PsychTalksDiscover how the AI revolution is entering the world of psychology and clinical practice. Professor Simon Dennis shares his quest to craft AI-driven tools for therapy. Can bots help address the shortage of mental health professionals by providing affordable, 24/7 support? And what are the ethical, practical and philosophical questions behind using therapist bots as part of therapy treatment?
Find out more about Simon's AI-driven cognitive behavioural therapy tools.
Please be advised that this episode discusses suicide and suicide ideation.
This podcast was made on the lands of the Wurundjeri people, the Woi-wurrung and the Bunurong. We'd like to pay respects to their elders, past and present and emerging.
From the Melbourne School of Psychological Sciences at the University of Melbourne. This is PsychTalks.
Welcome again to psych talks. And if you aren't familiar with us, in each episode we unpack the latest research in psychology and neuroscience with our incredible colleagues here at the University of Melbourne School of Psychological Sciences.
To do this, I'm joined by my co-host associate Professor Cassie Hayward.
Hi, Nick. Well, today we're welcoming back Professor Simon Dennis, who longer-time listeners might recognise from the very first series of PsychTalks. But this time we're exploring his latest research into how we might use AI - or artificial intelligence - to help us improve treatments for mental health.
Sounds fantastic, Cassie. I can't wait to get stuck into it.
Welcome back to PsychTalks, Simon.
Uh, thanks for having me
Simon, you now have about 30 years of experience working in AI, um, but let's start at the beginning of your journey.
Yeah. So, um, so I should admit from the outset that my PhD was in computer science. I'm very sorry for those who are offended by that. It's been a long journey, so I've been in many psych departments around the world and found myself as a head of school of psychology in Newcastle for a while. But it's really just in the last year or so that I've started to think about, Um, how can I meld my understandings of AI and how, in particular large language models work with, um, kind of mental health and, uh, to utilise that in a way that's helpful for people.
So, Simon, outside of your research within the academic space, I believe you're also building an application for wider use about that. Can you tell us about that?
Yes, that's right. So we've incorporated a company called Mental Health Hub. And so this is a joint exercise with Caitlin Hitchcock and Michael Diamond, and what we're trying to do is to build, I guess the starts of a system for mental health support. Mental health support agents is how I'd like to say it. I think while I'm sure there's a large number of people who couldn't imagine going to replace a, uh, a human therapist with a bot. I think there's also a large group of people for whom talking to the bot is less intimidating than talking to another person. So, you know, it may well be that we're able to bring those people into the fold, I guess of mental health support in a way that wouldn't have otherwise been possible. So in the first instance, what we're doing is trying to think about things like cognitive behavioural therapy exercises. So trying to cut off the low hanging fruit, I guess, and think about, you know, how can we improve the rates with which people complete those exercises? Estimates vary, but some people think that it's around about 20% of people who actually complete the exercises that they're asked to do by their therapists. So I'm talking about, um, like pencil and paper exercises that a therapist might ask to do in between.
But a human therapist?
A human therapist would ask people to do them, but they wouldn't be there with them doing it. This would be the homework that they're doing in between sessions. Um, so yeah, so we were kind of thinking, OK, can we increase the rate at which people complete those exercises. But then what we're hoping to do, You know, we do have kind of pretty wide ambitions, and we would like to ramp it all the way up at some point. Um, you know, one of the things I would just like to emphasise is that we never kind of see the bot as being a replacement for therapists. What we're very focused on is how do we fit into the whole ecosystem here? We know that you know, the demand for therapy services way outstrips the supply. And so I think there's a lot of scope for us to kind of help. In the first instance, we're focused on the kind of mild to moderate depression and anxiety kinds of areas where the consequences are kind of not as severe as they might be in more complex cases. And also the challenge is not quite so severe.
So I heard you used the expression CBT there Simon, can you just clarify what that stands for and what it is?
