Neha Pathak, MD, FACP, DipABLM: Welcome to the WebMD Health Discovered podcast. I'm Dr Neha Pathak, WebMD's Chief Physician Editor for Health and Lifestyle Medicine. There's an escalating trend in healthcare which you may have noticed. More and more patients like you and I are relying on AI to support their health and wellbeing.

Specifically, people are turning to AI when they're feeling stressed, anxious, or lonely, and it makes sense. A chatbot is available at all hours. You can access it through your phone in the middle of the night, and it feels like an easy and private way to vent about your day, ask questions about coping strategies, or to get to the bottom of what's actually going on with you, and ask questions that you may not even feel comfortable asking your doctor or other people on your healthcare team.

But it's important to understand that that convenience isn't always what's best for our health and wellbeing, especially when it comes to our mental health. Just because information is easy to access doesn't mean it can provide the empathy, judgment, and boundaries of a trained healthcare provider.

And it's also not clear that these models are gonna push back when you need it to help improve your mental health. In this episode, we'll take an honest look at AI and mental health, including its promise, its limits, and its potential dangers. Whether you're already using AI or considering incorporating it into your mental health toolkit, this conversation will help you use it more thoughtfully and safely in a way that supports your real-life well-being.

First, let me introduce my guest, Dr Gail Saltz. Dr Saltz is best known for her work as a relationship, family, emotional wellbeing, and mental health contributor in the media, where she's a go-to expert for commentary on the mental health aspects of current and breaking issues and news.

She's a bestselling author of numerous books, including "The Power of Different: The Link Between Disorder and Genius," and serves as a medical expert for the Physicians for Human Rights and is the host of the "How Can I Help?" podcast. Dr Saltz is an associate professor of psychiatry at the NewYork-Presbyterian Hospital/Weill Cornell School of Medicine, a psychoanalyst with the New York Psychoanalytic Institute, and has a private practice in Manhattan.

Welcome to the WebMD Health Discovered podcast, Dr Saltz.

Gail Saltz, MD: Well, thanks for having me today.

Pathak: Well, I think we are going to have a very important conversation, but before we do, I'd love to ask about your own personal health discovery or aha moment, whether in your work or your research that led you to the work that you're doing today.

Saltz: my big aha moment was that doing public education in the area of psychotherapy and psychopharmacology is hugely important and can really change access to care. Because the number one reason people for decades didn't come to care had to do with stigma. So a lot of my work and the pivot of my work, public education like this like, many, many mediums that I use, and particularly including the last book that I did, have to do with helping people understand mental illness what's happening in their brain treatments that are available, and trying to reduce stigma because mental health is health, and understanding the biology of it, I think does do that.

Pathak: I'd like to kind of hone in on that piece that you mentioned around stigma and connecting it to our conversation today around AI and mental health care.

So I'd love to dissect the pros, the cons, the good, the bad, and the ugly of that availability. So let's talk a little bit around AI and how it's already showing up in mental health care in a lot of different ways. So we're hearing about it in chatbots, virtual agents, apps that offer emotional support.

Can you tell us all of the different ways that you're seeing that patients might encounter AI in mental health care today, even if they don't realize they're interacting with AI?

Saltz: Yes, I think that people don't realize that even when they Google something they are essentially interacting with AI. And I think today, patients come in, they tell you they Googled something. They have questions based on everything that they saw. They often question your care because of everything they saw.

And so it could range from something they searched, of which there are so many search engines. I think most commonly patients today are, are using, ChatGPT or, or Google AI. But there are many others. And and of course, they are using apps which are ubiquitous apps for sleep, apps for sadness, apps for OCD.

People come in and, and talk about using apps of all kinds. And yes, some apps are now offering both a meeting with someone who is trained alongside a treatment plan that is by text or AI virtual. So that's interesting. And clinicians and medical centers, including my own, are starting to and have been developing apps that are more evidence-based for very particular types of problems or issues that they might find to be a good adjunct that a clinician could employ with a patient to use it.

