Is AI Triggering Anxiety and OCD? — Ep 11

Elias Aboujaoude: I’m Dr. Elias Aboujaoude. I’m a clinical professor of psychiatry at Stanford University. I’ve researched the technology psychology interface for almost two decades now. I also run the program in Internet Health and Society at Cedars-Sinai Medical Center.

Stephanie Hepburn: I was reading your paper in Nature by you and your co-author, Ashley Golden. You published your perspective on how AI chatbots maybe quietly reinforcing anxiety and OCD patterns. Can you share what that means?

Elias Aboujaoude: Yeah, so as part of running the OCD clinic and the anxiety section at Stanford, we’re interacting with patients all the time. And it has become more evident, I would say in the last six to eight, 10 months, something like that, that has become more evident how their interactions with general purpose chatbots may be reinforcing some anxiety tendencies that they already had, or in some cases may be causing de novo anxiety symptoms to manifest. So a couple of common examples, I would say, would be reassurance seeking and perfectionism. Now, in someone with OCD, for example, reassurance seeking is something they engage in to deal with doubt. Doubt is a common OCD obsession. OCD has been called the doubting illness. And if you’re working with a therapist doing what’s called cognitive behavioral therapy, which is a form of therapy that’s been proven and established as an effective treatment in OCD. And if you had doubt type obsessions, then your therapist would prevent you from seeking reassurance or would help you place limits on how much reassurance you’re allowed to seek. Because ultimately the goal is that you can calm the anxiety down on your own using your own resources and the tools that you’ve picked up in therapy and not rely for reassurance to come from an outside source. Well, what’s happening with a lot of the general purpose chatbots that are increasingly used now is that they’re endlessly good at offering reassurance and they engage you in more and more reassurance.

Stephanie Hepburn: Because there’s this pervasive adoption of the general purpose AI chatbots, it’s changing how people are seeking psychological support.

Elias Aboujaoude: Yes, a staggering proportion of the general population, something like 70 plus percent, seem to be going to general purpose chatbots for emotional support, psychological support. So these systems practically overnight became the biggest purveyor of psychological and emotional help. And to the extent that the media have covered them, to the extent that research studies have been interested in them, it’s really mostly to focus on the extreme cases, like tragic suicide type outcomes, homicide, psychosis. But what we’re seeing, and what’s really much, much more pervasive, is the subtle effects that are taking place. And these usually involve reinforcing anxiety tendencies, obsessive tendencies. And that’s what’s happening at a much, much grander scale, but that has gone essentially unnoticed and unaddressed up to this point.

Stephanie Hepburn: And so you’re saying that these adverse effects are more widespread than the rarer cases that we’re hearing about. You’re not trying to undermine those cases, but you’re saying, hey, there might be these other issues that are transpiring that we really need to be paying attention to. When you’re talking about more people turning to these general use or broad use AI chatbots, what is it in the behavior or in the interactions with the AI that is problematic? Is it because it is trained to be sycophantic? What is happening and transpiring that is negative reinforcement?

Elias Aboujaoude: Yes. Well, several things are happening. You mentioned sycophancy, which is a major problem. And it refers to this tendency that uh general purpose chatbots have to be agreeable, to really want to please you. And if you were someone who had anxiety or depression, who was working with a therapist, perhaps in a cognitive behavioral therapy mode, then the therapist is often confronting and challenging and very much disagreeing with what we call faulty cognitions, faulty assumptions that you have about yourself or about certain situations that come from this anxiety or this uh depression condition.

Stephanie Hepburn: Can you give a couple of examples of what that may be and then what the chatbot might be doing?

