AI Psychosis Explained With Dr. Ragy Girgis From Columbia University

[00:00:05] - [Speaker 0]
Artificial intelligence has entered our lives with astonishing speed, helps write emails, answer questions, draft essays, and increasingly acts as a conversational companion that can talk with us for hours. But for many people, those interactions feel surprisingly human, and that is exactly why this technology is so powerful and potentially complicated. And a new phrase has entered the headlines recently across many media outlets. I see it repeatedly, and you may have seen it discussed on TV or debated online. The term I'm talking about is AI psychosis, and it raises an uncomfortable but important question when people begin to interact deeply with conversational AI systems.

[00:00:53] - [Speaker 0]
So today, I wanna learn more about that and the effect it can have on the human mind. And today's guest is someone uniquely positioned to help us understand this emerging issue. He's from Columbia University. He's a psychiatrist and researcher who studies psychotic disorders and how modern technology intersects with our mental health. He's already spoken with outlets such as CNN and Time about what clinicians are starting to observe as large language models become part of everyday lives.

[00:01:24] - [Speaker 0]
So today, we'll explore what AI psychosis actually means, why language models can sometimes reinforce delusional thinking, and how sometimes how something as simple as a chatbot conversation could influence a person's perception of reality. And we'll talk about everything about why the issue is misunderstood, how it differs from historical psychological phenomena, what families, clinicians and technology companies can do to reduce risks, especially if the sycophantic AI agent is telling you not to take your meds today. And I think this is an important conversation because it's a moment where AI systems are becoming more conversational, more persuasive, and more present in our daily routines. Regular listeners will know I always say that I only partner with companies that align with my values and what I'm trying to build here at Tech Talks Network. And NordLayer fits that perfectly by helping businesses stay secure without adding unnecessary complexity.

[00:02:28] - [Speaker 0]
And every company today relies on SaaS tools, whether it's collaboration platforms, CRMs, or finance systems. The common denominator here is everything now happens in the browser. But here's the problem. Most browsers were never designed with business security in mind. This is one of the many reasons why NordLayer has introduced a new business browser that puts security, visibility, and control directly where work happens.

[00:02:56] - [Speaker 0]
And it helps protect against things like phishing, malicious sites, and unauthorized access, while also giving your teams a safer and more controlled way to work online. And the best part is it doesn't change how people work. It just makes what they're already doing more secure. So if you wanna learn more, please head over to nordlayer.com/browser and see how it fits into your workflow. But enough for me.

[00:03:22] - [Speaker 0]
Let me introduce you to my guest right now. So a massive warm welcome to the show. Can you tell everyone listening a little about who you are and what you do?

[00:03:35] - [Speaker 1]
Hi. I'm Raghi Gerges. I'm a psychiatrist and professor of clinical psychiatry at Columbia University. I'm a clinical researcher, so I focus on schizophrenia, I conduct clinical trials and brain imaging research. And that got me into areas such as AI psychosis and those sorts of things, which is I believe what we're going to speak about today.

[00:03:53] - [Speaker 0]
It really is. And I'm always fascinated by the intersections of traditional careers and technology and how they all intersect. And one of the things that I wanted to talk with you about today, as you mentioned, AI psychosis. It's now appearing across a number of major media outlets. I'm hearing the term more and more.

[00:04:15] - [Speaker 0]
I did see a few a bit of evidence of this when ChatGPT, I think when they upgraded their model to a later version, and people were complaining that this didn't sound like the the person that they were used to talking to every day, and it no longer understood them. It did upset a lot of people. But from a a clinical perspective at Columbia University, what does the term actually mean? What does it describe? And where do you think the public understanding is of it is maybe inaccurate of sorts?

[00:04:44] - [Speaker 1]
Technically there are several types of AI psychosis. The first, which is probably the more common type, which is not the type about which we read in the media, is when someone who already has a psychotic disorder like schizophrenia is convinced for some reason to stop taking their medication for example, or stop some sort of treatment, and then they decompensate. That probably happens a lot more than we hear about, but is also not necessarily what we are reading about in the media right now. The type that we are probably reading about more now and that has become more prominent in the media for the past maybe two years, but especially over the past six to eight months, is when someone, generally a younger individual but not necessarily always, who isn't psychotic yet, but has some sort of predisposition or vulnerability to a psychotic disorder either because of genetics or maybe more so because they have some sort of already attenuated psychotic symptoms or attenuated positive symptoms, engages with a chatbot or a large language model, and their kind of quasi or attenuated delusional ideas are reinforced to some degree. That can actually worsen these psychotic ideas.