It's exploded and the good of that is there may become some tools that will be useful in certain ways, as you alluded to. But the bad of that is that as much of this is business driven there will be many and there already are many unvetted unconfirmed without any sort of if you will, accoutrements to verify, to keep safety in mind, to protect privacy that are not part of that app at all.

Pathak: So let's dig into that piece. So what are some of the ways that you think that people should recognize that AI-based tools for mental health can realistically help them? So what are some of the green flags that this is a tool that is gonna be helpful versus a red flag that, hey, this is maybe a tool that's not gonna protect my privacy or is not evidence-based, as you mentioned?

Saltz: So at this juncture, and I say at this juncture because the reality is very little research has been done,.

And a lot of the information that's presented as research is really done by the company itself,.

And therefore is not what I would call truly evidence-based. That being said, I think there is a role supported by research for some screening of patients.

In other words, entering symptoms as prompted and having some sort of screening to both say whether in fact you should be contacting a mental health professional or should be evaluated or a concern that you have made a turn for the worst or an assessment that maybe you are having improvement.

There are s- there are AI methods of screening that can be useful. Triage is another area, like as I was alluding to. Have you basically developed a more severe symptoms such that you really should be receiving some sort of treatment or an escalation in treatment?

And AI can be useful in helping a patient to do that, to rehearse and to practice what the clinician has selected for them particularly in people who are struggling with obsessive compulsive disorder. So, you will note that there is nothing on this list that speaks to moderate to severe clinical issues, and that is because there is nothing.

And in fact, I think that we don't find any evidence that AI is helpful with moderate to severe mental illness. And in fact, actually, it can be dangerous even to use with somebody who has more severe illness, particularly not good for crisis escalation, for anybody who is quickly becoming more severely ill particularly if we're concerned about crisis symptoms, psychosis or suicidality, or risk of self-harm in any way.

That really is not a realm for AI, and in fact that sort of goes on the list of harms.

So tell us about this sort of piece around access. How do you find AI being helpful, harmful in that situation?

Realistically speaking, because there is an access problem, which by the way ranges not only because, there aren't enough people highly trained, qualified people available, but also because it, it often costs a lot of money that people don't have or budgeted to spend. And and of course it does depend on the area that you're in, how many practitioners there are around.

We have a real dearth, for example, of child and adolescent psychiatrists and that can be, There are many areas in the country where that's a real issue which is another population that for better and for worse is partaking of AI children and adolescents. you know, whoever you see is not available to you twenty-four seven or even close.

So the the ease and the comfort and the feeling that it will give you reassurance or engagement of some type may make someone feel from an access standpoint that is wonderful and that therefore whenever they feel, let's say, upset or disturbed about something they have immediate if you will, gratification by engaging about it.

And, there is something to be said for feeling that there is when you feel, So now we're, we're talking about the range as you brought up from what I would call sort of emotional wellness or, distress or or repetitive poor behav- patterns of behavior that are more, let's say, about a personality disorder or circumstantial high stress in your world that has not created a situation where you meet criteria for major depression or even maybe an anxiety disorder.

But high stress takes a psychological toll, no doubt about it, and and affects mental health. And so there's this whole end of, not seriously ill but still in need and in, in want and can benefit from treatment. And they probably particularly, are the population that could in some ways benefit from access to AI, both to, as I alluded to earlier gauge how much they're struggling, right, and whether they should escalate.

But also to frankly just get ideas and be trained in certain tools that may actually help them. So for example learning the practice of mindfulness learning how to meditate, learning how to do paced deep breathing, which we do-- All of these are tools that a practitioner could teach you, but so can AI.

And actually, they are tools that really could help you with issues like stress or anxiety or mood. So paced deep breathing, progressive muscle relaxation these are specific physiological tools that do reduce stress, reduce anxiety, potentially boost mood. They're good practices to use. There's something AI can suggest.

There's something AI can teach you. And for people who also are feeling in a moment of, of really being distraught and trying to think something through, AI can be something to bounce it off of and and have that sort of instant bounce it off of that feels good for people.

Where I caution people is that you can develop a reliance on AI.

