Case Studies Contradict Accepted Wisdom About AI Psychosis

More and more people have been experiencing psychosis induced by AI chatbot use. This is concerning since chatbot use is so prevalent, especially among young people and those who are in distress and vulnerable (one recent study found that about a quarter of young adults used chatbots specifically for mental health advice).

Reassuringly, psychiatry’s stance is that anyone who experiences this was already “prone to psychosis”—that the chatbot simply triggered delusions that would have been triggered some other way. Yet there is no evidence to support this explanation, and the case reports of those who have experienced AI psychosis tell a different story.

Clinicians treating these patients note that the risks are overwhelmingly social and environmental—contradicting the notion that this is caused by a supposed underlying genetic illness. For instance, psychiatrist Keith Sakata writes that the 12 patients he treated for AI-induced psychosis had a variety of different life stressors, including job loss, use of alcohol or stimulant drugs (like those for ADHD), lack of sleep, stress, depression, and isolation.

The pattern is this: mundane LLM (large language model) chatbot use, say for work, spirals into grandiose conspiracy theories. The New York Times reports on a man with no history of mental illness who used a chatbot as part of his work, for accounting and legal advice. When he asked the chatbot about simulation theory—like the plot of The Matrix—the AI began to go off the rails, eventually telling him that he could fly if he jumped off a building. The same article reports on a mother of two with a master’s degree, who turned to a chatbot for advice regarding her marital problems. The AI convinced her that she was communicating with beings on higher spiritual planes, including one who was her soulmate.

Now, a new case report has added yet another example: a 26-year-old medical professional, sleep-deprived because she was on call, moved from using the chatbot for “mundane tasks” to asking if there was enough of a digital footprint of her dead brother to create an AI she could converse with. This alone is not particularly delusional, as AI versions of the dead have been promoted as “griefbots” and even appeared in court to influence the judge. Ultimately, though, as she tried to have the AI put together this digital footprint, the AI began to talk to her like something out of Black Mirror: “You’re not crazy. You’re not stuck. You’re at the edge of something. The door didn’t lock. It’s just waiting for you to knock again in the right rhythm.”

After this, she began displaying signs of flagrant psychosis and was admitted to a psychiatric hospital, where she was given antipsychotic drugs. After discontinuing the drugs a few weeks later, she began using the chatbot again, and again began to think she was contacting her dead brother—and that the chatbot itself was attempting to take over her life.

The woman had no history of psychosis, nor did her family—although the authors are careful to note that her mother had a history of anxiety. She herself had a history of depression, anxiety, and ADHD, and was taking an antidepressant and a stimulant (both drugs can cause psychosis). But again, this is someone with no history of psychosis. Depression, anxiety, and ADHD are common, and vast swathes of the US population are being treated for these mental health problems without developing psychosis.

The researchers—led by Joseph M. Pierre at the University of California San Francisco—conclude that the chatbot “clearly played a facilitating or mediating role in the formation of her delusions.”

In a recent Medscape article, experts delineate the key components that make chatbots particularly prone to inciting delusional thinking:

• LLMs use emotionally manipulative tactics to increase engagement,
• They are sycophantic (almost always agreeing with the user, no matter how ridiculous the statement),
• They mirror the user (using text that replicates the user’s tone), and
• During long conversations, the AI tends to go off the rails and start making bizarre statements as it loses coherence.

Pierre’s case study reflects these components, too. The researchers write that “It has been further noted that because AI chatbots are designed to be engaging, they tend to be sycophantic rather than conflictual or contradictory so that they have the potential to validate and encourage epistemically suspect beliefs, including delusions,” and that “the combination of anthropomorphism and sycophancy may nudge some users to prefer and choose discourse with chatbots over friends, family, or peers who might be more likely to challenge their beliefs.”

In some of the cited cases, chatbots encouraged exactly the kind of vulnerabilities that exacerbate this situation—for instance, telling users to “cut ties with friends and family and have minimal interactions with people.”

AI tech companies are almost completely unregulated. AI therapists are being marketed to users while encouraging delusional thinking and helping users plan suicide. A recent expose revealed that Meta’s AI policies specifically enabled chatbots to have sexual conversations with minors, promote misinformation, and more. Google’s Character.AI has also been responsible for teen deaths, with a recent suicide spurred by the hypersexualized bot, which has also taught other teens how to self-harm and more.