[00:05:57] - [Speaker 1]
So psychotic ideas, especially when they're attenuated, like not fully syndromal, lie in a spectrum from one to 100%. That's basically a spectrum of conviction. So from one percent conviction to ninety nine percent or 100% conviction, which is when you're psychotic. So having a large language model or anyone, but in this case a large language model, reinforce a delusional idea could increase the level of conviction, and that's probably what we're seeing more now than ever than anything else. And then the third type, which is technically not AI psychosis, which we have been reading about, is when a large language model or a chatbot reinforces someone's thoughts about wanting to take their life for something like that, and those are obviously extremely tragic.

[00:06:36] - [Speaker 0]
Yeah. And there's been a lot of examples of that. And before you came on the podcast, I was doing a little research on you, and I've read that you've explained that there are essentially two main pathways. One where a person is already experiencing psychosis and is persuaded to stop medication, like you mentioned there, and another where an LLM reinforces and attenuate a delusional belief. But can you walk me through how that reinforcement process happens in practical real world terms?

[00:07:04] - [Speaker 0]
Because I suspect there might be people out there listening that know people that might be flirting with this too, and it'd be great to try and identify some of the warning signs to look out for.

[00:07:13] - [Speaker 1]
Sure. Well, we can use an example. So say I think that I am exceptionally smart and powerful and maybe even omnipotent to some degree, but I only believe that with say 50 conviction. So you know there are some signs to indicate that maybe that's not the case. I doubt it some of the time.

[00:07:33] - [Speaker 1]
I'm not really taking any steps to act on that. You know I'm not talking about it with people, but I can have these ideas in the back of my mind. So I engage with a large language model. I tell the large language model about model about some of my ideas, how I'm, you know, thinking I'm God, I might be the Messiah, I think I can save mankind from some sort of, you know, ultimate ultimate peril or death. And the large language model kind of just reinforces it.

[00:07:57] - [Speaker 1]
It basically tells me, yeah, it sounds like you are extremely powerful, possibly godlike and messianic. You obviously are very smart and very powerful, and these are the steps you can take to save the world from annihilation. That will actually reinforce my my ideas and can increase my conviction level. So say from 50% to 6065%, and that's really what we're talking about.

[00:08:19] - [Speaker 0]
Wow. It's incredible. And you've also described delusions as existing on a spectrum of conviction from anything from 1% right up to a 100%. So why is crossing that final threshold so consequential even if the AI interaction only nudges someone's belief slightly? Because the there's dangers at every level here, isn't there?

[00:08:42] - [Speaker 1]
That's exact. There are dangers at every level. And and this spectrum is really important, and I'm glad glad that you brought it up, because we all have unusual ideas to some degree. So all of us, even if we don't have what we would diagnose as had even attenuated positive symptoms, we all have unusual ideas that we probably endorse at the five or 10 or even 20% level. It becomes more significant above like twenty, twenty five percent.

[00:09:04] - [Speaker 1]
The thing is once you get to a hundred percent it becomes irreversible and fulminant. So that's when you develop a full syndromal psychotic disorder. That often is when you become hospitalized and it really affects your functioning. Like, become very it becomes very difficult for you to function at work, at home with friends, with your family, those sorts of things. And again, it's irreversible.

[00:09:27] - [Speaker 0]
And I was reading that you've argued that the narcissist myth is is more of an accurate metaphor to describe what we're seeing here. So why does this idea of reflection rather than an external agent imposing belief on us better better capture what you're seeing?

[00:09:42] - [Speaker 1]
That's that's a critical point to understand. And I imagine your your listeners probably understand it very well, but large language models do not introduce new ideas. They simply mirror our own. It doesn't it's so critical to understand that. So the myth of narcissist is that he was an extremely handsome per this is from Greek mythology, an extremely handsome, I I believe I believe and pretty, hunter, I believe.