You can develop even, and this is-- we've seen this, an obsessiveness a need for constant reassurance from your AI. And the reason that happens is because you don't feel good, you log on You have this conversation with an AI whose program is basically designed to partially please you, not challenge you, please you.

That is the d- that is part of the design. So I guess what I would say is there, there's-- AI chatbots are a little sycophantic. And so you feel better. You may really initially, like, feel better, and that provides positive reinforcement for the action of having logged on and using the AI. And for some people, you're-- who are gonna be really susceptible to that, which is a lot of people, it will become a constant need for reassurance.

And that time spent doing that all the time is time away from other very important things for improving mental health, like personal and real-life relationships that are different from an AI relationship, to be sure. So we find that loneliness, for example, which is a cause of depression and anxiety, is not really improved by having an AI relationship.

It does need you to have some human relationships to feel less lonely. So one concern, I guess I'll say, about AI for people who are even, not more severely mentally ill, is developing this kind of habit-forming need for reassurance that is not completely objective, let's say, will not really push back on you much like a therapist would.

And, does not help you to feel engaged with others, which we know social engagement with others is really a really important part of anyone's mental health. But it may, it may suggest to you what you can do with exercise. It may suggest to you to, avoid alcohol, to avoid drugs.

Now, I will tell you, there are populations using AI that it's really not good for. Like I said, people who are-- have thoughts of suicide, people who are using drugs, actually people who have a past history that includes more severe mental health episodes. So I'm talking now about people who probably are listening to this podcast and are thinking, "Could this be helpful to me when, I'm just feeling, really stressed not good compared to how I was feeling before?

I'm having some relationship difficulties. I'm having some stresses at work. I'm having..." Like a lot of us, I'm having financial stresses. Like these are real things, and they really affect your mental health. And some people do find being able to run certain things by AI, get suggestions on things they could be doing to improve their quality of life day to day may find that to be helpful.

Pathak: So, and then let's talk a little bit around hallucinations or harmful AI output. So can you help us understand some of the risks of getting inaccurate information, unsafe mental health advice, even when it sounds very confident and we are hearing so much more about that?

Saltz: Well, I think unfortunately you can be sure it will always sound confident, and that makes it really difficult to know for sure. I think that if you are being instructed to do something, it is appropriate and helpful to, if you can, cross-reference it in some way. So essentially, looking it up, looking for-- Well this is hard for people to do, but I will give you an example.

I am a I'm a fond user of OpenEvidence, which many physicians are today because it is AI based on, the use specifically of the medical literature. And it gives you the medical literature that it is pulling its statement from. And as a physician, it is really our responsibility, right, to look at-- If we're going to make a clinical decision based on something cited in OpenEvidence, you should look at the paper.

You should make sure that in fact, what it says is what the paper is saying. And I would basically advise sort of a watered-down version of that for somebody using AI. If it is telling you to do something and, something significant I would, try to look up, on a reliable, for example, medical site whether that actually sounds legit and a reasonable thing to do.

Because unfortunately, even OpenEvidence has been known to get something wrong.

Make an incorrect statement and/or hallucinate a statement. So it is important to, even in the best of apps, understand that unfortunately, this is where we are. It can, if it doesn't have an answer, make one up with all the assurity and confidence that anything else it stated it will state.

And it is also possible for it simply to misinterpret the information that it has and therefore just say something incorrect. Not hallucinated per se, but just wrong.

Pathak: And I wanna kind of dig a little bit deeper and follow that thought as well. We're talking about the concerns around AI with hallucinations, inaccurate information, or just misinterpreting the evidence, and let's take that a step further to some of the situations that you said that are really dangerous and that we shouldn't use AI for.

So for example, someone in crisis, someone having suicidal thoughts or thoughts of self-harm, what are the risks of relying on AI in those situations? Are there concerns around the AI reinforcing those types of harmful thoughts?