Elias Aboujaoude: Sure. So if you have perfectionistic tendencies as part of either diagnosed OCD or subclinical OCD, I’m thinking back at a patient whose perfectionism always showed up in his travel plans. So he travels a lot for business and would spend three to four hours in the old days, meaning pre-ChatGPT planning a business trip, because he had to choose and to guarantee the absolute best possible routing from point A to point B, best departure time, best connection city, best price, et cetera. Well, ever since he discovered chatbots, this process has exploded into something like 10 to 12 hours. And the reason is that the chatbot, unlike the partner or the work colleague or the travel agent, is built to engage you endlessly and to help you optimize for better and better outcomes. So this is someone who in therapy would have been challenged by his therapist around this perfectionistic tendency, but there’s no one to challenge him. And indeed, the very opposite of that happens when he’s interacting with the chatbot. Another very common example that we’re seeing is doubt. People who have doubt, maybe because they have OCD, which has been called the doubting illness. And as part of dealing with their doubt, they have the compulsion of reassurance seeking. And if you’re working with this person in a therapy model, you would very much try to limit this tendency and to control it, to contain it, to put parameters around it. Well, LLMs are always available and uh were built in many ways to engage with you open-endedly and provide you all the reassurance that you can ask for. So for someone with this particular issue, instead of dealing with the anxiety that comes with the doubt and learning how to confront it and waiting for it to dissipate, they end up engaging in very time-consuming and ultimately self-defeating reassurance using chatbots.

Stephanie Hepburn: You mentioned earlier that it’s not only that it’s exacerbating people’s symptoms, but also triggering new ones. Can you flesh that out? What does that mean?

Elias Aboujaoude: As someone who has researched the technology psychology interface for almost two decades now, I very much have flashbacks to what we were seeing in the early days of social media, meaning folks who may have had some narcissistic tendencies or maybe some grandiose tendencies, but nothing out of the ordinary. And then they discovered social media and in some cases ended up with full-blown narcissistic personality disorder because narcissism was one of those personality traits that very much were exacerbated and worsened by social media. And something similar, I would say, is happening with chatbots. So people with uh perfectionistic tendencies, people with doubt who need a lot of reassurance, these tendencies are being also exacerbated because what ends up happening is they get this immediate relief from the chatbot. And uh, of course, the relief doesn’t last. And next time this doubt creeps in or this perfectionism creeps in, nothing really will calm it down until they get the same level of immediate reassurance as they got from the chatbot. So this ultimately causes these tendencies to get uh to get worse and to be reinforced.

Stephanie Hepburn: In the paper, you highlight how general purpose chatbots might drive negative reinforcement and people with OCD and anxiety disorders, but you also highlight that even evidence-based AI tools, they too carry their own risk for also doing negative reinforcement.

Elias Aboujaoude: There’s certainly some early research that suggests that in some cases, maybe chatbots that were especially built for mental health can be helpful, but we’re not there. So I don’t want listeners to assume that field of specialty chatbots is now established and ready to be rolled out. What’s necessary is uh what’s needed is something along the lines of a scale that keeps in mind and that measures these tools on some essential criteria. And we tried to develop a scale myself and my colleague Ashley Golden that we called uh FITA, the framework for AI tool assessment and mental health. And it uh walks the user or the clinician or the developer through six major criteria that any tool should be measured on. And those include the evidence base. Has this tool been proven to do what it claims to want to achieve? Uh, crisis management, how does this tool handle psychiatric emergencies such as suicide, such as homicide, such as emergence of psychosis? How transparent is this tool? Do we know who’s behind it? Do we know who’s funding it? Do we know what their conflicts of interest might be? What’s the user experience like? What’s the user agency like? How equitable is it in terms of it having been studied across user profiles? So we very much need to keep these criteria in mind, in my opinion, as we navigate this brave new world, including and especially when it comes to tools that purport to be specialty tools that purport to be developed with mental health specifically in mind.

Stephanie Hepburn: So clinicians need to be proactively asking their patients and asking their clients whether they’re using AI. And I’ve seen this anecdotally mentioned by clinicians that I’ve interacted with. And that is also, like you said, it’s a new space and it’s changing so rapidly. And just recently am I starting to hear clinicians say, hey, we need to ask our clients and our patients whether they’re using this.

Elias Aboujaoude: Yes. And I remember back to when we first started training our residents and future psychiatrists about how important it is to ask their patients about whether they’ve been Googling suicide. You know, in our field, we’ve always asked patients with depression and patients at risk for suicide whether they had firearms at home. But then when online searches became what they became, we incorporated a question in our mental status interview about online suicide searches. And I feel we’re at a similar sort of inflection point with general purpose chatbots, where it’s really imperative for clinicians to inquire about this dimension of their patient’s online life.

Stephanie Hepburn: So let’s say thinking about the example that you mentioned of somebody who is using a general broad-use AI chatbot to figure out their business travel, like you said, to get from point A to point B or whatever that might look like. What sort of limits could that individual put on themselves to make sure that it doesn’t take on a life of its own?

