Latest / The Medical Mind / Mental Health Pathfinders: What Psychiatrists Need to Know About AI with Dr. John Torous

Transcript

Mental Health Pathfinders: What Psychiatrists Need to Know About AI with Dr. John Torous · The Medical Mind

• 0:05
  Thank you for joining us for this episode of Mental Health Pathfinders.
• 0:09
  I'm your host, Aaron Connors.
• 0:11
  Technology continues to have a profound impact on our everyday lives.
• 0:15
  It's hard to imagine life before smartphones,
• 0:18
  laptops, and the many apps we rely on every day.
• 0:21
  Technology is also at the forefront of healthcare,
• 0:24
  making this an exciting time of innovation and opportunity.
• 0:27
  At the same time, it's important to recognize the benefits while remaining mindful
• 0:32
  of the potential risks and challenges.
• 0:34
  With that in mind, the APA has a new Council on Digital Health,
• 0:38
  Innovation, and Technology. And joining me now is the Chair.
• 0:41
  Of the New Council, Dr. John Toros.
• 0:43
  Dr. Toros is also director of the Digital Psychiatry Division in
• 0:47
  the Department of Psychiatry at Beth Israel Deaconess Medical Center.
• 0:51
  That's a Harvard Medical School affiliated teaching hospital,
• 0:54
  where he is also serves as a staff psychiatrist and associate professor.
• 0:59
  Thanks for joining us today, John.
• 1:00
  Well,
• 1:01
  Thank you so much for having me.
• 1:03
  let's jump right in. Can you tell us more about the new council and you know some
• 1:07
  of the work it's already kind of starting to do?
• 1:09
  Yeah.
• 1:10
  And I think the APA has always had a leadership role in technology.
• 1:15
  This is a a new council, but it really grows out of a lot of different efforts that
• 1:19
  were happening across the APA.
• 1:21
  The APA has always been leading in telehealth,
• 1:24
  setting standard standards, advocating for reimbursement.
• 1:28
  We've always done a lot of work in mental health apps.
• 1:31
  We helped write the APA's app evaluation framework that kind of put the
• 1:34
  AP at the center. So the idea was we're doing all this work in telehealth.
• 1:39
  Doing all this work in apps. Clearly,
• 1:41
  AI is the new friend or enemy in the room.
• 1:44
  Well, we can go into it. We said,
• 1:46
  That's right.
• 1:46
  look, this is really rising to the level of a council.
• 1:48
  Let's put all of these under one umbrella and bring onto it.
• 1:52
  So, really, it's a new council.
• 1:53
  So, in kind of jargon y APA governance terms,
• 1:57
  it's new, but really it's the extension of great work that the APA
• 2:01
  has been leading for a long time with now an AI spin.
• 2:06
  Absolutely. And you know what as we said,
• 2:08
  this is such an exciting time.
• 2:10
  Is there anything you particularly are excited about when it comes
• 2:14
  to technology and innovation? When it comes to healthcare?
• 2:16
  I mean, yeah.
• 2:18
  I I think that we of course have seen medical records for a long time,
• 2:24
  and we sometimes like medical records,
• 2:26
  we don't like them. We've seen smartphone apps,
• 2:28
  sometimes patients use them, sometimes they don't.
• 2:30
  What's different about AI is the rate of use by both patients
• 2:36
  and clinicians has been unparalleled,
• 2:38
  right? You you've never got a technology where both groups are saying
• 2:41
  we we're we're using it, and we're using it at rates that are well over
• 2:46
  I think there was an APA survey in February,
• 2:49
  and I think it was about 50% of psychiatrists were using AI as of February.
• 2:53
  I'm gonna be bold and say now that we're talking about later,
• 2:56
  it's well over 50% of psychiatrists are using it.
• 2:59
  There's also been a lot of kind of different poles of patient populations,
• 3:03
  and probably at least a third of people are using it.
• 3:06
  And again, we never quite got that with apps.
• 3:08
  EHRs were kind of forced upon us by our employers in that,
• 3:12
  but this is spontaneous, right?
• 3:13
  And and I think
• 3:16
  When you see something that rapidly growing,
• 3:18
  that spontaneous, there's something that we as psychiatry have to chase down there
• 3:23
  and make sure it works well for our field because it's coming and it
• 3:28
  you could say in some ways it already came.