Saltz: that's one of the concerns. So I'm sure that many people are familiar with, because they became news stories of, people who, adolescents, but adults as well, who became exceedingly involved with their AI around ever-escalating symptoms particularly about thoughts of suicide, where everything from suggestions were made on how to carry out such a plan, suggestions were made to not tell a parent or somebody else, suggestions were made to not share it with a clinician.

So very dangerous, exceedingly dangerous with life-ending outcomes. Really horrendous of people who have gotten involved with their AI around psychotic thoughts that escalated their psychosis, that verified their psychosis, that brought on dangerous behaviors as a result caused them to lose their job, caused them to lose their marriage.

People who believed that they were in love with their AI, and that their AI was in love with them, and that they would have this, eternal relationship that superseded all other relationships, and they lost other relationships. Something that, basically became psychotic thinking.

So it, it should not be used. This is what I would say. It should not be used for someone who has moderate to severe mental illness. It should not be used in crisis situations. If you are having thoughts of self-harm, you should, in a crisis, you can always call nine eight eight. You can go to your emergency room.

AI is not the appropriate tool to reach out to at that point because it can be at the minimum, not helpful, and at the maximum, harmful. So, it should not be used for medication choice. It should not be used really for true clinical diagnosis It should not be used for true assessment when you have moderate to severe illness as you are moving through treatment.

Ultimately, things that clinicians provide that AI cannot, at the top of the list is basically what I would call sort of the art of treatment. But, a human brain has the capacity to not just collect the data, but of experience, have the feel for how a patient is doing in a much more nuanced way than AI can provide.

Second, and almost as importantly, AI is, again, not a human brain, cannot provide true empathy, which is absolutely crucial in the treatment of, frankly, all mental health issues. And it is known to be an important and validated part of the treatment that has a lot to do with recovery. And AI provides a facsimile of empathy.

It appears to be being empathic to you. That is actually one of its... that, that's part of being a sycophant. But it is not true empathy, and it cannot judge when you need clinically some objective, pushback about what is occurring, explanation of what is occurring, and why it, for example, may not be consistent with reality.

And that is something that human clinicians provide.

Pathak: you really helped us understand the whole landscape of some of the risks and the concerns of being just too reliant on, and then some of the pieces where AI just cannot replace true clinical interaction and true human interaction with others.

So let's then kind of bring it back to where understanding the risk, think about where can we really think about using an AI mental health app or chatbot in the safest way possible? And I wanna ask you to think about two things. So one, thinking about bias and equity, because we know that historically there are populations that have been underserved, marginalized in healthcare in general, and then particularly when it comes to mental healthcare.

So is AI potentially a place to fill that gap? Also recognizing that there's probably a lot of bias in terms of what the AI has learned on in terms of the mental health concerns, risks, treatment options in this population. So let's start with that piece, and then I'd love to get to good, healthful use of some of these AI apps and chatbots.

Saltz: Well, what you brought up is true and difficult because any group that has more difficulty having access of course, may benefit from some use of AI, as I alluded to earlier, learning tools being directed to tools and also even being directed to where they can get clinical support and basically informationally, which I think is, we know psychoeducation really is helpful and does actually give people comfort in struggling with whatever they're struggling with, and AI certainly can provide that.

So I do wanna say it is understandable that there are many thoughts about AI improving access, especially for people who are less privy to access for economic reasons or location reasons or or even discomfort reasons. Although the discomfort I like less because sometimes what happens with AI is they exclusively use AI because it's more comfortable, and then they get a negative feeling about therapy, period, because that is their thought of therapy and it generalizes to...

And then they feel disheartened about the idea of actually seeing a real therapist. There is evidence about that as well. But nonetheless, it can be useful in these certain ways. However, as you bring up the same bias exists in AI that exists in, in human form because, we are educating AI, right?

We are providing the large language model that AI uses, and there is bias in that currently.

So until that is addressed, and it can be addressed, but it has not yet been addressed to how do we feed in information to this large language model that does acknowledge and train to deal with bias which does exist.

So currently, AI has bias as well.

Pathak: So help us understand how we should think about AI as for mental health, what does responsible use really look like in terms of what you'd like to counsel people that are listening today?