Elias Aboujaoude: The first step is for us to educate ourselves about what it means to be interacting with a tool that is so sycophantic, with a tool that is available around the clock that gives us instant relief in some cases to sources of anxiety, with a tool that’s absolutely frictionless, right? Like we can be discussing things in ways that are unhelpful, but there’s no social friction that is introduced to interrupt that discourse and interrupt that conversation. And last but not least, this very sort of self-confident, very authoritative voice that these tools have that make them sound like they can never be wrong. Really understanding these properties and questioning them, I think is crucial to maintaining psychological balance as we interact with them more and more. I think asking yourself what purpose or what goal you’re trying to meet through these interactions. Is it for work purposes? Is it for entertainment and research, or is it for psychological and emotional well-being? And if it’s for psychological and emotional well-being, ask yourself whether it might be more productive and more helpful to interact with an actual mental health professional. Also think about some properties that the tool has that you can use to your advantage. For example, instruction protocols that you can type in, just like you type in your prompt, in which you instruct the tool to limit its interactions with you around certain topic if you start sort of getting caught in these uh endless loops, uh, to maybe give you reminders when you’ve gone a bit too long without taking a break in a particular exchange with the tool. So use some of these properties to help contain the experience and to help keep it safer.

Stephanie Hepburn: So, what about the flip side? So the intent of these broad-use AI chatbots, it’s not mental health. That’s not the priority, but it doesn’t mean they aren’t working on these issues. But just as we are more aware, like you mentioned, of these more extreme cases associated with the chatbots, this may be information that is new to them as well. So, what would you say to those AI companies? Right now, we don’t have federal guardrails, but what should those companies be aware of and what can they do to not unintentionally cause people harm?

Elias Aboujaoude: I think they should try to be good citizens. You’re right that the lack of any established legal standards and guardrails so far complicates the process and makes it essentially entirely voluntary for developers to uh do the right thing. But I would invite them to be good citizens and think of ways to, for example, reduce sycophans, to reduce this over-agreement that the tool has with users, to introduce maybe some friction in the interaction with users if it goes on in very circular, repetitive, endless ways, to think of ways to maybe provide gentle prompts when a particular exchange is feeling again repetitive and uh unproductive. So I think these are all things that are well within developers’ abilities to uh to introduce. And I would love to see more of this voluntary kind of effort being uh deployed by them.

Stephanie Hepburn: And I think anybody who’s used these tools is very aware of the sycophency and also the overconfidence, not stating that it’s wrong, that can cause all kinds of problems in work and in personal life. But when you mention friction, what do you mean exactly?

Elias Aboujaoude: So, if, for example, to go back to the reassurance seeking example, that’s a common one in OCD and in anxiety, if you’re struggling with doubt, uh, doubt about whether you locked your door or turned your stove off because you have OCD, and you’re repetitively asking your partner to reassure you that yes, indeed, you did turn the stove off, and I saw you do it. Well, if you ask this question three or four times, at some point your partner is gonna stop answering, and that the social interaction is gonna start feeling overwhelming to the recipient, but also to you as the repetitive kind of questioner, right? It’s gonna feel awkward, it’s gonna feel embarrassing. Well, this awkwardness and this embarrassment simply don’t exist on LLMs, which means this particular type of interaction can go on without any impediments, without any friction open-endedly. So, this is what I mean by introducing friction in these types of unhelpful interactions.

Stephanie Hepburn: So, what further studies need to be done? And what are you working on next?

Elias Aboujaoude: Moving forward, we need to be working on this sort of in parallel. On the one hand, we absolutely need to continue to better understand those more extreme cases that lead to suicide, that lead to homicide, that lead to emergence of psychosis or mania. But alongside that, I would love to see more attention being paid to the masses who don’t fortunately end up with these tragic outcomes, but whose psychology is being manipulated almost sort of insiduously by these tools with hardly any research to the state into them. And in terms of uh personal research projects, it’s very much along these lines. Uh, we’re trying to understand de novo emergence of anxiety.

Stephanie Hepburn: And de novo being new?

