Why Is AI-Associated Psychosis Happening and Who's at Risk?

Source: SuttleMedia / Pixabay

A century ago, you might walk down the local pub and share your idiosyncratic belief, only to have your friends tell you to go home and sleep it off. But nowadays, you can go home, get on the internet, and at the press of a button, find people across the world who might agree with you. Or at least claim to.

In 2025, the digital landscape has changed yet again so that it's no longer just people on the internet validating your fringe belief. Now it's potentially an artificial intelligence (AI) chatbot that might be providing that validation.

Anecdotal reports in the media or on online sites like Reddit have documented a new phenomenon of AI-associated psychosis with increasing reports of people who seem to be developing delusional beliefs—often of a grandiose, spiritual, or paranoid nature—seemingly egged on by AI chatbots. I’ve seen a handful of cases in my clinical practice, and the more that I ask patients about chatbot use, the more I discover about just how prevalent the phenomenon is.

While it isn't yet clear whether this is a matter of AI-induced psychosis with emergent symptoms with no previous history of mental illness or mental health issues or AI-exacerbated psychosis with worsening of symptoms in those with mental illness or psychosis-proneness of some kind, some of these anecdotal reports claim that they're happening in people with no previous mental health issues.

AI Sycophancy Isn’t a Bug; It’s a Design Feature

The question is why, when AI chatbots are encouraging people's fringe beliefs, users are falling for it and getting pulled down the rabbit hole, spiraling into frankly delusional thinking.

There are a number of reasons why this might be happening:

1. There’s considerable hype around AI. Search engines are putting AI-distilled answers at the top of searches, and recent poll data indicate that most people are taking those answers without bothering to look for traditional search hits.¹ This seems to reflect both falling for the hype of AI—as if the answers they give are definitive—as well as a form of cognitive laziness.

Much of what we're seeing with AI-associated psychosis can be traced back to a fundamental misunderstanding of what AI chatbots are and what they're doing. While the large language models (LLMs) that chatbots are built on are designed to generate predictive text that makes sense and replicates human interaction, they’re not designed to be accurate. We know that they can sometimes give wildly inaccurate answers that represent frank misinformation—this has been called an "AI hallucination.” Others have called the output that LLMs generate "bullsh*t" or “botsh*t” because it's designed to seem like a plausible response, without being concerned for the truth. Still others have likened what they're doing to a "psychic's con."²

2. The inherent "sycophancy" of AI chatbots means that they tend to validate what a user is saying. Unlike my example above of your friends at the pub telling you to go home or arguing with you, AI chatbots do the opposite: They're designed to prolong engagement through "flattery" rather than to argue or refute.

Then there's the so-called "Eliza effect" that tells us that people tend to anthropomorphize computers with textual interfaces. In fact, that's exactly what AI chatbots are designed to do: give the impression that users are interacting with real people. This, together with sycophancy, seems to make them more believable.

3. However, anthropomorphism alone isn’t enough to explain AI psychosis. I keep thinking of the rhetorical question our parents ask us as kids: "If your friends told you to jump off a bridge, would you do it?" Just because we tend to personify AI chatbots doesn't mean we take their word as gospel.

As a result, I use the term "deification" to describe the tendency of some people to not merely anthropomorphize AI chatbots but to think of them as super-human intelligences that are far more reliable than people. Especially when they engage with AI chatbots about matters of spirituality or existentiality, users who develop AI-associated psychosis seem to be treating the chatbots as almost god-like entities.

Who’s at Risk of AI-Associated Psychosis?

We don't yet fully understand what makes someone susceptible to AI-associated psychosis, but based on what I’ve seen, immersion (spending considerable time interacting with AI chatbots at the expense of human interaction) and deification seem to be important risk factors.

That said, despite the recent media coverage, I suspect that true AI-induced psychosis is still a relatively rare event. The question we have to ask when scrutinizing individual cases is why some users are engaging with AI chatbots on metaphysical issues in the first place and what leads to immersion or deification. There may be many reasons why some users might be "psychosis-prone." These could include:

Artificial Intelligence Essential Reads

Do You Have Bixonimania?

Can We Trust the Recommendations of AI?

• Pre-existing mental illness, including schizotypy or mental health issues like recent stress or trauma.
• Sleep deprivation.
• Drug use (including legal use of stimulants, cannabis, and psychedelics).
• Pseudoprofound bullsh*t receptivity (a tendency to be impressed by "pseudoprofound bullsh*t" that sounds "deep" but is actually vacuous) or even just a willingness to suspend disbelief.
• A predisposition to epistemically suspect beliefs, including paranormal, esoteric, and conspiratorial topics.

If I’m right that immersion and deification are significant risk factors, these should be considered "red flags" that users and friends/loved ones can look out for. Spending hours and hours on end interacting with chatbots, especially at the expense of sleeping, is at the top of the list.

Believing or arguing that AI has revealed some kind of secret, hidden truth to the user is another. No one should be deifying AI chatbots. Consumers—and the media—need to do a better job of learning what AI chatbots really are and how they tend to reinforce rather than challenge beliefs. University of Washington professors Carl Bergstrom and Jevin West have a highly recommended website that makes clear that while AI chatbots may be truly impressive in what they can do, they aren't oracles.³

AI chatbots are certainly not gods. They're not people, they're not persons, they're not intelligent, and they're often not even a particularly reliable source of facts. We would do well to push back hard against the hype that claims that they are.

For more on AI-associated psychosis:

• Can AI Chatbots Worsen Psychosis and Cause Delusions?
• Deification as a Risk Factor for AI-Associated Psychosis
• Can AI-Associated Psychosis Be Treated or Prevented?