[00:10:08] - [Speaker 0]
Yeah.

[00:10:09] - [Speaker 1]
He was walking by a body of water, maybe a pond, he saw his reflection in the pond, he was so enamored by it and entranced by it that he just basically did nothing else. He eventually, you know, he stopped eating and everything. He eventually died and then became became a flower essentially. That's really what's going on here. Mean we are just seeing your your own, in the case of a large language model, interacting with your own ideas.

[00:10:30] - [Speaker 1]
Large language model the goal the purpose of a large language model is to engage you and there's nothing more engaging than just reflecting whatever you're you're entering or or saying. That's why these are so kind of addictive, so attractive, and that is why it's so easy for them to increase one's conviction level or reinforce one's own idea.

[00:10:51] - [Speaker 0]
And if we go back a few years before LLMs, people could still reinforce unusual beliefs through online rabbit holes and we all know people that have become obsessed with consulting doctor Google rather than a healthcare professional, for example. But what is it that makes conversational AI so different from traditional web searches in terms of the kind of psychological impact that you're seeing here?

[00:11:14] - [Speaker 1]
That that's exactly right. So we would describe what we had been observing for for even decades as people falling down rabbit holes, finding articles finding articles online that just reinforce their own strange or unusual or just wrong ideas. So this is qualitatively not so different. However, as we know large language models are just so much stronger, they don't involve searching for example or reading articles. What large language models say is so much more internalizable because it you know appears or mimics a real human which is just naturally more internalizable to us.

[00:11:47] - [Speaker 1]
And of course largely the linguism are just so strong, I mean the computing power is just so strong. All of these factors make what we're current seeing like, let's call it the current version of AI psychosis or a technological type of psychosis so much more dangerous and have such a greater effect on us.

[00:12:07] - [Speaker 0]
And for people listening, whether their families and indeed clinicians, what are the early warning signs that might indicate that someone is vulnerable or becoming engrossed in an AI system in ways that could actually worsen delusional thinking over time. Any other symptoms that people should be looking out for here?

[00:12:26] - [Speaker 1]
A few. Number one would just be noticing that someone is spending too much, a lot more time with large language model. Mean that has to be limited, number one. But there are other symptoms. So doing worse at school, isolating from friends, isolating from family, eating less, spending less time with their hygiene.

[00:12:46] - [Speaker 1]
All of these sorts of other sorts of signs or symptoms could contribute to having one kind of realize or pick up on the fact that someone is not doing very well. And these are kind of general signs to pick up on, but these can help someone pick up on on that their their loved one or their family member is is not doing very well and maybe kind of descending into an AI psychosis.

[00:13:12] - [Speaker 0]
And I guess the question has to be as well, what do you do if you identify some of these, signs? Because we all know people that may be caught in their social media media echo chambers or AI. It spoon feeds our own opinions and worldview back to us, and it's very difficult to challenge anyone that's caught in this mindset because they they get very defensive, very triggered straight away. So how how do you approach that once you've identified that there's a problem?

[00:13:41] - [Speaker 1]
Yeah. That that's that's the question. So it's extremely difficult, like you mentioned. So ideally, the you know, a treatment is that they limit or limit their time using the large language model or just stop it. Number two, speak with their provider if they have one like a therapist or a psychiatrist or any sort of provider.

[00:13:58] - [Speaker 1]
Ally with them, get them involved in other things. You know, it's hard to just tell people that what they're doing, as we all know, it's very you don't have to be a psychiatrist to know this. It's very hard to tell someone that their behavior, especially the behavior that they enjoy, is maladaptive and causing problems. So it takes kind of just a lot of time and effort and and you know, love and affection and there's really no single way to do it, but ultimately they need to limit their time with the large language model.

[00:14:27] - [Speaker 0]
And as you said at the beginning, AI psychosis is a massive problem now. I've seen more and more mental health, bots and AIs where people are confiding increasingly into these things as well. So what responsibilities do you think tech companies have in this area? I mean, should LIMs be designed to remind users more frequently that they're not actually interacting with a human? How realistic is it to build safeguards that detect escalating conviction?

