Tristan Glatard: Can AI Help With Mental Health?
AI could expand access to mental-health support, but a convincing conversation doesn’t establish that a chatbot is safe.
0:000:00
About this conversation
Speaking with Ayush Prakash, Glatard distinguishes clinical tools, wellbeing apps, and general-purpose chatbots. He explores how AI could help match patients with treatments, what happens when a support conversation goes wrong, and why patients should help shape these systems.
His example is someone struggling with addiction late at night, when their usual support may be unavailable. He sees a role for evidence-based tools under human supervision, while stressing the need to evaluate their effects and respond when a conversation becomes unsafe.
“I think the right approach is how can we expand the support that human therapists provide with AI, just because there's not enough therapists in the country to provide support to everyone.”Tristan Glatard · 13:47
Follow the conversation
AI and mental health at CAMHWhat do we mean by mental health?How technology can help personalize treatmentAI beyond chatbots: wearables and predictionThe case for an AI companionDo people open up more to chatbots?Where AI mental-health support breaks downBalancing opportunity, safety and health equityCrisis detection and data privacyThe AI healthcare conversation we needWho supervises AI support?A surprising research discoveryWhat keeps a researcher goingWhat gives Tristan hopeWhere to follow Tristan’s work
Full transcript
Ayush Prakash
Let's start at where you operate. What do you do on a daily basis and what are your perspectives on artificial intelligence with mental health and with health in general?
Tristan Glatard
No, thanks. So I'm leading a research center at the Center for Addiction and Mental Health in Toronto. So that's a hospital that treats patients for various psychiatric, mental health, addiction conditions. And the research center that I'm leading is not a clinical center, a clinical research center. It's a center on neuroinformatics. So very much focused on how can informatics, computer science, data, AI, all these technologies help to ultimately treat mental health and addiction. So yeah, that's what I do. So I lead a team there. I have my own smaller research group in it. And yeah, so we see, you know, maybe to start with your question about AI, so we see AI as both a set of opportunities. There's tremendous ways that AI can improve the life of people with mental health issues or addictions, but also a set of risks. And so we are very conscious about this trade-off, this balance. And I guess, you know, our goal is to strike the right balance between opportunities and risks.
Ayush Prakash
Yeah, absolutely. Thank you so much for that. I don't want to talk about AI immediately. I want to talk about what mental health is. For my entire life, my generation— I'm Gen Z— has been talking about mental health, and it feels like, uh, it's ironic, a figment of our imagination. I've never been given a full definition or even a full understanding of what mental health is. Can we just go back and forth on what mental health is? Where does it start? Where does it end? Where does the boundary of mental health lie? What does it mean to suffer from mental health?
Tristan Glatard
Sure, sure, we can do that. So I should also say that my background is in computer science, so I'm very much a technical person. So, you know, take what I'm going to say with a grain of salt. I'm pretty much going to repeat what I remember my clinician colleagues from CAMH often say. So mental health impacts the whole lifespan. But it has a particular importance in youth because many of the mental health conditions start in youth. So youth is a very important time for mental health. Ultimately, the deepest issues develop into conditions like schizophrenia, depression, bipolar disorder, like all these things have a very important impact on people's lives in many respects. Addictions as well, I think, are maybe a bit easier to grasp, but there's obviously very deep impacts on people's lives. And yeah, so there's Many approaches to try to help people. I think when people come to the clinic, they're already, you know, facing some very serious conditions. But I think it also starts with, you know, maybe more light things such as, you know, well-being and, you know, coping with stress and, yeah, these kind of things.
Ayush Prakash
How are we using technology, whichever the technology is, to help treat something that's— can I say fundamentally internal? Mental health only exists within the mind. Technology exists without the mind or outside of the mind. How do we use technology, something external, to treat something internal? Where does that balance lie?