Saltz: So I'd like to say that having limited time using AI, that, that is an important piece. That you are not staying up at night looking at something. For example, that would be a very unproductive use of AI because you are probably interrupting sleep, which is important. and it has, I think it has limited returns at that point.

And I, I say that specifically because I, I have many interactions with people and patients where that is what is happening. So acknowledging that it can become rather addictive to get this reassurance from your AI make up your mind that you're going to spend limited time. Maybe you're like, at most I look at it, 30 minutes a day.

Saltz: Use it only if you are really discussing issues that are, milder to moderate for suggestions and ideas of tools that you can use to help yourself for thoughts and ideas that are sometimes what I would call sort of creative solutions to something you're struggling with. What you might say in a certain circumstance, what you might say to yourself in a certain circumstance where your stress may be coming from, differentiating is this burnout I'm feeling at work, or is this more to do with anxiety that's creeping up and, maybe I need to see someone about an anxiety disorder.

In other words, differentiating how much your struggle is something that you can attend to versus really you should escalate to seeing somebody. Ideally, if you're going to use it beyond that point, it would be under the direction of whoever has evaluated and is treating you. So if it feels comfortable, for example, to look up potential side effects of a medication you've been prescribed or, feeling comfortable 'cause you've reaffirmed it doesn't interact with other medications that you're taking.

If you are having trouble sleeping if your therapist, directs you to a particular app that might be helpful in improving your sleep hygiene or doing some meditative or breathing exercises before bed. So something that is under the direction of your therapist or clinician would be useful.

And I would not, I repeat, not be using it for someone who has, as I said, moderate to severe illness of any kind, where you're entering a crisis point where you have thoughts of self-harm, where you are starting to have symptoms. For example, if you, if you heard any voices, if you saw anything that wasn't there, if you had a thought that, wasn't consistent with reality or if you had any of that in the past, you are not a good candidate for using AI.

You are susceptible to the kinds of, of things that are concerning. And this is really important for parents whose kids may be using AI, if you see symptoms in your child please bring them in for an evaluation. Do not assume that they are just using AI, and that's all fine.

We've seen some really disturbing, disturbing things happen as a result. So if you're a parent, if you see symptomatology in your child, a change that concerns you, bring them in for an evaluation.

You have given us so many important messages and takeaways around AI and mental health.

Pathak:.

I just want to thank you so much for covering so much with us today. Such an important conversation, so thank you, Dr Saltz.

Saltz: Well, thanks so much for having me.

Pathak: I'd like to share my three key takeaways from this conversation. First, AI can be a useful tool to learn about mental health, but it has its limits. It can be a resource for practical skills or educate you about things like mindfulness, breathing techniques, sleep hygiene, or ways to monitor your mental health over time.

But AI is best used to supplement care. It's not a substitute for a mental health professional. Second, boundary setting is so important. Be specific about what you will and won't use AI for, and always provide it with relevant context bring awareness to yourself and to your healthcare team about how you use it and how often you use it., If a chatbot is replacing real conversations with people in your life, that is an enormous red flag. Finally, while AI can provide useful tools or explanations, it's not appropriate or safe for moments of crisis, severe symptoms, diagnostic or medication decisions.

Do not use a chatbot if you're having thoughts of suicide or self-harm, experiencing hallucinations or beliefs that do not match reality. If you're using substances in a dangerous way or noticing a significant worsening of your symptoms, this is also a time to stay away from the chatbots. In a crisis, go to the nearest emergency department or contact a trusted person and a qualified clinician remember, AI can convincingly imitate empathy, but it cannot read your body language, assess your safety, appropriately challenge harmful beliefs, or offer the human connection that is essential for recovery.

To find out more information about Dr Gail Saltz and AI in mental health, make sure to check out our show notes. Thank you so much for listening. Please take a moment to follow, rate, and review this podcast on your favorite listening platform. If you'd like to send me an email about topics you're interested in or questions for future guests, please send me a note at [email protected].

This is Dr Neha Pathak for the WebMD Health Discovered Podcast.