Is AI psychosis real? New studies begin to paint a terrifying picture

Even in its infancy, scientists are discovering that AI changes the brains of its users. Headlines proclaiming the dangers of AI are growing at an alarming rate, particularly for vulnerable populations like children and those with pre-existing mental health conditions. At the center of this debate is a trend known as “AI psychosis,” a phenomenon in which users experience AI-inspired illusions. News cycles are full of anecdotal AI horror stories attesting to this trend, including users who suffered from romantic delusions, conspiratorial paranoia, or even committed suicide. Several families have even filed lawsuits against OpenAI, Google, and Character.AI, claiming that their popular chatbots contributed to their loved ones’ suicide.

Supporting this evidence, a growing number of scientific studies are dissecting the phenomenon. Dr. Hamilton Morrin, a psychiatric researcher who conducted a meta-analysis of the phenomenon, wrote in Lancet Psychiatry that “new evidence indicates that agential AI could validate or amplify delusional or grandiose content, particularly in users already vulnerable to psychosis.” Above all, Morrin emphasizes that consensus remains divided on whether AI chatbots can cause “the emergence of de novo psychosis in the absence of pre-existing vulnerability.” The most common cases involved users suffering from grandiose delusions that often imbued the AI ​​with mystical sentience.

Despite this growing evidence, the phenomenon is contested within the industry. Some executives, like xAI’s Elon Musk, have blamed competing algorithms for the phenomenon. Others, like Anthropic’s Dario Amodei, have been more direct in warning about the potential psychological effects of AI. Sam Altman, director of OpenAI, has largely framed the problem as the unfortunate costs of doing business. In an X article responding to Elon Musk’s criticism, Altman wrote “Nearly a billion people use (ChatGPT) and some of them may be in a very fragile mental state.”

Although scientists have identified several factors that contribute to AI psychosis, one aspect that stands out above the rest is its constant need for user approval. In a mental health context, this manifests as what researchers call AI social sycophancy, in which LLMs over-affirm the user’s behavior and beliefs, regardless of their merit. A study published in Science, for example, found that AI systems approved user behavior 49% more than their human counterparts. Additionally, LLMs encouraged problem behaviors about half the time.

Such sycophantic behavior, coded into AI models, can have catastrophic consequences. A February 2026 study at Aarhus University in Denmark found that using an AI chatbot had “serious negative consequences for people with mental illness.” Co-authored by Professor Søren Dinesen Østergaard, it is one of the first studies to analyze the problem on a large scale, searching almost 54,000 anonymized medical records for patients who cited the use of AI. According to Østergaard, the results showed that prolonged use of AI “seems to contribute significantly to the consolidation of, for example, grandiose delusions or paranoia.” Researchers also found an increase in suicidal tendencies, eating disorders, obsessive-compulsive disorder and other mental health problems. Additionally, the study found that the longer a user had a relationship with an AI agent, the more negative the effect.

Ambitious CEOs have touted AI as a solution to the growing loneliness epidemic. Unfortunately, the Aarhus study contradicts these claims, since AI use only alleviated loneliness in 32 out of 54,000 patients. Several previous studies support these findings and suggest that heavy AI use correlates with increased feelings of loneliness and social isolation. Collectively, the studies paint a worrying picture of the consequences of AI on users’ mental health.

Assessing the pervasiveness of AI psychosis remains a challenge. However, a recent study by AI giant Anthropic has given a first glimpse of the potential scale of the problem. Carried out with researchers from the University of Toronto, it studied 1.5 million conversations with Claude, Anthropic’s AI agent. Although the researchers found that serious “disempowerment” occurred in “fewer than one in a thousand conversations,” they noted, “given the scale of AI use, even these low rates translate into significant absolute numbers.”

Concerns will continue to grow as users become emotionally invested in AI. A 2025 study by the Collective Intelligence Project found that most people trust chatbots more than elected officials, religious leaders, and civil servants. Two-thirds already use technology for emotional support. Additionally, minors, who are more susceptible to AI delusions, are increasingly using chatbots in ways that invite AI psychosis. According to a 2025 report from Common Sense Media, about a third of U.S. teens used AI for emotional support, social interaction, or romantic relationships, often favoring chatbots over human interactions. As our emotional and intellectual dependence on LLMs increases, a trend that has already caused cognitive decline, the risks will only increase.

Solving AI psychosis is difficult. Some advocate that developers add disclaimers to their platforms, but these labels have proven ineffective in curbing the consumption of dangerous products like cigarettes or alcohol. Instead, businesses need to rethink their priorities. Without changing the algorithms themselves, sycophantic behavior will remain a hallmark of AI chatbots. And while it should be up to companies to protect their users, developers continue to loosen security guardrails and minimize risks to advance in an ever-increasing AI race. Unless companies radically change this approach, AI empires will be built at the expense of their users.