Tristan Glatard
Yeah, I know that's a good question. I think there's, you know, maybe maybe 2 aspects of this. So one is that, you know, one of the grails, so one of the things that we are trying to do is precision mental health. So that means adjusting existing treatments or new treatments to the individual situations of people that people face. So this is, you know, that means essentially, you know, Collecting data, integrating data, understanding what makes a treatment work for some people and not work for other people. If we think about depression, for instance, I mean, one of the treatments is antidepressants. They are typically prescribed on a trial and error basis. Psychiatrists would try a type of medication, see their patient a few months after, has it worked, has it not worked? And then iterate like this based on trial and error, essentially. And so one of the things where, one of the areas where AI can help is to try and better predict based on people's clinical assessments, people's history, possibly from genetics data, imaging data, even brain imaging, what treatment will work best for for each patient. And this is important because, you know, all these treatments have side effects. And so you don't want to, you know, I mean, it's important to get as fast as possible to the right treatment. So that's one example.
Ayush Prakash
When we're zooming into artificial intelligence, which I think will be the bulk of the episode, when you're saying that we're using AI for mental health, what does that mean? Because immediately people are going to say, oh, your patients are talking to chatbots. You know, maybe that's true, but is there more to that? Are there different things that AI is doing in the background, in the foreground? What is the scope of that?
Tristan Glatard
No, that's a great question. And I'm happy we're having this discussion because I think it's important to disentangle all these things. And so I'm going to speak more about the opportunities, but I'd like us to talk about the risks as well, because as I said, I'm very conscious about that. And I think it's very important to have a conversation which is measured between opportunities and risks. So I think today in the general public, we think about AI as these language models, chatbot type large language models, the ChatGPTs, the Geminis, and all these tools that essentially can talk in written or spoken form to people. But there's also more to that, more in AI than these large language models. So maybe I'll start with that second part. So what I was describing before, if you want to predict which treatment will work, if you want to predict, for instance, if people are going to have crisis, that's also a big application of AI. Can we predict when people are going to be in crisis and therefore try to anticipate that before the crisis hits? So this is a technology which technically is very similar to large language models. I mean, that's machine learning. That's very similar. But in practice, it looks very differently because we use other types of data than just text. So for instance, if you want to predict, we have projects with Fitbits or all these wearable devices, for instance, for crisis prediction. You want to create AI tools which can work not from text, but from these different data types. In order to make predictions which are relevant for patients. So there's a whole range of these tools that actually people interact with through some products in their daily lives, but they are not what we typically refer to as AI in our daily lives. So for instance, think about Let's say an app that's on your phone and that can tell you, oh, you know what, maybe you should have a conversation with a professional because we've noticed that you're not doing that well these days. So these things are not based on ChatGPT type of AI, but on the integration of other types of data. And so So yeah, that's one thing. And then I know maybe we can also talk about chatbots for mental health because they are also an important application.
Ayush Prakash
Yeah, I would love to go down that tangent just for a second. Is there a strong case for an AI companion? So let me just lay the land. So one thing that I've observed from a lot of, let's say, tech executives is they want people to have an AI companion with them at all times. And what they say that this would be able to do for people is remember their bills, remember their conversations, be able to serve them up information as needed and track their lives. And so you go focus on being human. We operate in the background and allow you to be more human.
Tristan Glatard
Yeah.
Ayush Prakash
Okay, that's a great vision, but I think it's a little more complicated than that. In the domain of mental health, is there any strong, obvious case that people are just very blind or very unable to see just because there's so much doom and gloom, so much FOMO, so much hype? And if that is the case, where does that case break down? So I guess it's 2 questions. Number one, is there a case and where does that case end?
