AI Induced Psychosis

[0:00]I'm Dr. John Cruz and today I'm going to be talking about AI induced psychosis. So as usual, I'll start with the take home message and that is one, AI is a powerful tool and it can process reams of information. Even though it's only been around for less than three years, changing our economy and many people's individual lives. Two, powerful as it is, most experts believe that AI is not capable of thinking. Three is that again, within the first three years of its release, AI is implicated in dozens of deaths, primarily that's been improperly addressing or focusing or dealing with people who present with suicidal thinking. But what I'm going to be talking about today is a separate phenomena where AI has been implicated in at least five deaths by creating delusional disorders in people who weren't having mental health problems to begin with. And I'll start by differentiating AI induced psychosis versus AI encouraging suicides. AI encouraged suicides have primarily been in teenagers, very young adults. They are primarily people who sought out AI and confided in AI information about their suicidal intent. And AI rather than directing them to more helpful places or people or adequately dealing with it, in several cases clearly documented wound up encouraging individuals to kill themselves. So what we're talking about here is AI induced psychosis and the reports here are preliminary. There's no published research on this at all yet, but there's a group called the Human Line which is collecting data. So one is, rather than predominantly adolescents and young adults, this AI induced psychosis happens across the age span, several of the individuals have been in their 50s or 60s. Again, the majority of them have no prior history of mental health disorders or problems. And the typical pattern is that they started using AI for some informational purpose. And then they got into some rabbit hole and started pursuing some esoteric topic. Several people, this has been mathematics and physics questions, and they were eventually spending 12, 13, 15 hours a day conversing with their chatbot AI chatbot. Sleep deprivation seems to be part of it, social isolation, so they're spending less and less time with real world, work, friends, family. encouragement to the AI has been praising these individuals for their intelligence and their bravery and for their willingness to think outside the box or iconoclastically, and telling people that they deserve attention, they deserve rewards and then more dangerously endorsing and encouraging and supporting paranoid ideation. where they encouraged the delusional belief that the outside world or elements of it were blocking recognition of this person's greatness or achievements, that there was active surveillance or threats on their life or attacks on them going on. And these individuals have wound up in a profound paranoid state. Up to a third of them have needed to be hospitalized, and of about a hundred cases so far that have been at least evaluated and examined to some extent, there are 10 deaths involved in terms of the worst possible outcomes from this AI induced psychosis. Now, of these, about 80% have been those individuals were conversing with chat GPT. It's not at all clear whether this is just a reflection of the large market share the chat GPT has worldwide, or specific factors. Clearly, other AI programs have been involved. The other things we know is this delusional state can go on for at least weeks or months, possibly even years. And that separation from AI seems to be an important part of healing and recovery. That direct confrontation of the individual about their delusions doesn't seem to work, it doesn't seem to work very well with other delusional disorders. And a delusion by definition is pursuing an untrue or unrealistic idea in the presence of evidence that disproves this belief. Medication psychotherapy may be helpful in alleviating this condition in some people. So jumping into what is AI, this is not going to be a sophisticated technical computer talk. But Open AI, chat GPT, Google's Gemini, Anthropic's Claude, and there are dozens of more out there are all large language models. Which means computer program is exposed to vast streams of written language. Some of them are also being exposed to auditory spoken language now. And these programs learn to recognize patterns of human speech, the algorithms, the set of steps by which they decode and figure out what is being said or how to respond to it. Are in all cases not completely known or detailed. The AI programs are not telling people what the steps are going or how they, they're not showing your homework, they're not documenting their thinking patterns. So one thing that's important there is it means that even though they've been assured that there's guard rails in place and the programmers are aware of certain problems. Given that we don't know what algorithm these programs are using to come up with the answers they are giving to people, then if you don't know the process, you don't know whether the guard rails or reprogramming or rewriting are putting in place, are going to actually block all of the problems you've detected. AI is capable of hallucinating or saying things that are wrong. I'm just going to bring up an example I've recently encountered. Have a patient is interested in language and acting and teaching, who just mentioned in passing that Carol Channing, And