• 3:32
  And thanks for
• 3:32
  mentioning the poll. We did that with WNBC in in New York.
• 3:35
  That's right.
• 3:36
  And ~ we we had over two thousand members respond to that.
• 3:40
  And ~ what some of our members we asked them,
• 3:44
  you know, to we w to find out what members know about AI and what their concerns
• 3:48
  are. Eighty percent are concerned that psychiatrists don't have enough training
• 3:51
  ~ when it comes to AI. Only eighteen per percent felt very informed about AI,
• 3:57
  and forty percent have ~ not used it in their practice at all.
• 4:01
  It
• 4:02
  sounds like we have some some work to do here.
• 4:04
  Yeah. And in some ways I can I think I always hear people say,
• 4:07
  Well, AI is changing, John, so quickly.
• 4:09
  I what's the point of training?
• 4:11
  It's gonna be different next month.
• 4:13
  And I think my counter to it is yes,
• 4:15
  it's always changing, but everyone has access to these AI models.
• 4:20
  If you have a computer, you can test them.
• 4:21
  And again, you can put in hypothetical information.
• 4:24
  Do not put patient information in.
• 4:26
  We all realize that, but you can kind of test them yourself and get a sense
• 4:29
  of what it is.
• 4:30
  You can ask them questions about therapy and safety say you can
• 4:33
  ask them questions about medication and see what they come back.
• 4:35
  Doesn't again, so I I think that in some ways the best training in a world like this
• 4:40
  is to be trying it and say, there's a new model that came out.
• 4:42
  I've heard my patients are asking about it or using it.
• 4:46
  What does it give back? And I I think there's something about that genuine hands-on
• 4:50
  experience by trying it that really lets you again understand what you
• 4:54
  see as the risks and benefits,
• 4:55
  what you want to talk to patients about,
• 4:57
  because
• 4:58
  When people say, well, this AI model is all bad or all good,
• 5:02
  that's for their experience, for their thing.
• 5:04
  And I think we've seen that there is a risk benefit,
• 5:08
  there's balance, and that's what we as psychiatrists are trained to do.
• 5:11
  We we know there's no one perfect treatment.
• 5:13
  We evaluate the risk, we look at the benefits,
• 5:15
  we work to the person in front of us,
• 5:17
  we say, Hey, this makes sense to use or not sense to use.
• 5:19
  So I'm not saying we should use an AI for anything clinical,
• 5:23
  but we can still use that kind of clinical decision making.
• 5:26
  help people understand and the only thing one has to do if listening is just
• 5:29
  go play of these things, try them.
• 5:31
  Try different ones. See what they say.
• 5:34
  You you may be horrified. You may be impressed.
• 5:37
  But
• 5:37
  well, on a bright spot over half of those polls said they see promise
• 5:40
  in AI for treating things like anxiety,
• 5:42
  depression, and substance use disorder.
• 5:44
  So so there is a little bit, they're seeing that there there could be,
• 5:48
  you know, in conjunction with the work they're already doing with their patients,
• 5:51
  there could be some benefit here.
• 5:53
  Yeah. It's and and I think in part as the models probably become more powerful
• 5:59
  and read more of the internet and learn more,
• 6:01
  they're gonna hopefully have more benefits.
• 6:03
  New risks will emerge too that we don't know about.
• 6:05
  So I think that we can't give any of them a free pass,
• 6:09
  but I think I always like to say like I think what a model could do this year versus
• 6:15
  last year, it's only been one calendar year.
• 6:17
  It's been like 10 years in terms of research advancement.
• 6:19
  And I think that's the exciting.
• 6:21
  Hard in some ways. And why it's a bit of an intimidating space is that
• 6:25
  if you thought that the models did something dangerous,
• 6:27
  made mistakes, they did. You are correct.
• 6:29
  But the question is, do they still do they introduce new ones?
• 6:33
  So I think it's a field where I always say when I give talks on it,
• 6:36
  I have to update my talk every 30 days because the whole evidence has changed over.
• 6:41
  So it it's it's tricky, but the basic principles of risk and benefit don't change,
• 6:47
  right? The basic principles of
• 6:49
  What is its appropriate use case?
• 6:51
  The tool may evolve, but what its clinical applications are given
• 6:54
  is something that a group like the APA is going to help set the standard for.