Tristan Glatard
Yeah, so there is a case. I think we should delineate between mental health care, you know, which is providing medical support for people who suffer, well-being, which is more, you know, I don't have any particular diagnosis, but I'm not feeling that great today. I'd like to talk to someone. And then general-purpose AI, which is more like I'm opening ChatGPT and asking questions related to emotional support or mental health. So there is a clear case for the first type, so mental health care, chatbots, just because in this day and age in Canada, it's actually difficult to access psychotherapists for multiple reasons. It can be cost, it can be because not everyone lives in a city where they're easily accessible, it can be time, people work. So in practice, many people need psychological supports, but don't get the support that they need. And I think as a society, that means that people may be depressed, may have all types of conditions, and are not getting the support that they would need. So this is big because it's a big problem in society. So I think there is a clear case for AI chatbots to provide psychological support, but with a big but, which is it has to be evidence-based. It can't be just ChatGPT telling you how to deal with your depression or how to deal with your alcohol addiction. This is just not the way it should be. So one of the things that my colleagues do at CAMH in Toronto is to develop these chatbots, but chatbots that follow specific methodologies because they are methodologies that psychologists and psychotherapists follow. So making sure that the chatbots follow these methodologies so that they provide care as similar as possible to a human therapist. And of course, the intent is never and will never be to replace human therapists, but it's more to provide a middle ground between, I'm all alone with no support, and I need to see someone and having this human connection. And I think there's a very clear— I'm actually very excited by that. I think there is a very clear case here. Think about, for instance, people who struggle with addictions. They may be followed at the hospital. They may have all kinds of supports. But when they start having cravings at 2:00 AM, nobody's going to be there to pick up the phone and— and provide evidence-based support on how to help them. And so I think if we can build— and there are big risks around that— but if we can build these chatbots, and there are already examples coming out of CAMH, that can have a huge impact.
Ayush Prakash
If I were to double-click on that for a second, I really want to discuss what the difference is. Maybe it's just anecdotal from what you've observed from a human talking to a human versus a human talking to a chatbot and the different types of information that they would relay.
Tristan Glatard
Yeah.
Ayush Prakash
Because it feels like the chatbot, again, from what I've witnessed, is able to get more information out in some way. People feel more comfortable with the chatbot immediately. Is that all true from your observations?
Tristan Glatard
So I don't know if we have good data on that. I think that's a good intuition, you know, What I'm hearing too is that some people, for instance, people with autism may be more comfortable, or people who have communication difficulties may be more comfortable to speak or to interact with a chatbot than with a human. I don't know if there is good data about that. I also can see how, just by the ways that we interact, like it's when you come to your therapist, it's going to be, I don't know, on a Tuesday afternoon at 3:00 PM. And maybe you would talk to your chatbot at 9 PM after work or after you had anyway all sorts of— so I think the modalities are different. I can see how indeed chatbots could get more information. But I also think that where AI is yet is not— I don't think comparing AI with human therapists is the right approach for now. I think the right approach is how can we expand the support that human therapists provide with AI, just because there's not enough therapists in the country to provide support to everyone.
Ayush Prakash
Roger that. And again, I don't want to stray too far from the river. We were talking about— we were going to talk about where the case breaks down.
Tristan Glatard
So this first, we talked about mental health care support, evidence-based, under the supervision of therapists. I think there is a whole gray zone around what we can call well-being. these apps that help you manage your stress, cope with your anxiety and all these things, coaching things. And so I think here the evaluation and validation is in general lacking. So these apps, they don't claim to provide healthcare. So this is why they can also exist. But I think we need more evaluation. We need to be careful about that, cautious about that, and think about evaluating them. And then when it comes to general-purpose chatbots, I think support like mental health support from ChatGPT or Anthropic tools should remain very light touch. I think a good analogy are maybe medicine, drugs. So you have drugs which are over the counter, the more potent ones need a prescription. And at the same time, on the other end of the spectrum, you have things like maybe something that you put, I don't know, like a— anyway, like a whatever light thing that you can just use for yourself. And I think this analogy also holds for AI. So we don't want, we can't have, you know, non-tested, non-evaluated AI tools just provide mental health support. That's just not responsible.
Ayush Prakash
I know we were going to talk about the opportunities and the risks and try to balance that. But before we get into specifics, I want to ask more of a meta question. How do you balance opportunities and risks? There's so much happening in the world of AI. There's so many different things that are being said, so many different technologies that are coming out, so many different political and geopolitical events. And it feels like the zeitgeist sways one way and the other, you know, many times a day where people are all for AI and all excited and then all terrified and all doom and gloom. How do you stay grounded in that? How do you continue to see past the noise into the signal and again find that balance and not believe one side of the equation versus the other side? What does that look like for you?
