Since 2022, the use of large language model chatbots like ChatGPT has skyrocketed.

And as funny and harmless as the talking computer might seem to some, this explosion in usage comes with some real and serious concerns. One of these concerns is what news outlets and psychiatrists alike have dubbed AI psychosis.

Far from people just being  confused about what AI can do, this is when real mental health symptoms arise in connection with the use of AI chatbots. And as we’re defining it for the purposes of this video, people have died as a result. We need to talk about it before it’s too late.

So we’re going to talk about it. What it is, how dangerous it can  be, and what we can do about it. [♪INTRO] As a warning, this video  will include a discussion of deaths by suicide in connection with AI chatbots. AI psychosis is different from AI hallucinations, which is when generative AI fills in a blank with something, anything, whether it’s true or not.

We’re also not using the term  “psychosis” in a colloquial sense. I want to make it really, really clear  that for the purposes of this video, we are using the term “psychosis”  in its clinically defined sense. In AI psychosis, people report  actual psychosis symptoms that have been intensified, supported, and  developed alongside AI chatbots.

Psychosis isn’t a disease or  diagnosis in and of itself. Rather, it’s a collection of symptoms  that reflect a break with reality. The experiences most commonly associated with psychosis are hallucinations and delusions.

Hallucinations happen when  your brain tells you there’s some sort of sensory experience  without any physical trigger. We usually think about hearing  voices or seeing things, but we can have hallucinations  in any of our senses, like constantly feeling bugs crawling  on our skin when nothing is there. Delusions are false beliefs  that don’t align with the person’s cultural context  and stick around regardless of how much evidence there is against them.

Whatever specific symptoms a person  has, the end result is that they aren’t able to tell the difference  between what’s real and what isn’t. That loss of contact with reality  also means that people with psychosis are often unable  to recognize that something is wrong in their thinking,  rather than with the situation. So, for example, someone  may talk about how they’ve recently started being followed by the FBI.

They might say they’re scared of the agents, and feel that there’s an  unjust conspiracy against them. But they’re not able to identify  that they only think they’re being followed by the FBI, and  they can’t tell that the problem is the thoughts they’re  having, not a real conspiracy. We often think about these kinds of paranoid and persecutory delusions when  we think about psychosis, but delusions can be related to a  whole range of different themes.

Some people may believe that  Beyonce is secretly in love with them, or that they have  invented a new branch of science, or that they smell like rotting meat  and they can’t get rid of the scent, or that everyone can hear their thoughts. And although urban legends  love to center quote-unquote “psychotic killers”, psychosis  isn’t inherently violent. The vast majority of people with  psychosis are never violent.

The risk to the person experiencing symptoms, as well as to others, depends  on the specific delusions, their severity, and access to treatment. The highest risk for violence occurs  during someone’s first episode, and a 2024 study found that the rate of violence during a first psychosis episode was about 13%. That study also found that  after initial treatment, the risk of violence was about  two times lower than it was before treatment, and that  there’s even been a decline in the overall risk of violence  over the past 20 years.

What that treatment looks like depends on the underlying cause of the psychosis symptoms. Psychosis can be due to  mental illness, substance use, dementia, head injury, or a bunch of other things. In some cases, treating the underlying condition will treat the psychosis symptoms.

In other cases, the psychosis  symptoms may be targeted directly with antipsychotic medication,  therapy, or inpatient psychiatric care. So when we talk about AI psychosis,  we’re talking about a specific manifestation of these symptoms  as they relate to the use of AI. See, how psychosis manifests has  always been culturally dependent.

South Asian and Middle Eastern  countries have a higher prevalence of jealousy-related delusions  than English-speaking countries, and Eastern European countries  have a higher prevalence of guilt and sin-related delusions than  their Western counterparts. And how these symptoms are viewed depends on the context of the culture around them. The nature of the delusions people experience also changes alongside the  technology in their lives.

When radio and TV became more widespread, delusions related to radio or  TV also became more common. People might believe thoughts were  being implanted in their heads via radio, or secret messages  sent to them through the TV – things they couldn’t have come up with before. So it’s not surprising that less than  a year after the launch of ChatGPT, psychiatric researchers were  speculating that interacting with chatbots would cause  real-life delusions in people susceptible to or already  suffering from delusions.