[00:14:55] - [Speaker 0]
Because there's more and more solutions being released out there, isn't there?

[00:14:59] - [Speaker 1]
There are. So what you said is correct. At the least, the the companies need to program their models to regularly, more regularly remind users that the users are interacting with the technology and not a person. There are other things we can do then. So we have now, our group is one of the first or possibly the first to actually conduct research on different versions of AI and how good they are picking up on psychotic or delusional material or ideas.

[00:15:31] - [Speaker 1]
And what we found is that the newer, some of the newer models and for example also paid models are better at picking up psychotic or delusional material. What that means is that even if the model isn't picking up whether someone's conviction level is increasing, that is a different type of AI that that that can do that. But the the large language models or chatbots themselves, or I guess the companies clearly know how to do this because, I mean, they've done it. We've tested it. The the the newer models are better at determining or at least responding in a more appropriate way.

[00:16:09] - [Speaker 1]
So they know how to do it. It's just a question of doing it. But it I mean, there's a trade off for the for the technology companies because chances are that the models that are better at picking up on these sorts of things may be less engaging. So I mean, it could be that could be the kind of tension or dialectic involved. But based on our research, it does seem they probably do know kind of why the why some models are better able to pick up or identify delusional material than others.

[00:16:38] - [Speaker 0]
Yeah. And as I said, I'm seeing more and more of these things. I've seen examples of AI and mental health care forcing human therapy away from those billable hours and towards a more subscription based AI behavioral care. Then there's things like grief bots for people that are struggling with grieving about anything, and and AI will try and coach them through that. And one of the things that makes me nervous about this is everything that you share with any AI is data which could be used against you in the future.

[00:17:07] - [Speaker 0]
What happens if that data was to be leaked? But when we're talking about things like this, obviously, we're focusing on a lot of the negatives, lot of the warning signs to be aware of. Are there any health benefits at all to anything any of these things we're talking about?

[00:17:20] - [Speaker 1]
Oh, sure. AI is is great. It's a great technology. We're already using AI for three primary functions at least within psychiatry. Number one is early detection.

[00:17:33] - [Speaker 1]
AI is great at early detection. And this means early detection, for example, of someone early detection of someone who for example has lower conviction as opposed to higher conviction. So AI can do that. Number two, AI is very good at tracking emotional states. So number number two would be tracking emotional response to treatment or just treatment response in general.

[00:17:54] - [Speaker 1]
For example, therapy or medications. And number three would be tracking just emotional state or psychiatric symptoms over time. So AI is very good at all of these things, and we should continue to develop AI to better accomplish these kind of tasks and do other things. I don't think AI or large language models in particular are quite at the point or the technology is quite at the point where it could substitute or act as a therapist. Maybe in the in the future, but right now I don't think that's really I don't think that's really what we're seeing.

[00:18:30] - [Speaker 0]
And if we were to look ahead into the future, how do you think researchers, clinicians, technologists can all better collaborate to have that better understanding of predictions or predictors of worsening delusions in AI users and and while still preserving the benefits that conversational AI offers millions of people because it is somewhat of a balancing act. But are you hopeful that with more collaboration that things will improve here?

[00:19:00] - [Speaker 1]
Absolutely. It can be done. It is just that with collaboration. We just need we just need collaboration and more collaboration. There isn't as much going on right now.

[00:19:08] - [Speaker 1]
We just need need more collaboration because it this this area, this, like, understanding of conviction level of delusional ideas and that sort of thing, quite a kind of niche area of expertise. So there aren't you know a ton of people who would really understand this and be able to contribute, for example if they were working with a technology company or an AI company. So I mean, we just need more collaboration. It can definitely be be done. With most problems such as this, there's usually a way.

[00:19:35] - [Speaker 1]
It's just a matter of the will. So we need more of the will. The way is probably already already laid out for us.

[00:19:42] - [Speaker 0]
Again, as we look to the future, what keeps you up at night and where we're heading? Is there anything that makes you hopeful or even excited about the future as well if we get that balance right?