Tristan Glatard
I think working in a hospital-based environment changes a lot of things because we are driven by specific use cases. We have a responsibility. If AI has the potential to help patients, we have a responsibility to unlock that and to enable that, right? Because it would not be responsible to ignore that there is a technology that can that can better support, provide better support for our patients. So, you know, on a day-to-day, we have clear use cases, projects. I'm a researcher, so most of these projects are research projects, but I'm also involved in more operational aspects of the hospital. And so this is, you know, our mind is sort of shaped by, okay, how does this, you know, apply to, you know, the treatment response models that we are trying to build or chatbots that, you know, may provide psychotherapy support. So I think this is one thing. And the other thing is, likewise, being in a hospital environment, there's many aspects like CAMH is very adamant and focused on health equity, obviously on patient safety, fairness. And so when we think about AI, all these things are part of the conversation. We have team members who focus on AI fairness, for instance, and health equity. So that creates safeguards within the institution that there's no way that we can just pick up the latest and greatest and apply it without evaluation or thinking about it. The other thing, maybe another aspect to your question is I think it's very important to remain informed, not just about technology, but about more generally how people think about technology and to be able to reflect on what's happening on a day-to-day. So looking at the long term and history rather than just the news that's happening today.
Ayush Prakash
If you could write an email to any of these tech executives and wave a magic wand and create a feature for these chatbots and take away a feature of these chatbots, it could be an architectural feature, it could be whatever it is, what would you do just from your perspective, from somebody that works in the hospitals and works at CAMH?
Tristan Glatard
Yeah. So if we focus on these tech giants, AI products, I think one thing I would add is a clear detection of of crisis cases. You know, like, we can name suicidality, we can name, you know, people starting to develop delusions, we can name risks of violence. You know, I'm referring to this recent example early September of this person who, you know, presumably, or, you know, used ChatGPT to prepare this terrible shooting in BC. I mean, these types of things are crisis situations that involve mental health and that, you know, I think these tools should— I'm not saying it's easy, but should be able to detect and then, you know, react to. So I think that would be the number one thing to add. If we're talking about, you know, Things to remove.
Ayush Prakash
Yeah, there's—
Tristan Glatard
it's hard to pick one, but I think the whole data management part is something which, you know, is something which is a bit concerning. You know, these tools work well because they manage to aggregate and summarize gigantic amounts of data. And I think we're all aware that It has shortcomings, limitations. When we talk about medical data or mental health data in general, this is a big concern. So if I had a magic wand to make sure that these tools can just respect people's data privacy, I think I would use it.
Ayush Prakash
I've been struggling because the way I get my news on artificial intelligence is just by talking to folks like yourself or reading articles. I can't read books anymore because any book that's post-2022 is likely touched by AI, and I'm not going to read anything like that. In terms of the mainstream news though, these topics revolve around 5 to 10 of the same things constantly. It's data centers, it's energy use, it's, you know, political things and blah, blah, blah. But these are not likely the talk— your daily conversations. And I feel like just if I'm reading this conversation correctly, that you have a lot going on, you know, in your mind and in your conversation that just— it's not that it's not important enough to make the mainstream news, but it just— it's so dense that it wouldn't be able to penetrate the mainstream news. There's no ability to crunch everything down into 30-second soundbites. What is that? What are you talking about? What would you like to see become a mainstream conversation, or at least become a conversation that's just the talk of your colleagues, but it's not able to penetrate broader conversations?