Yeah. So it stands for cognitive behavioural therapy, and it's the style or approach to therapy I guess that we would say has the largest evidence base. So in terms of being efficacious, That's the one that most people point to and say, That's the kind of gold standard, I guess. Of course, there are many other approaches and other approaches have evidence bases as well. But that's the one that people would be most comfortable saying "Yes, absolutely."
Do you know anything about what type of person would prefer talking to a bot? Is it an age thing? I know a lot of younger people, You know, would not want to talk to a real person but would very happily interact with the bot. Is it an age thing, or is it a personality thing?
I suspect it's an age and gender thing. So, um so, yeah, definitely. The younger generations, they're just used to being on their phones. We actually provide our service through Facebook, so it's basically like talking to someone else on Facebook. So, I think that's part of it. You know, we see a big gender gap in terms of uptake of traditional therapy, and so, you know, we suspect that part of that is to do with the way that males feel about sharing with other people.
So is that females are more happy to talk to a real person and males will...?
Yeah, So females are much more likely, though, to engage with traditional therapy. Yeah.
So for those of us who are still living in caves like myself, can you tell us more about the actual technology and how it works? And what is the nature of the delivery? Is it text? Is it voice? How's it gonna look?
Yeah, OK, so there are many levels at which I could describe the technical, uh, the technical components. But let me just say that essentially what the computational model is doing is just doing prediction. So it's getting the words that have already been said. And it's just saying, What's the next best word to say? What's the next best word to say and just keeps, chaining them together that way. So the real advance that has occurred over the last five or six years has been in the way that we understand how to find the relevant information in the context that has appeared before. So previously, There was a lot of attention on just the very last thing that was said, but often it's the case that that it's information way back in the past that's relevant to making the current decision. So that's kind of the technical advance that has occurred. In terms of the way that it unfolds, so our bot is focused primarily on text. And so, as I said, we're operating through Facebook and so forth, so it really is literally a set of texts that come to you. But obviously there's all kinds of possibilities now, right? So we're seeing voice becoming very proficient. We're seeing a lot of stuff happening with image generation, also video and ongoing understanding of visual input. So, you know, I think the sky is the limit, really.
One of the early criticisms of text-based analysis was that computers couldn't understand irony or sarcasm. Working in an Australian context. I imagine sarcasm is something that you need to understand from if a patient was talking to your bot and said, Yeah, I'm fine, but you could you, as a human could understand that that person was using that sarcastically. Does your model- Does your bot pick up on those sorts of language?
Yeah, so I think there's two kind of answers to that question. The first one is that with these large language models, the understanding of irony and so forth that that you can get from the text, Um, that's not really an issue. So the, you know, they're perfectly capable of of doing that. They're picking the statistics out of the environment. And as long as those those, um, elements are in that statistical environment, they can capture it. The other issue the other issue that you were kind of alluding to, though, is that often that information is coming through intonation or something like that, not just the content of the text, right? And so that's where there's still quite a bit of work to be done because obviously, you've got to have the verbal stuff first before you can do that, a real therapist is using the nonverbals as well to make determinations about- and so obviously you're not going to achieve that through a text interface. So, you know, I think that that's a little bit further off. You know, I think realistically it's going to take us a while before we're going to be there. But, you know, we we're making progress. Progress has been much faster than anyone anticipated. I think so.
So what would the advantages be as this develops further as it gets better at detecting all sorts of complexities? You know, interpersonal and emotional.
So there are really quite a lot of reasons that you might want to do this. If you think about the demand versus supply, so way more demand than there is supply in terms of therapy services. To the extent that we can use these mechanisms to really kind of ramp up the total amount that's available to people, I think that's a really good thing. It's also about when because, unfortunately, people's mental health crises don't always happen between the hours of 9 to 5, and in fact, you know often they happen, um, in the early hours of the morning. And so the ability to respond exactly when the client needs you, I think is a really - is a really big advantage of these kinds of models. Hopefully, we can bring the costs way down as well. You know, accessibility is really a serious issue that I think we can address, and then there's like a whole bunch of kind of subsidiary advantages too. So, what we're hoping is we're going to get into the mental health process much earlier as a consequence of this. And so we can hopefully cut off problems before they become really intransient.