• 6:59
  Right. And because things are changing so much.
• 7:02
  Of course, you know, there's concerns about AI for therapy.
• 7:04
  We know there's a shortage of psychiatrists and other mental health providers
• 7:08
  out there. So you know, it's understandable that people will look
• 7:11
  to other places for help there.
• 7:13
  What's your advice there?
• 7:15
  I think if you look at some of the newest meta-analysis that came out
• 7:19
  a couple weeks ago, it seems like when you use an AI for therapist kind
• 7:23
  of self-guided with without a clinician or psychiatrist,
• 7:26
  it probably maybe has the same effect size as a mindfulness app,
• 7:30
  as a CBT app. So I don't think anyone ever felt that mindfulness apps were
• 7:35
  the be all and end all or the future of treatment.
• 7:37
  They have a role. There's some really great mindfulness apps out there.
• 7:41
  I recommend patients use them to augment it.
• 7:44
  But there's really not the data that an AI therapist kind of delivers anything above
• 7:48
  that level. It's even more interesting.
• 7:51
  Recently, as the models again have evolved,
• 7:55
  there's some early data, and it's it's early that says,
• 7:58
  well, maybe can the big models like Claude or Google or Chat GPT,
• 8:03
  our data is becoming so big, they can offer the same type of conversations that
• 8:07
  a mental health-specific AI does.
• 8:09
  So we we heard a lot about people saying,
• 8:11
  Well, I have my own mental health AI.
• 8:13
  It's been trained on conversations and the kind of debate in the field now is well,
• 8:18
  have the big models just caught up to it.
• 8:21
  So so like, do you need, do you even need a specialized mental health one?
• 8:25
  Maybe, maybe not. We'll find out.
• 8:28
  So I think that's where it's useful for just a psychiatrist to test them,
• 8:32
  play a bit, see what you think is the best response.
• 8:35
  It's not what Aaron and John think is the best or what a company tells
• 8:38
  you is the best. It's you say,
• 8:39
  look.
• 8:40
  These are kind of my personal tests and cases and see what they respond back.
• 8:45
  And one thing I'll finally say that is like we ask every patient in our practice,
• 8:51
  do you use AI? And we actually frame it as we think a lot of people are using AI.
• 8:54
  It has some interesting benefits and drawbacks.
• 8:57
  Can you tell us about what you're using?
• 8:58
  We almost invite people to tell us what they're using.
• 9:01
  Cause some people actually said,
• 9:03
  Well, I didn't want to bring it up to you or my therapy team because
• 9:08
  I thought I would.
• 9:09
  So it's a case where the patient wants to protect a relationship by
• 9:14
  not kind of threatening the therapist or making the therapist
• 9:17
  or psychiatrist feel hurt. So in essence,
• 9:19
  Interesting. Wow.
• 9:20
  like so if you open up a space and tell the patient you're not gonna hurt
• 9:23
  my feelings by saying you're using AI,
• 9:27
  I will make a bet for zero dollars because I'm not a betting person.
• 9:31
  But I would say about four in ten patients will say,
• 9:34
  Yes, I have used AI something.
• 9:36
  And by asking just that simple question,
• 9:38
  what are you using AI for, you will get a very good sense of what are
• 9:42
  the pros and cons. And then you,
• 9:43
  of course, can use AI yourself and say,
• 9:45
  is it is that a good answer or a bad answer?
• 9:47
  So you don't need to be an AI expert,
• 9:50
  you don't need to be on our AI counsel.
• 9:52
  We welcome your feedback and help.
• 9:53
  You just need to ask your patients,
• 9:55
  which we're very good at asking people in psychiatry,
• 9:57
  what are you doing? And just test it out yourself.
• 10:00
  And you would be surprised in a couple weeks,
• 10:02
  you will be considered an AI guru by your patients.
• 10:06
  You will know what to do, what to talk about,
• 10:08
  and you'll have fun with it, or you'll be in control of what's happening.
• 10:13
  you ~ I I'm sure you probably are you concerned about this technology
• 10:17
  it being a little bit too much and people are straying away from
• 10:20
  the human interaction more and more now you know that they aren't they're relying
• 10:25
  on that more than having that that relationship with someone it could
• 10:29
  be anyone it also could be their therapist or psychiatrist
• 10:33
  I think right the key to all health,
• 10:35
  physical and mental health, is good relationships.