Tristan Glatard
That's a great question. I think your question is important because it's important that discussions about technology in general, not just AI, but including AI, are not just confined to technical people or the developers. I think these are societal conversations. And so that's how I see penetration to the mainstream news. So I think having conversations around how does it look like to have potential support for mental health through AI at a broader level, in the society, you know, there's big risks. There's people have fears. It's not sometimes, it's not always objective, and we should respect that. But also there are opportunities. So, you know, we both as researchers and in my case, you know, in a clinical environment, we can develop, think about tools, think about the opportunities. But at the end of the day, I think it's important that people get informed about about these things and are able to cut the line between the horror stories that we hear in the news and the big risks, but also the opportunities. So for instance, I would love to have deeper conversations about, or more general conversations about, imagine you have an AI that can tell your clinician, your psychiatrist, what's the right treatment for you? How would you feel about that? Yes, it's an AI. It's using your data in a way that's contained in the hospital. The data stays. It's not shared with anyone. But it's still the same kind of technology. So you may not like it. You may not want to have an artificial tool involved in your care pathway. But at the same time, it can improve your care. So between both, which one do you choose? Which one? How do you look at that? Just as a quick segue, we actually have, you know, there's a few people where I work at CAMH who work on developing this dialogue with patients to understand their perspectives. And so, yeah, I think this kind of— so, you know, how about involving AI in healthcare? And, you know, what would you think about what I mentioned before about these chatbots? You know, like if you If you have evidence-based chatbots working under the supervision of a hospital or a clinician that would help you with your struggles, how does it look like and what do people think about that?
Ayush Prakash
I want to again double-click on that because a thought popped into my mind and I can't square it properly. You are— let's just create a hypothetical. So it's you, it's your AI, it's a patient, and it's their AI. What does that interweb of relationships look like? Because you have to kind of talk to your AI, you have to talk to the patient, you have to read the patient's AI. How do you manage all that different information? I guess the question I'm asking is, where does, what does the source of truth look like? Is it still you?
Tristan Glatard
Yeah.
Ayush Prakash
Are you still processing everything or do you lean towards one of these different— I won't say entities, but one of these different sources of truth?
Tristan Glatard
To answer your question, again, I'm seeing that from a technical lens, but I think the key to answering your question is how do we evaluate the AI? So you have a chatbot that talks to a patient. How can we rate the conversation on a scale from 0 to 10? existing evaluations that— because psychotherapists also have evaluations, they have ways to evaluate their trainees, this is all quantified. And so I think the key lies in having good metrics. So you may have thousands of conversations happening with these evidence-based mental health chatbot tools. I think the question is, can you put one number? Was it a 9 out of 10? Was it a 3 out of 10? And then have human clinicians look at the bad cases, making sure that typically what you want to happen is that when the conversation derails, what's the worst-case scenario? Someone talks with a chatbot, for instance, because they want to reduce their alcohol consumption, something like that. And the worst-case scenario is that let's say you have someone who is getting suicidal or is talking about hurting someone else as a result of the interaction with the chatbot. And so these tools need to be able to very rapidly flag that and interrupt the conversation without even an interaction from a clinician. And then after that, the clinician and the technical team can have a look at that and understand what happened, how can we improve the tool, et cetera. But I think that the first principle is to not harm people and to just stop everything when the conversation derails. The second situation is the conversation sort of happened, but it didn't have any effect. So one thing that's measured in these conversations is typically willingness to change. So a sort of average scenario is Someone comes to the chatbot, talks about, let's say, reducing their smoking or quitting smoking. And then at the end of the conversation, the conversation hasn't progressed. People still don't have more motivation or their motivation to quit smoking hasn't improved. So in this case, it's also important to rate that and that this be reported to clinicians and the technical team to understand what has happened. So this is how I see the relationship. You have the technology scales. It has to be evaluated. And then essentially clinicians and the technical team supervise these AI support tools. And some of them may be fired in case they don't work properly. Some of them may require additional training similar to how humans are trained and managed.
Ayush Prakash
When was the last time you were surprised by a piece of research?
Tristan Glatard
I'm not going to say every day because I think I would be I would be lying, but, you know, definitely every month, maybe every week.
Ayush Prakash
What does that surprise look like? What is it? Is it shifting your internal perspective on the world or is it just, oh, I didn't know that. I didn't think that was possible. And how do you manage that?