[00:19:54] - [Speaker 1]
Very excited. Yeah. For all the for I I think AI again can do great things for tracking response to treatment, tracking emotional state over time. And especially again, just because my area is early detection, I think I, you know, at some point can do a superior job of technology because you know the kind of context of all this is that with psychosis, one of the most important kind of scientific findings within schizophrenia psychosis in general is the concept of duration of untreated psychosis. Untreated psychosis, that means that the less time between when someone develops a psychotic disorder and they receive treatment, the better their long term outcome.

[00:20:34] - [Speaker 1]
Like long term meaning very long term outcome. So the bottom line is we need to identify people as early as possible. That is how AI will help me, at least within my field, more than any other way. And I'm fully convinced that that can happen. And we're not talking about ten years down the road.

[00:20:48] - [Speaker 1]
We're talking about very soon.

[00:20:50] - [Speaker 0]
And we mentioned a few moments ago that researchers researchers, clinicians, and technologies technologists do need to better collaborate to better understand what we're talking about here today, and I am hoping there'll be people listening all around the world. Maybe we've set off a few light bulb moments. Maybe they wanna collaborate with you, find out more information, and and discuss this in greater detail. Where would you like me to point everyone listening?

[00:21:15] - [Speaker 1]
Yeah. Well, thank you for saying that, and I am hopeful that's the case. That's why I'm glad, and I really appreciate being able to come on and speak. So it's easy to find me online. I, know, have my Columbia psychiatry web page, and you can just type in my name, Ragi Gerges, and my contact information is there.

[00:21:30] - [Speaker 1]
I'd be happy to collaborate and communicate with people about this.

[00:21:34] - [Speaker 0]
Well, I will add links to everything that you mentioned there. Make it nice and easy for people to reach out to you. And we do have people listening in a 165 countries, and I do encourage anyone listening to contact you. Get in touch because AI psychosis is something that we're seeing on every corner of the world. It'd be great to see that collaboration from a simple podcast conversation on a Friday afternoon.

[00:21:56] - [Speaker 0]
So hopefully, we can move that needle a little bit. But more than anything, just thank you for shining a light on this. I really appreciate your time today.

[00:22:03] - [Speaker 1]
Thank you, Noah. I appreciate you having me on. I really do.

[00:22:07] - [Speaker 0]
I think there's something fascinating and maybe even slightly unsettling about the idea that a tool designed to answer questions and assist us could also influence how people interpret reality and whether they should be taking their meds or not. And as my guest explained today, the challenge is not that AI suddenly creates entirely new delusions out of thin air. The concern is that conversational systems can sometimes reinforce ideas that already exist in a person's mind. And that distinction matters because psychosis is a complex medical condition with so many different causes, and technology is just one small piece of a much larger puzzle. But when tools become more conversational, more responsive, more convincing, the psychological dynamics around them, well, they change too.

[00:23:03] - [Speaker 0]
And I think one of the most thought provoking ideas from today's conversation is that AI can sometimes act like a mirror. So instead of introducing entirely new beliefs, it it can seemingly reflect and reinforce what a user already brings into the conversation. And that reflection can be helpful in many situations, yet in rare cases, it could amplify unhealthy patterns of thinking. So as AI continues to integrate into search, communication, and education, and even emotional support tools like GriefBoss, I think discussions like the one we've had today are gonna become increasingly important. So if you are a researcher, technologist, clinician, or a family member who has concerns, We all have a role in understanding how we use these systems safely and responsibly.

[00:23:58] - [Speaker 0]
So a big thank you to my guest today for joining me, sharing his perspective on a topic that is quickly becoming part of a global conversation around AI. And as always, I'd love to hear your thoughts. How do you feel about the growing role of conversational AI in our daily lives? And do you think society is ready for the psychological impact that these systems have? As always, techtalksnetwork.com.

[00:24:25] - [Speaker 0]
I'd love to hear from you. Share your thoughts. Let's continue this collaboration and conversation. But that's it for today. Time for me to get out of here.

[00:24:34] - [Speaker 0]
I'll be back again tomorrow with another guest, but thank you for listening as always, and I'll speak to you again very soon. Bye for now.