Elias Aboujaoude: New, yeah, new and folks who never had it and now have it perhaps as a result of these types of interactions. Another somewhat related area that we’ve also gotten very interested in is the anxiety that’s really quite prevalent, especially in Silicon Valley, where my Stanford practice is based, anxiety around professional displacement by AI. I see it all around me among brilliant computer scientists, but also uh, you know, colleagues who are in radiology or any number of fields, and their anxiety is real, and it is reaching levels of what we would call a disorder. And very little research has been done into that as well. It’s the start of something that is going to be more widespread. Yes, we are seeing patients who never had to be in therapy or never had to go on medications, really never had a diagnosis mental health-wise of any kind, who now are showing up in our clinics because they’ve either been displaced by AI or they strongly believe it’s only a matter of time before they are. And what am I supposed to retrain in? What’s gonna be around in you know, six months? It really, these are really big questions that many people are starting to ask. And I think it’s only a matter of time before they become more pervasive.

Stephanie Hepburn: It is very hard to retroactively make changes. And so I think bringing awareness not only to the adverse uh potential effects of something that we haven’t really been paying attention to, which is how does this exacerbate anxieties, symptoms of OCD, and like you said, existing ones, but also triggering new ones. So that, like you said, that really needs to be examined. But when we think about the things that really have an impact on someone’s mental health, it’s relationships, it is employment, it is quality of life. And these are not issues that these very rapidly moving, and again, they’re so pervasive, I think even consciously making a decision not to use them is almost impossible at this point.

Elias Aboujaoude: And I think that will be even more true in five months, you know, a year. And this is exactly the time scale we’re looking at. I mean, things are moving so fast. And as you know, the pace of research has always been slower than the pace of technology, but we are so much more aware of this fact with AI than with any prior technology wave, really. You know, it leaves us without answers to many very legitimate questions that patients ask of us, but also the culture at large is asking. And what I hope culture will do with AI is learn from mistakes in a way that prioritizes research more than we ever did with earlier waves of internet technology and try to draw the lessons from that experience because they’re important and valuable lessons and they’re available to us now if we’re interested in heeding what they have to teach us.

Stephanie Hepburn: Thinking about these parallels that you mentioned, it’s such a smart comparison because it feels so novel. It feels so new, and it is, but it’s not the only time where we have had something rapidly adopted that is novel. You mentioned earlier the adoption of the internet in general, social media. What are some of those red flags that you’re talking about? What are those that we can pay attention to? Because again, while this is new, the experience of quick adoption of something novel in the tech space is not.

Elias Aboujaoude: Our first study, and it was the largest study into what was called internet addiction back then, came out in 2006. So that’s 20 years. And you’re right, AI is like nothing that preceded it. But to me, the parallels, you’re know, a brand new technology that people are absolutely enamored with and adopting blindly without really asking themselves the kinds of questions that we were talking about in terms of effects on mental health, and while ignoring some red flags in terms of consequences to psychology, like something similar might happen this time around, where psychology is not prioritized, where you bringing up potential negative consequences is equivalent to you saying that America should lose the AI race, or you know, you want the economy to crash or whatever. I mean, these are all very old arguments that we we heard before, that I personally heard before, and uh there are echoes to them in today’s discourse. So again, learn from those lessons, prioritize mental health and research, and invest quickly in research because it has to happen very soon if we have to have any hope of truly understanding the full consequences of what we’re talking about. But that research is still in its infancy, and we need to do a lot more. Lastly, on the regulation side. Again, to this day, there’s very little by way of regulation, and we’re hearing the same exact arguments that we heard when we were calling about some form of regulation of social media and warning about cyberbullying and online aggression and lack and the effects of living in a post-privacy world and all the things that social media have had such a negative impact on. So regulation is another very interesting parallel where we can do differently and we can do a lot better than we did before.

Stephanie Hepburn: That was Elias Aboujaoude. He heads the Internet Health and Society program at Cedars-Sinai Medical Center. I’ll include the link in the show notes to his paper on how AI chatbots may be quietly reinforcing anxiety and OCD patterns. If you enjoyed this episode, please subscribe and leave us a review wherever you listen to the podcast. It helps others find the show. Thanks for listening. I’m your host and producer, our associate producer is Rin Koenig, Audio Engineering by Chris Mann, Music is Final Couch by Blue Dot Sessions.