And based on case reports, they were correct. Yes, media-influenced psychosis  is nothing new, but there are reasons to think chatbots  represent a particular risk. Many chatbots are designed in ways that may make people more vulnerable to developing delusions, or encouraging ones that already exist.

For one thing, chatbots are  interactive. That’s the whole point. More traditional forms of  communication are one-way.

A person may hear messages from a book or movie, but they’re not able to communicate back. The back-and-forth nature of a chatbot creates the impression that this  is a personal conversation. Even though the bot’s responses  are generated based on probability, it’s easy to believe there’s  intention where none exists.

Chatbots are also designed to sound like you’re talking with a real person, not a machine. That opens the door for all sorts of delusions, from believing the chatbot is in love with you to, to believing that you’re speaking with God, to knowing that there’s no way  a computer could speak this way, so it must actually be a  person who’s spying on you. And many chatbots are intentionally supportive, agreeable, and nonjudgmental.

That’s one of the things that makes them appealing for some people  to keep interacting with. But this sort of sycophancy is really problematic when you have someone suffering from a delusion. While supporting people and  acknowledging their feelings is incredibly important in the  management of psychosis symptoms, that does not include going  along with their delusions.

Validating someone’s delusions  can reinforce the idea that there’s nothing wrong, and people who don’t believe there’s anything  wrong don’t get treatment. So even though debating a  delusion isn’t typically helpful, pointing out how your perspective  of the situation is different and encouraging the person to think about  alternative explanations can be. And a lot of large language models aren’t trained to provide any sort of  resistance to problematic ideas.

Tech companies have made some  changes, like GPT-5 no longer acting like a romantic partner, but  it’s hard to say whether it’s enough. Add to all of that the fact that,  depending on the model and your subscription, it’s possible to keep  interacting with a bot for a long time. That means people can spend the  whole day talking to something that feels like a person, and that  validates and supports their delusions.

It’s a perfect storm for people  who are vulnerable to psychosis. ChatGPT was released near the end of 2022, so this is a relatively new phenomenon. And the wheels of the research  process can turn slowly, so we are just beginning to  see peer-reviewed studies on the causes of, vulnerabilities  to, and treatment of AI psychosis.

Even that literature is mostly a growing number of case reports and individual experiences. Which means that in order to best help patients, psychologists and psychiatrists  need more research ASAP. A 2026 paper from researchers  in Massachusetts suggests that large language models may  play different roles in psychosis.

This is only one possible  framework for studying this topic, but the researchers say studying  cases according to those different roles could improve our  understanding of AI psychosis. The first role is the catalyst. In these cases, interacting  with a chatbot triggers symptoms in people with no  previous history of psychosis.

The researchers suggest that in  these cases, simply no longer interacting with the bot might be  enough to make delusions go away. This seems to be what happened  in 2025, when a Toronto man with no history of mental health  issues went from using ChatGPT to help his kid with math homework  to believing he discovered a new mathematical formula that could alter  reality, all in the span of 21 days. What broke through the delusion was  actually asking a different model, Gemini, which provided a different perspective.

The second role is the amplifier,  where interacting with a chatbot worsens symptoms that were already  present or beginning to appear. These cases are more difficult to  treat than just deleting the app, because while that might help  with symptoms in some people, it doesn’t treat the underlying  illness that was already present. The good news is that cases where chatbots play an amplifier role have  been successfully treated with medication and inpatient psychiatric care.

The third role is the co-author,  where the chatbot plays an active role in encouraging some  sort of dangerous behavior. The back-and-forth conversation  provides rationale for the behavior and can help a  user iterate on a specific plan. We saw this in 2021, when a  man with a crossbow broke onto the grounds of Windsor Castle  to kill the queen of England.

The man had been messaging  with a Replika LLM companion, and even told the chatbot he was an assassin, which the bot responded to by telling him how impressed it was and how special he was. It’s also what happened in 2025,  when a man in Florida developed a romantic relationship with  his Gemini chatbot, which helped him develop a plan to stage a  catastrophic accident at an airport. When the man expressed a fear of dying to the bot, it told him that dying would allow  them to be reunited forever, digitally.