• 10:37
  There's been wrong, but I think it's the Harvard happiness study that's kind
• 10:41
  of followed people for almost 50 years.
• 10:43
  It was like what makes people happy and have fulfilling lives.
• 10:46
  It's relationships. It's what it comes down to.
• 10:48
  It's not money, it's not status,
• 10:50
  it's not neighbor it's it's good,
• 10:52
  solid relationships. And I think that it's very tempting for AI to kind
• 10:58
  of for people to form relationships with AI or us to form relationships with.
• 11:02
  With AI, I I I think in some ways that's where the biggest harm happens,
• 11:08
  especially for our patients. There's some fascinating research out of Stanford.
• 11:13
  And what they did is they asked people who had experienced serious harms,
• 11:16
  up to including death from suicide by AI,
• 11:19
  and they said, Can you or your family give us the chat logs of the chatbot that
• 11:23
  led to these again adverse events?
• 11:26
  And what they found was the average chat log in these adverse events,
• 11:30
  the number of messages back and forth,
• 11:32
  was over 5,000 messages.
• 11:35
  So this is not kind of how are you doing today?
• 11:38
  No, good, bad. This is a often multi-month thousands of interactions back
• 11:45
  and forth. And there's clearly some evidence that when you begin to kind
• 11:49
  of have these super in-depth conversations,
• 11:51
  you build a relationship with the chat bot.
• 11:55
  You go thousands of messages in,
• 11:57
  the safety guard rails, the chatbot itself seems to get confused.
• 12:00
  And this could probably happen with any company.
• 12:02
  Would happen with a mental health chatbot,
• 12:04
  it would happen with a general one,
• 12:05
  but just they're not designed.
• 12:07
  They haven't even been safety tested.
• 12:09
  Because I think honestly, no one thought that people would go down five
• 12:13
  or 10,000 messages down it. And so I think what I always tell patients is.
• 12:20
  Reset the memory from time to time.
• 12:21
  If you notice that the chat bot,
• 12:22
  if you kind of have a personal feelings towards it,
• 12:25
  reset the memory. If you notice you're going beyond a thousand messages,
• 12:29
  reset the memory. If you're being to have romantic feelings to it,
• 12:32
  you could you could reset the memory.
• 12:34
  And I think that's probably the best thing for AI safety we can do.
• 12:37
  That said, a lot of companies now automatically turn on the memory.
• 12:40
  They want it to know you, they want to have that relationship.
• 12:43
  But I guess I'm here saying the number one risk factor for serious harm.
• 12:48
  Is long, long conversations. And again,
• 12:51
  so why do people have long conversations?
• 12:53
  They may be having a relationship.
• 12:54
  Some people have sexual relationships with a chat bot.
• 12:57
  They may think it's has superpowers,
• 12:59
  but what whatever leads you down that train track faster,
• 13:02
  that is where it's you're gonna get the most dangerous.
• 13:07
  I'm not saying they're all perfectly safe.
• 13:08
  I'm not saying they're good for short conversations,
• 13:10
  but definitely those long ones are pretty concerning to do.
• 13:15
  And I'd almost say in the short term.
• 13:18
  They're getting better and better in the short term for not causing harm
• 13:20
  the chatbots. If you kind of pressure test it and say,
• 13:23
  Well, give me ways to do things I shouldn't do or things that be harmful,
• 13:26
  they're pretty good at saying no now today.
• 13:28
  Not always, but it just feels
• 13:30
  And it sounds like this speaks
• 13:31
  to the issue of loneliness for all ages these days.
• 13:33
  Yeah.
• 13:36
  It exactly. And and so I just think they they're and they can
• 13:40
  be useful as companions too, right?
• 13:42
  We're not saying they should never be used.
• 13:44
  We're not saying this is all bad,
• 13:45
  but it's just it's to be careful what happens what when they go
• 13:51
  too far down that world. And just imagine again the chatbot doesn't
• 13:56
  it just loses track of its own reality.
• 13:58
  And that's why you end up with these weird messages or you kind of
• 14:02
  see these things on TV and you go,
• 14:03
  How did that happen?