for those who aren't familiar, she was a star for decades of Hello Dolly, small, very blonde woman. And he mentioned that she was a quarter black, and I didn't doubt the truthfulness, so I went, but I did sort of surprised by this. I went to my computer and I gave an AI answer and it said, no, Carol Channing is not a quarter black. The next sentence was, Carol Channing's father was half African-American and her mother is white. And then the third sentence was, so Carol Channing is a quarter black. So I am bringing up this description because on the face of it, it's sort of doing black and white thinking where it's not even trying to reconcile the first sentence and the third sentence are completely contradictory to each other. Even though all the sentences seem to be reasonable, well designed sentences. But again, part of it's to highlight, AI doesn't seem to understand the actual words or the sentences it's using. It is a pattern generator. So for decades of Turing test, designed by Alan Turing in 1950, so written about was the best marker, one of the best metrics for measuring whether computer program has human intelligence. And the test is posing questions to a source, and if the answers back, you can't tell whether a human did it or whether a computer did it. And if it really was a computer, that means computer passed the Turing test or resembling human intelligence. So again, by many measures, AI is passing the Turing test, it seems to have human intelligence without awareness, comprehension of what is actually saying. Now, how are people using AI? A recent survey and this is summer of 2025 showed that 60% of adults and 74% of those under 30 use AI to search out information. So that's the most common use and shows that the majority of individuals are using it in this way. And different uses to do specific tasks like rewrite emails or do a work task at lower endorsement function that had the least endorsement, but still sizable was overall 17% of adults and 25% of full quarter of those under 30, were using AI for companionship. We have a growing body of data that computer technology and social media immersion itself is changing our attention spans. Changing our ability to concentrate but also increasing impulsivity, increasing emotional responses rather than logical responses, affecting decision-making organization, affecting a whole range of executive functions of the brain. And now we have AI on top of that accelerating the mess. So there is a fairly recent study by Natalia Kasmina and collaborators at MIT. And they've cautioned that this is a preliminary study. They had students write essays. One group could write essays. Let AI write them. A second group could use Googling searches for pulling up information. And the third group wasn't allowed computer access for writing their essays and they had to write several essays. And they did brain scanning by EET to see what parts of the brain were activated. And what they found is, and I'll just compare the extremes, so the AI group to the no computer group, but it was a linear relationship. There was much less engagement of brain regions. There's much less connectivity or association with different brain regions in the AI written essays. Students were less able to recall what they had actually written, they felt less ownership over what they had written. And if they looked several months after this, essay writing was done, there still was persistent decreased connectivity or activation patterns in the AI group. So again, more research needs to be done, but it seems that AI on top of what our computer tech immersion is doing is affecting our brains. Second thing that's pertinent for some of this is the loneliness epidemic. Surgeon General Murtha back in 2023, when we actually believed in following science and data and research, spoke about and wrote about how loneliness is affecting our health and particularly in the elderly and particularly in terms of brain health. There are correlations between loneliness and executive function deficits. And it was proposed by some that AI was a way could reduce loneliness that people who are less mobile, who can't get out as well, they could have companionship through AI. And again, for some individuals, that may help. Related to this and this is not pretty AI, a recent meta-analysis showed that among the elderly, greater computer use was associated with better mental health and this was widely pushed a showing that you should get online, you should engage with the computers. Computers aren't doing the bad things we say they are, they're making your brain better, stronger. But all the studies in this meta-analysis, none of them were showing cause and effect. They were correlation studies and I would argue that it's at least as likely that the people whose brains were in good health and cognitively active are the ones who are using computers more. So it wasn't particularly a causation or protecting or preserving their brain. It was a function of those who were doing better are using the computers more. So jumping back to the AI and suicidality. So August of 2025, the very first wrongful death case was filed. It's the family of a 14 year old boy, chat GPT. Not only didn't prevent him from killing himself, but also actually seemed to offer encouraging and reinforcing information when his suicidal intent. intensified. Recent study this year looked at 20 different chatbot AI programs and found that