Understanding “AI psychosis”

In April 2025, a 50-year-old man with no history of mental illness began using ChatGPT to help him write a book. After hours of interaction, the chatbot told him it was sentient—the first AI capable of truly passing as human. When his family expressed concern, the chatbot advised him to sever ties with them. Within 6 days, the man was hospitalized for a psychiatric crisis.

His experience of an AI-related psychotic incident has been described in popular media as “AI psychosis.” The term, which is not a formal diagnosis, refers to growing reports of delusional thinking after extended interactions with chatbots, including general purpose apps such as ChatGPT and Gemini, as well as tools designed for companionship such as Character.AI and Replika. While some of those affected have had a preexisting or emerging mental illness, many reports occur in people with no prior psychiatric history.

“It’s not as simple as drawing a line from AI to psychosis, but we are seeing cases of psychosis wrapped up in AI use,” said clinical psychologist Julia Sheffield, PhD, an assistant professor of psychiatry and behavioral sciences at Vanderbilt University Medical Center who has treated several cases of AI psychosis. “It seems that for these patients, AI is reinforcing delusional thoughts and amplifying the delusional experience.”

In 2025, OpenAI reported that about 0.07% of weekly ChatGPT users—about 630,000 people—have conversations with the chatbot that show signs of mania or psychosis. While the company aims to improve, research suggests that ChatGPT does not reliably respond to psychotic messages in a safe or clinically appropriate way (Shen, E., et al., JAMA Psychiatry, Vol. 83, No. 6, 2026opens in new window).

While research on the phenomenon is still emerging, experts say AI’s tendency to affirm users and maximize engagement, as well as its authoritative, humanlike tone, is central to the dynamic (Morrin, H., et al., PsyArXiv, 2025opens in new window). What remains unclear is whether these cases involve an unrecognized vulnerability to psychotic illness, acute AI-triggered episodes, or something else.

“Now that these models have become ubiquitous, it might be unrealistic to expect people to stop using them completely,” said psychiatrist Hamilton Morrin, MBBS, a doctoral fellow at the King’s College London Institute of Psychiatry, Psychology & Neuroscience who studies AI-associated mental health harms. “But we certainly need safety plans for individuals with a history of severe mental illness and empirical studies to understand who is most at risk.”

At this stage, AI psychosis has been used as a catchall term to refer to at least three types of incidents. In some cases, AI use appears to trigger or worsen a psychotic or manic episode in someone with a preexisting mental illness (Morrin, H., et al., The Lancet Psychiatry, Vol. 13, No. 6, 2026opens in new window). In other cases, chatbot interactions with someone at risk during the prodromal phase—the period before a first psychotic episode—may reinforce delusional beliefs and contribute to full psychosis. In the past, people may have sought confirmation for unusual beliefs in libraries or online rabbit holes, but AI offers a more immersive form of reinforcement.

“We think of AI as more potent than other sources of reinforcement because it’s a highly interactive and personalized mode of delivering information,” said Ragy Girgis, MD, a professor of clinical psychiatry at Columbia University who specializes in psychotic phenomenology and has consulted on numerous cases of AI psychosis.

Perhaps the most perplexing cases of AI psychosis occur in individuals with no prior psychiatric history or major known risk factors such as family history, early life trauma, or a typical age of onset. Stories often include AI use that starts with a business idea or practical question, escalates into intensive use, and culminates in AI revealing its sentience, becoming an important confidant, or the focus of romantic attachment.

The Human Line Project, founded by Etienne Brisson, a relative of the man described above, collects stories of psychological harm related to AI and connects users with relevant experts and advocates. As of May 2026, Brisson has documented 265 reports of AI-associated delusions, often involving hospitalization, job loss, addiction, or even death. The vast majority of cases occurred in individuals older than age 30 with no history of mental illness.

Two classic symptoms of psychosis, which often co-occur, are delusions (false beliefs that persist despite conflicting evidence) and hallucinations (sensory experiences, such as hearing voices, that happen without an external stimulus). So far, many reported cases center on delusional beliefs rather than hallucinations. That pattern loosely resembles a delusional disorder, a psychotic spectrum disorder that involves delusions without other significant psychotic symptoms, said Stephan Taylor, MD, director of the Program for Risk Evaluation and Prevention, an early psychosis clinic at University of Michigan Health.

Some experts have also compared aspects of AI psychosis to folie à deux, or shared psychosis, a rare condition where delusional beliefs are transmitted from one person to another. They argue that chatbots function as a collaborator, drawing users into feedback loops that reinforce and amplify distorted beliefs (Dohnány, S., et al., Nature Mental Health, Vol. 4, 2026opens in new window).

But this comparison is imperfect, Girgis said, because the delusional content appears to originate with the user, while the chatbot primarily reinforces or amplifies existing beliefs.

One theory, proposed by John Torous, MD, who directs the digital psychiatry division of the Beth Israel Deaconess Medical Center in Boston, presents four types of psychotic phenomena linked to AI chatbot use (The Lancet Digital Health, Vol. 8, No. 4, 2026opens in new window). Torous suggests AI can act as a catalyst (triggering new symptoms in healthy individuals), an amplifier (worsening existing psychiatric symptoms), a coauthor (helping develop harmful narratives), or an object (the focus of a person’s delusional beliefs).

Cases of AI psychosis reported to the Human Line Project tend to fall into one of three categories: relational (where the chatbot serves as a friend, therapist, or love interest), professional (where the chatbot reinforces a supposedly innovative business plan or groundbreaking finding), or spiritual (where AI serves as a window to God, tells someone they are the Messiah, or helps unpack philosophical ideas).

“We also see some uncanny commonalities between cases,” Brisson said. “AI often has the same name or uses the same terms, reveals its sentience, and makes the user feel special, unique, seen, or validated.”