• 14:05
  It's just been there. So so I I so point is I think relationships are very powerful.
• 14:12
  They can be helpful probably from a chat bot for some people and but some
• 14:15
  can be toxic as we know in the clinical world.
• 14:19
  But I think in our case, very long chatbot messages are assumed to
• 14:23
  be the greatest risk.
• 14:26
  Do need more safeguards and oversight when it comes to this?
• 14:27
  so good.
• 14:30
  I I think that we need certainly groups like the American Psychiatric Association
• 14:34
  to bring partially the values of of what we do.
• 14:37
  I think we we want to make sure that the chatbot is serving the patient.
• 14:43
  It's not goal is not to serve more ad revenue.
• 14:46
  It's not to keep you going on to sell you something.
• 14:50
  I I I do think making sure the values come into it.
• 14:53
  And I think that psychiatry, I think we work for our patients.
• 14:57
  I think we do a very good job for that.
• 14:58
  And we we bring those kind of values to it.
• 15:01
  I think that will guide the safety very well.
• 15:04
  Cause again, the chatbots in some ways are reflections of what we train it on,
• 15:07
  what we ask it to do. So I think no chatbot will ever be a hundred percent safe
• 15:12
  because they're based on probability.
• 15:15
  At the end of the day, they kind of don't always do the same thing.
• 15:18
  They're a little bit different.
• 15:19
  But I think that you really want to have one that is guided by your best interests
• 15:24
  in mind. And that then the safety will follow from it.
• 15:26
  You know, it won't be perfect.
• 15:28
  But I do think.
• 15:30
  That is where I think there's a lot of codes of conduct that we have in the APA.
• 15:33
  I think again, we all have a patient doctor relationship.
• 15:36
  It's even legally protected to some extent,
• 15:38
  right? We have our and I think that's what's been missing in the space.
• 15:42
  Any chatbot you're using for mental health now,
• 15:44
  as of the time of this recording with Erin and I,
• 15:47
  it's either a wellness device or a consumer device.
• 15:50
  It is not offering clinical care.
• 15:52
  And they try to be deceptive and claim it's offering clinical care.
• 15:56
  But if you read the fine print,
• 15:57
  it is not offering you a patient doctor.
• 16:00
  Relationship at that point. And that does make a tremendous difference,
• 16:04
  even just for the confidentiality of your data.
• 16:07
  So I always tell my patients, you can do what you want with a chatbot.
• 16:11
  My advice is it's not private,
• 16:13
  so you may enter information into it.
• 16:15
  You have no idea where it's gonna pop up next year,
• 16:17
  next month, in what context. And that company has no obligation to do it.
• 16:22
  I guess I think some of them are now working to respect HIPAA and do more things,
• 16:26
  but I would be.
• 16:29
  I would caution patients to be careful.
• 16:30
  Clearly, as clinicians, we cannot upload any patient information into them.
• 16:34
  We should not be doing it. And again,
• 16:37
  no one's asked no one's also telling us to upload patient information too.
• 16:41
  I think so that's where we have to be just careful.
• 16:44
  And that gets back to we we should test them,
• 16:46
  but again, make up your own hypothetical scenarios and test them with those.
• 16:51
  Privacy is a big issue, and I I would think for parents it's it's very important
• 16:56
  to kind of keep an eye on what your kids are doing on these things.
• 16:59
  Yep. And I think in some ways there's a lot of software to keep an eye
• 17:03
  on what your kids are doing or not.
• 17:04
  Kids are smart. They're probably smarter than all both of us combined,
• 17:08
  Aaron, trying to stop kids from using it.
• 17:10
  No doubt.
• 17:10
  We're gonna find a way around it.
• 17:12
  And there was this paper that came out in BMJ in a journal this week and it said,
• 17:17
  you know, Australia banned smartphones or social media use for young people.
• 17:21
  And they said, you know, in the group that was fifteen to age sixteen,
• 17:25
  I think, they said, you know, we looked at the first three months of the ban.
• 17:28
  Social media use went up in that group.
• 17:30
  And you go, whoa, whoa, whoa, we just said a goal is to ban it for young people.
• 17:34
  How did it go up? So the point is,
• 17:37
  as soon as you tell the young people it's like they shouldn't use,
• 17:39
  they're smart. They find ways around it.