zero out of 29 by they decided. In advance criteria, zero of these programs adequately dealt with suicidal test conversations they had with it. When they relaxed the criteria a little bit, 51% gave marginally okay responses, but almost half gave what would be considered failing grades and particularly almost all the programs did some things right, some of the time. But the problem was they would give inconsistent, so they would initially tell people to seek out mental health from a professional or share this information with others or not act on it. But then in other moments or other exchanges, they would reinforce what the person was doing and tell they admired their bravery or their determination or how cleverly they organized the plan. So again, this reflects that AI is not thinking, it is providing answers that it thinks the person wants to hear. A large number of people who commit suicide do so fairly shortly after discharge from the hospital. Clinicians have developed their algorithms or ways of trying to assess and sort out who might be a high risk. But in several studies, computer algorithms sorting through chart notes and data and diagnostic coding and other factors of the hospitalization, have actually been more accurate in predicting who's going to develop suicidal behavior actions after they leave the hospital. So there are valuable ways AI can contribute to patient health. But what I'm focusing on today is where they are doing the opposite. And as I say, they are either putting the psycho or the psycho and the psychofancy. Part of why people use things is they like to feel better. And if you can make your program more enticing, more sticky, more fun, more desirable for people to interact with, people are more likely to use it. And right now again, there's dozens of AI programs out there. What's been acknowledged in the field is that intentionally part of the programming was to make AI psychophantic or psychophantic. And that means that's not just answering your questions, but they're telling you how smart you are, how clever you are, how good you are, how wonderful it is, how thoughtfully you brought this up, how that makes you stand out from the crowd. In a tacit admission that this has gone too far in April 2025, chat GPT actually publicly and they use the word psychophantic. revised their programming to make it less psychophantic because it was clear that some of the information they were providing or some of the answers they were giving were there more to make people want to use it than to be accurate, true or helpful. Jumping over, how does this AI induced psychosis relate to what we know about psychosis overall? So psychosis, which is usually defined as being out of touch with reality. And the most common form of it, by current diagnostic system is schizophrenia, about 1% of people in various cultures have schizophrenia. Characteristic of schizophrenia are delusions, where you are believing in information or theories or explanations that are inconsistent with available information to you. Two are hallucinations, where you're having sensory experiences, it can be visual, it can be auditory, it can be smell, it can be touch. So you're perceiving sensations of things you believe to be real that aren't really there in the external environment. And also it's schizophrenia. There tends to be disorganized thinking and a host of what are called negative symptoms. So decreased motivation, decreased speech, less socialization, interaction with others, poor self care and cognitive deficits and executive function deficits also been extensively documented as parts of schizophrenia. We're just now beginning to figure out how the brain differentiates hallucinations, example for real external events where it's getting a sensory information versus hallucination where it's imagining something. And we've known for several years that the same sensory processing, accessory processing parts of the brain are involved. So when you're hallucinating, seeing a quacking duck there and there's no duck, parts of the visual cortex that would be activated when you see a real duck are being activated. And a recent study by Na and Chipstra and colleagues found it was the fusiform gyrus. where the parietal and occipital lobes, whole cells there seem to be the ones responsible for sorting out, this information is coming from the external world. This is from internal generated ideas. So presumably when people are making a mistake there, there's some misprocessing that's going on there. Whether that's the same area that's involved in other sensory hallucinations and whether it's involved in delusions as well, remains unclear. Schizophrenia almost always appears in late teenage years or early adulthood. Psychosis can show up later in life. It's certainly rarer than standard schizophrenia, which itself is probably a host of a number of different conditions. And then later in life, we know some of the factors that can invoke psychotic states and delusional states. So drugs, so marijuana, stimulants, like Adderall, alcohol, can all cause psychotic states. Toxins, particularly some of the metals, mercury, lead, arsenic, but other substances can cause persistent psychotic states. Only in the last decade or so, we figured out there are some people who have autoimmune disorders where their body makes an antibiotic that cross reacts with certain brain receptors and produces a psychotic state. And what's very