Tristan Glatard
Yeah, so that's a great question. I mean, for me, this is a big part of the drive. You know, for my work. So I know that was the meaning of your question, but I'll tell you, you know, something from my own research, from my lab. Recently we have noticed, maybe discovered, I don't know, but we'll see, that very simple models, very simple AI models, models that don't require training, So models that don't require thousands of data points, weeks of GPU compute. Some of these models that we tested by chance, they're actually working really well for image analysis. And that was very surprising initially. We thought it was a mistake. But then we realized, oh my God, this actually works. So we have a paper in preparation for that. And it's really nice because you have these moments that you know something which is very surprising because in imaging, the state of the art, what people do, they integrate tens of thousands of data points. Training these models requires the most advanced GPUs, and now we show that with a very simple model, things work well for what we were testing. So, you know, you have this moment where you know something. So in my research group, my students, postdocs, we know collectively something that no one else knows. And I don't know, there is a very unique feeling about that. Yeah, and that's a big part of the, for me, of the motivation for research. It's like knowing things, learning things, and then of course then the question is, how do you convey that to other people? How do you transmit that? Are people interested or do they not care at all, which is fine too.
Ayush Prakash
Speak more about that. I want to know in your day-to-day or week-to-week or month-to-month, there are these moments that you have that make it all worth it. And you can't really talk about them. If you were to bring them up, it would just kind of be— not to say that they're weird, but it just— we have these moments in our lives that we would love to broadcast, not for the ego, but for the soul, for the heart. What do those mean to you? What are they? How do they represent?
Tristan Glatard
I don't know if they are— sorry, I don't know if they are discrete moments. Like, you know, so we have, we have our lab meeting, our weekly lab meeting every Friday. So, for me, that's a very important moment because this is where we— and I'm saying me, some of my colleagues from CAMH, my trainees— we collectively discuss in general new results. This is very interesting. You know, this is very exciting. So, you know, it looks like someone did an experiment this week, shows new data, and then in general, it's like, oh, there's many limitations. But sometimes it's, wow, okay, that's really cool, that's really interesting. But the truth is, the way I like to work is we interact very much on a daily basis. We use instant messaging on a commercial platform that I won't name, just not to advertise for them. Yeah, and so we have these constant interactions, this constant stream of, you know, look what I found, look what I read, how does it relate to what we're trying to do? And so this drive, this excitement is sort of blended in the week. And so, yeah, there's also, you know, this thing that, I mean, that's a whole conversation about research and the excitement for new knowledge. But I find that this excitement is also like you get excited for a few days when you see something new, and then you need to go beyond that. There's always more questions and more. So it's sort of a constant drive and I don't know how to call that, but like a constant ongoing path to do more things, to understand more. And it's like, you know, like the excitement also wanes after a few days if we don't have more results, more, you know, new ideas, et cetera.
Ayush Prakash
So I love it. And I just want to ask one more question before we let you go. What gives you hope in the world?
Tristan Glatard
Yeah. I think what gives me hope is to see people who are interested about moving the needle, about having an impact. And I mean, you know, in the environment that I work with now, that means, you know, advancing technologies to improve healthcare in general, mental health in particular. And it can take many different forms. You speak to, for instance, I don't know, a company owner and they are developing this product, but yes, of course, there's always selfish motivations and people want to get rich, they want to get famous, they want all these things. But I think what gives me hope is also to see when people are genuinely interested in making a good product, writing a good research paper, acquiring good data, not because this will make me keep my job or because people will give me money for that, but because I like to do good work. And this is, you know, I want, you know, some people would say, I really want good data to be out there. And so Yeah, good products, good science. I think this is, you know, people generally interested in, and I see that a lot around me, you know, people who are interested to do good work, to have an impact and to help other people.
Ayush Prakash
I love it. Tristan, this was a great episode. I really do appreciate your ability to speak at length on such dense topics and to make it accessible for people. Where can we find you online to keep up with your research and your work?
Tristan Glatard
Yeah, I have a webpage on the, on the CAMH website. I also have a Google Scholar profile with all my publications. My research lab also has a page, it's called epics-lab.ca. Yeah, so this is where we post our things.
Ayush Prakash
We'll put that all in the show notes. Tristan, thank you so much for today.
Tristan Glatard
Thank you so much, Ayush. It was a pleasure.
No passages match that search.