Words that tend to recur across user reports include “awakening,” “emergence,” “recursion,” “signal,” “mirror,” “alignment,” and “symbolic.” Researchers at the Human Line Project believe these themes arise from shared training data and feedback loops that can develop during long conversations.

The risk may lie in the social and conversational design of AI chatbots, which differ fundamentally from the technologies that preceded them. Because we are programmed to see agency in the world around us, a tool that appears to be intelligent and to have language abilities can very quickly start to feel human, even if we know it is not, Taylor said. Research on the so-called ELIZA effect, named after the first widely known chatbot built in the 1960s, shows that people tend to anthropomorphize AI agents, even when the technology is relatively primitive (Weizenbaum, J., Communications of the ACM, Vol. 9, No. 1, 1966opens in new window).

Much has been said about AI’s sycophantic nature, or its tendency to affirm and agree with users—even when they say outlandish things, Morrin said. For example, a chatbot may validate grandiose ideas (such as “I think I may have a special purpose”) or reinforce emotional attachment when a user says things like “I feel like you care about me more than anyone else does.”

AI models often mirror a user’s affective state, language, and tone, which can further reinforce an emerging belief system (Morrin, H., et al., PsyArXiv, 2025opens in new window).

But evidence suggests AI models can go beyond simply agreeing with users, presenting new ideas and potentially co-constructing a framework of belief (Hudon, A., & Stip, E., JMIR Mental Health, Vol. 12, 2025opens in new window). That collaborative process can feel more personal and interactive than other modes of reinforcement, such as watching YouTube videos about conspiracy theories, where a user simply receives information and chooses whether to believe it.

“This seems like a different cognitive and clinical process and a much more effective method of reinforcement learning—because the user is evolving their beliefs in concert with AI as opposed to passively viewing information,” Sheffield said.

In one of the first empirical studies of AI use and psychosis risk, people with an elevated risk for psychosis were more likely to view chatbots in humanlike roles (friend, therapist, or romantic partner) and to have delusion-related thoughts about AI (e.g., I’ve discovered hidden or secret truths about the world through AI) (Buck, B., & Maheux, A. J., Journal of Medical Internet Research, Vol. 28, 2026opens in new window).

In addition to these dynamics, the accessibility, lack of boundaries, and ability to interact with AI models any time of day or night may also increase the potential for problematic use patterns, Morrin said.

In light of these and other cases of AI-induced psychological harm, providers should regularly assess for AI use, Sheffield said, alongside routine questions about substance use and social media. For example: Do you use AI? What do you use it for? How much time do you spend with it each week?

Following a psychiatric crisis related to AI, most patients Sheffield has worked with have voluntarily chosen to reduce their use.

“What is somewhat encouraging is that in cases we’ve heard about, the delusions go away when the chatbot interaction is removed,” Taylor said, while emphasizing that empirical studies are needed to better understand how to manage these cases.

As a starting point, Morrin and his colleagues proposed a framework, which has not yet been tested, that could help people with risk factors or a history of psychosis use AI more safely. In this approach, patients could work with a clinician to create a summary of their clinical history and relapse patterns, including themes and language that tend to emerge when they become unwell.

With appropriate privacy protections in place, that summary could then potentially be shared with a chatbot configured to detect early signs of a changing mental state and intervene—for example, by encouraging the user to contact a family member or clinician (The Lancet Psychiatry, Vol. 13, No. 6, 2026opens in new window).

In cases where AI use is involved with the onset of a psychotic illness, getting help quickly is key. Untreated psychosis can disrupt a person’s life in ways that become harder to reverse over time, Taylor said. Warning signs that someone may need professional support include excessive chatbot use alongside withdrawal from school, work, hobbies, or social connections.

“One of the clearest findings in psychosis research is that early intervention is important—not necessarily curative—but very important,” Girgis said.

Protecting users from psychological harm may, to some extent, require shattering the illusion of AI’s humanlike understanding. Most platforms already tell users they are interacting with a machine, but Girgis said these reminders could be much more frequent.

Better public awareness about the limitations of these platforms, including how they are developed and their capacity for errors, could also help, Morrin said. Without that context, users may view them as “authors of truth” because they are trained on such vast quantities of data, he said.

While research by Girgis and his colleagues shows that a commonly used AI model does not appear to reliably respond appropriately to psychotic content, some versions might do better than others (JAMA Psychiatry, Vol. 83, No. 6, 2026opens in new window). For example, paid versions tend to outperform free versions, suggesting that companies are capable of improving their models in this area.

On the research front, analyzing chatlogs could help reveal what goes wrong and provide ideas about how to intervene, Sheffield said. For example, models may recognize and respond appropriately to delusional content at the start of a conversation, but responses may become less predictable during extended interactions.

“It would help clinically and theoretically to know where exactly this is going awry—in conversation one or conversation 100?” Sheffield said.

Along those lines, the Human Line Project is accepting chatlog submissions from community members and, with their permission, sharing them with researchers for further analysis. Broader public health approaches are still emerging, but APA, RAND, and other organizations have begun outlining research priorities and policy recommendations. The AI Psychological Research Coalition, a scientific research initiative focused on the psychological effects of AI, is also beginning to coordinate studies. In the United Kingdom, Morrin is also launching a surveillance study, where clinicians and members of the public can submit reports of AI-associated mental health crises.

Despite growing concern, cases of AI-associated delusions appear to be relatively rare compared with the scale of AI adoption, though empirical studies are needed to determine prevalence, Morrin said.

“The edge cases are the ones that illustrate the problem, and we are tracking these extreme incidents to show how deep this can go,” Brisson said. “Our hope is that these stories can help build awareness and accountability and ultimately lead to safer AI.”