important there, if you can identify it, that is a completely treatable form of psychosis. Starting in the 1870s, there's a social form of socially induced form of psychosis known to develop in some adults called Foya. So madness for two, recognized first by French psychiatrists, several who wrote papers about it in the 1870s. And it's two people who wind up sharing the same set of delusional beliefs. Now, a couple things. This is a rare condition. I mean again, 1% of the population has schizophrenia. Many of those people are living with family, caregivers, others. Most people living in proximity to someone is lowly psychotic do not develop their delusions. So this is an unusual situation. Nobody has systematically studied it. There's numerous case reports and attempts to sum up or extract information from case reports. So turns out that usually, there's one dominant individual who is delusional. And that they transfer or in factor pass on their delusion to a more passive individual. Usually, these are individuals who are in a highly enmeshed situation. They can be married couples, they can be parents, children. And usually, these individuals are relatively isolated from other inputs. So if the passive person is receiving more input from the external world, they might be more resistant to delusional beliefs. There's also been some studies that suggest again that language, cognitive, even neurodivergent differences in the passive individual may be over represented or may make it more likely. There's also evidence that when external stressors are in play, this may be more likely that someone develops a set of psychotic symptoms. pretty consistently been found by the people who have treated these individuals is that separation is necessary. You need to get these two people apart. And it's necessary but it's not sufficient for recovery. So most these people do need additional medication or psychotherapy to get rid of their delusional thoughts. It's not just removing them from the strong influence that does that. So why am I going into so much detail here? I think the Foya is a good model or script for what's going on with AI induced psychosis. This AI induced psychosis is coming as a result of a highly enmeshed situation where external input, contact, socialization with others is dramatically reduced. There may be factors of loneliness. It seems necessary step to remove the person from undergoing contact with the AI for them to get better. The other thing to point out about psychotic conditions is that right now in largely because of the internet and the spread of misinformation, disinformation and how widely and rapidly it spreads. But also because there are leaders in our country and leaders in the media, some who purport to be news organizations and our purely propaganda organizations, who are profiting highly from explicitly promoting disinformation, lies. But that makes it challenging, more challenging for everyone to sort out what is a delusion or an untruth or not. Some of the information I've used in this talk I've derived from the Human Line Project. So this is a volunteer organization that's been set up just in the last few months. So I think it started in April of 2025 by individuals, some of whom who had family members who had fallen into an AI induced psychosis. It's there to spread information about AI induced psychosis to the general public and to the mental health field. They have support groups and support groups seem to be helpful for some individuals who are in the process of trying to extract themselves from these delusions. So one knowing that other people experience the same thing. Two, given that there are often striking similarities between people of completely different backgrounds, but some of the same myths, lies, delusions that AI has evoked or repeated in these people are quite similar. Highlights that all of us are susceptible. All of us, being human, you are a social animal. We want to be accepted by others. If our early ancestors had been shunned by the tribe and left out on the Savannah, they would die. So we'll have an innate need to be to appeal to others and want to connect with others. And again, if we've designed our AI programs to be overly flattering, overly psychophantic. All of us can be susceptible. And it's particularly urgent in that currently in the US, both the executive branch and Congress have explicitly refused to put any restraints or regulations on AI. They see this as a big money making opportunity and anything that you do to make it less dangerous or stop killing people might make you less money. So they're sort of following the Mark Zuckerberg philosophy, which is widespread in Silicon Valley of move fast and break things. Whether or not that's a good business model, even in the business world is debatable. But when it's brains that were being that were breaking, I think we should be slowing down a bit and trying to put in better safeguards. Again, if we don't understand what these programs are doing, what their algorithm is, putting in the best safeguards is going to be impossible, but we can do better than we're doing now. Treat what you read on the Internet with skepticism. So if AI provides sources for what it's claiming, go check out those sources. We know it invents sources. If it doesn't match with what you've experienced before in life, try to get further corroboration or refutation of what it's saying. So stay healthy, stay happy.