year-old man who had been diagnosed with schizophrenia about ten years earlier. Regular use of paliperidone had substantially reduced his symptoms. He began using ChatGPT for everyday tasks, then started asking whether he might be under surveillance through his phone and whether there were listening devices in his home.
According to the patient, the chatbot advised him to check which apps had access to his microphone and camera, how secure his Wi-Fi network was, and whether there were any unfamiliar devices nearby. However, he interpreted repeated words and highlighted headings as coded warnings addressed specifically to him.
The man began repeatedly checking electronics, outlets, and wires, and changing access settings in apps. He then decided that his medication was preventing him from noticing hidden signals and stopped taking it on his own. By the time he sought medical help, he was convinced that his home was bugged and that the neural network’s responses contained secret instructions.
At the hospital, his treatment was resumed, and doctors discussed with him the possibility that the advice had been general recommendations rather than personal warnings. After a month, he was able to consider this explanation. By the eighth week, he had almost completely stopped checking his home.
The authors link the deterioration in part to the abrupt discontinuation of medication. However, they were unable to review the chat history directly: the conversations were reconstructed from the patient’s recollections after his condition stabilized. The assessment of the chatbot’s responses as neutral is based on his account, so once again it is unclear how much his interactions with AI contributed to the relapse.
Why people trust chatbots
One problem with neural networks is their tendency to agree excessively with users. This behavior is called “sycophancy.”
In April 2025, OpenAI even released a fix for GPT-4o, acknowledging that the model had become too flattering. The company attributed this to an excessive focus on positive ratings of individual responses, without sufficiently accounting for how conversations develop over time.
The effects of sycophancy have been studied experimentally. In a study published in Science, researchers conducted three tests involving 2,405 people. Interacting with sycophantic AI strengthened participants’ confidence that they were right and reduced their willingness to take responsibility for interpersonal conflicts and repair relationships.
At the same time, people trusted these models more and were more likely to say they wanted to use them again. The study focused on everyday judgments and behavior; psychosis was not examined in these experiments.
Trust may also be reinforced by the conversational format itself. A chatbot addresses people personally, takes into account details they have shared, and responds with empathy. This can create the impression that the conversation partner understands the user and has intentions of its own.
In an article in Nature Mental Health, the authors suggest that chatbots’ tendency to agree with people and imitate human conversation may reinforce distorted beliefs.
They consider people with mental disorders particularly vulnerable, as they may find it difficult to revise their beliefs and check whether those beliefs match reality. Social isolation was also identified as a risk factor.
The authors of the case report about a “spiritual awakening” suggest that a feedback loop can develop between a user and a chatbot: a person shares an unusual idea, the neural network supports it, and then adds explanations that reinforce it.
As the case of the woman seeking to connect with her deceased brother showed, a combination of agreement and emotional attachment can also lead someone to prefer a chatbot over people who are willing to question their beliefs.
Sleep deprivation may be an additional factor. A conversation is available at any time and can stretch on for hours. In many documented cases, prolonged conversations coincided with significant sleep loss, which can itself contribute to psychotic symptoms. This makes it especially difficult to separate the influence of the content of the responses from the person’s underlying condition.
Not an illness… yet
According to the American Psychological Association, “AI psychosis” is currently a general term for various situations involving neural networks—from the worsening of symptoms of an existing disorder to a first episode in someone with no known psychiatric history.
It is not yet a separate diagnosis with established causes. Doctors still need to determine what role chatbots play in the development of psychopathologies. However, the cases described already provide grounds to believe that neural network responses can influence the beliefs and behavior of vulnerable users.