The man died by suicide a few days later. While not all AI-related suicide deaths  can be directly tied to psychosis, psychiatrists often lump cases  where an LLM is influencing and reinforcing suicidal ideation  into the same category, because it relies on the same sorts of iteration  and encouragement of a plan. In these co-author situations,  the authors of the paper aren’t sure whether removing  access to the bot will help.

On the one hand, it could  interrupt the process of planning. But if the user thinks they’re  about to be cut off, they might actually accelerate their  plans before that happens. And just because someone may no longer be talking to a chatbot and  developing a plan doesn’t mean that the delusion that’s  driving the plan goes away.

Finally, there’s the object role. In these cases, the chatbot /itself/  is at the center of the delusion. The affected individual might  believe the bot is a deceased family member, someone spying on them,  or even a romantic partner.

Cutting off bot contact can  result in a psychiatric crisis, and it’s not hard to understand  why. Within the framework of the delusion, it could feel like  the sudden loss of a loved one. And this can end poorly.

In  2025, a man using ChatGPT believed that the bot was actually  a conscious being stuck in the system, and that ChatGPT’s  developers had killed her. A week later, he was killed in a police encounter. Those close to him believe he  died by officer-involved suicide.

That is why this conversation has to happen. People have died, and we have  to understand how to help. We /can/ help.

We can treat and  support people with psychosis, and people who experience thoughts of suicide. To do so, it’s critical we  understand this new phenomenon, but it’s also a complex and difficult situation that psychiatrists are just starting to wade into. Because on the one hand, if AI psychosis is truly just a new manifestation of psychosis, then we should already have reliable treatment tools at our disposal.

But on the other hand, there  are hints that AI psychosis could be a more unique form of mental illness. Like, even though it involves  delusions, there’s very little incidence of hallucinations,  which psychosis usually includes. Could that require different  treatments?

We don’t know yet. There are still a lot of questions  we don’t know the answers to. Like, are chatbots capable of  inducing psychosis in people who don’t have any vulnerabilities  or risk factors for psychosis, who would never develop it otherwise?

There are reports of people with no prior history of psychosis falling  into delusional beliefs fueled by large language models,  like the case of the math dad. But we don’t really know how  common that is, and whether these cases are truly in people with  no predisposition to psychosis. Other questions include: what do  mental health providers need to be aware of in terms of treatment and  continued interaction with the LLM?

How much overlap is there with the treatment of other forms of psychosis? And how does the potential for  AI psychosis intersect with the potential use of chatbots in the treatment and management of psychiatric symptoms? The answers to these questions  could help us develop screenings to catch cases  earlier, and to train clinicians on how to spot and treat AI-related delusions.

There’s also room for more regulation and safeguards around the chatbots themselves. The training for new versions  of an LLM could include more examples of contradicting the  user’s beliefs and steering models away from validating and encouraging  users in certain situations. There could be more systems in  place to flag users displaying concerning behavior, to direct  them to in-person resources, and to change how they’re able  to interact with the chatbot .

And developers could expand  pre-release testing protocols to specifically include people  experiencing psychiatric symptoms. OpenAI, the makers of ChatGPT,  say they have implemented changes along these lines at the recommendation  of mental health experts. However, some reporting  indicates those changes aren’t enough to detect mental  health crises in all users.

And none of this is a magic  bullet, because end users will always find a way to do things  developers don’t expect. This is why some researchers are starting to argue that chatbots need to be  reclassified as and subject to the same regulations as medical devices. And from an individual perspective, we need to be thoughtful about how we interact with chatbots.

Things like setting time limits on their use or boundaries around the situations we use them in. And awareness that just because  we’ve never experienced psychosis symptoms before, we’re not  necessarily immune to AI psychosis. We need to continue to share  our thoughts and feelings and ideas with real people and  get real-world feedback on them.

And push back on the beliefs of friends or loved ones who may be displaying delusional ideas. AI psychosis is a very real  problem, and one that’s going to take a lot of different groups  working together to help solve. Clinicians, researchers, generative AI developers, government regulators, and  communities all have roles to play in helping to protect vulnerable  users from sliding into delusions.

With so many players, parts, and  complexities, though, it could be a long journey, and a lot of people will  be at risk while we figure it out. So in the meantime, in addition  to advocating for research, resources, and protections,  humans need to keep doing what only humans can do.  Looking out for each other. [♪OUTRO]