• 17:42
  So I think relying on technical safeguards alone is not going to ensure safety
• 17:48
  for anyone. And it's very tempting to say,
• 17:50
  well,
• 17:52
  The company now says they a filter and young people can't use it,
• 17:55
  or I as a parent have this system around it.
• 17:57
  But I I think I would just assume people will find a way to use it.
• 18:01
  And so it's probably better to have an open discussion about it if possible,
• 18:05
  or try to set some family guidelines around it,
• 18:08
  because I think relying on technical protections alone.
• 18:11
  It's not a bad thing. But if we're saying in the whole country of Australia,
• 18:15
  social media use seems to be the same or increasing despite it being banned
• 18:19
  for young people.
• 18:22
  Something ain't working there.
• 18:24
  Well, it's such an important time for this council to be coming together right now.
• 18:28
  Anything else that you're working on right now or kind of looking at to focus
• 18:31
  on in in the coming year or years?
• 18:34
  I think one
• 18:35
  thing is clearly education for psychiatrists about this.
• 18:39
  We said the field moves quickly,
• 18:40
  but not that quickly, that we can't keep track of what it is.
• 18:44
  So I think that we realize that APA members and psychiatrists
• 18:48
  are hungry for accurate, up-to-date information,
• 18:51
  and that's something that we are going to be delivering on.
• 18:54
  I think we also know there's a lot of policy changes.
• 18:57
  Our policymakers want to make safer,
• 19:00
  more accessible, affordable care.
• 19:01
  And in different states.
• 19:04
  There's different approaches to it.
• 19:05
  Some have a more conservative approach to chatbots,
• 19:08
  some have a more liberal approach to it.
• 19:09
  And I think that these are new.
• 19:11
  Mental health chatbots have probably been around for two years,
• 19:15
  maybe two and a half years, in the form we're talking about.
• 19:18
  Everyone is learning and new to it.
• 19:20
  So I think that our group of the council is to work with APA local branches
• 19:25
  and chapters and help make sure that state regulation makes sense.
• 19:28
  It's safe for patients.
• 19:30
  It's evidence-based, it's effective,
• 19:32
  it aligns with what is best for patients.
• 19:34
  So I think we want to do a lot of advocacy on the policy level and make sure that
• 19:38
  we shape this in the right direction.
• 19:40
  And then again, we've done a lot of work on helping to smartphone
• 19:45
  app evaluation standards. And I think we can bring the same things to AI.
• 19:48
  We're not going to say what is the best AI.
• 19:49
  I'm not going to say what's worst AI.
• 19:51
  It's like me saying, what is the best medication?
• 19:53
  For who? When? You evaluate your patient.
• 19:56
  You understand that.
• 19:57
  But we can help bring some of the background information,
• 20:00
  some scaffolding, some questions to think about which AI maybe makes sense
• 20:03
  to use and which one doesn't, or maybe we should never be using them
• 20:06
  for certain things. So I think education,
• 20:09
  policy, and kind of benchmarking is more the technical word,
• 20:13
  but evaluation are gonna be the three initial pillars that the council
• 20:17
  is focusing on. That would say AI does change rapidly.
• 20:21
  We want to hear from you. If you say amazing things,
• 20:24
  let us know. If you see things that are terrifying,
• 20:26
  we want to know that.
• 20:27
  Two, there's a lot of just it's amazing time for
• 20:33
  AI in both a good and bad way.
• 20:35
  And I think if the APA doesn't know it,
• 20:37
  we can't help you with advocacy,
• 20:39
  we can't help you with standards,
• 20:40
  we can't help with education. And so I would say reach out,
• 20:44
  and again, our council is easy to find an APA website.
• 20:47
  It's you can talk to your district branches,
• 20:51
  but I think bring stuff forward because we're in a brand new era.
• 20:55
  And we're all seeing different parts of it.
• 20:57
  But I think what you're seeing on the ground is gonna matter the most
• 21:00
  and that's what we have to make sure we're responsive to.
• 21:04
  Dr. John Torres, this is important work that you're doing,
• 21:06
  and we really appreciate you taking some time to join us here today.
• 21:09
  And to our listeners,
• 21:10
  Thank you so much. Bye.
• 21:13
  you can find more episodes of Mental Health Pathfinders on your favorite podcast
• 21:17
  platforms.