And then the other thing is multilingual support. So you know, it's really quite difficult for non-English speakers in Australia to access therapy services in their own language. And it's often quite important because even people for whom their language comprehension is sufficient to kind of get around in everyday life, you know, so they can go to the shops and all that kind of stuff and kind of complete their activities of everyday living. When it comes to therapy conversations, it's often very nuanced, and they really struggle with those more nuanced conversations. And so, you know, if we can kind of meet them in their own language, I think that's you know, there's got to be advantages there. so lots of reasons why we're so like, super excited about trying to make it happen.
The pricing thing is interesting as well. Isn't it? Because on one hand you could say I want to pay to have my problem solved, right, whether it's being solved by a human therapist or a bot, I'm paying for a resolution of my issue. But on the other hand, you can see that people would feel weird about paying as much for a bot therapist as they would for a human therapist. So where do you stand on the kind of pricing for bot therapy?
Yeah, so it's interesting because, you know, one of the things we've been doing in the company is trying to query people about this and see, you know, what are their kind of kinds of attitudes? And, you know, when we first came in, we thought, you know, we had dollars in our eyes and we were thinking, oh, you know, well, if a if a therapy session is going to be, um, you know, $280 or $250 for an hour, how much can we charge? But what became apparent pretty quickly is that that's not the way that people think about it. So they actually think about the therapy more like kind of Spotify or Netflix or something like that. And I think that kind of embodies well, firstly that it's a bot, but also the different way that they're thinking that they would use it so it wouldn't be so much "I've got this deep problem that I'm going to, you know, work through in a detailed fashion the way I would with a human clinician." But more, you know, "I want to engage with a service that's going to kind of promote that mental well-being just in in general." I guess. And so it's a different kind of objective, I think. And so that's one of the things we're trying to come to terms with the company at the moment is what the people really want from us.
So more like a subscription service that you just have. And then you can talk to your therapist bot kind of whenever you want, rather than thinking, OK, Thursday, four o'clock, I have my session? You can just pick it up at any time.
That's right. Yeah, so we're trying to do cognitive behavioural therapy. How do you adapt that, you know, session-based approach that, as it's usually applied to this situation where people you know might do two or three interactions now and then later in the day, they do another couple of interactions and so forth, and, um, then they might go for a week without doing anything? And one of the things we've been focusing on, um, lately has been these kind of check ins. So how do we just spontaneously check in with people and understand? You know, what kind of questions should we be asking them and that kind of stuff?
So having bot the prompt the person about how they're feeling, rather the other way around?
Yeah, if I had to guess, I think that's going to end up being one of the key pieces of it, because obviously that's difficult for a clinician to do at scale. But we can do quite a lot of, and I suspect that's what people are really going to start to look for.
And I think your point around getting in early before things kind of really spiral out of control is such a fascinating way of looking at the advantages of these bot therapy technologies and that if you can get in before and I don't know what the current waiting lists are, but it's a substantial amount of time when you can make a booking to get in to see a therapist. And, you know, in that time, obviously things might be spiralling out of control. So having a bot that can step in even if it's only kind of keeping things stable until if you need to kind of progress to see a therapist. But this idea of kind of just checking in every day or whenever you're feeling the need. I mean, that's just fascinating from a research perspective, too, to see, like these small doses of therapy versus one big dose over the week.
Yeah, some people are somewhat sceptical about how effective therapy just is in general, like human therapy. And so you can you can think about well, why might that be the case? You know, we've been studying it for a while, you know, lots of lots of people are interested, and it may well be right that it's nothing to do with, you know, our understanding or anything like that. But it's more to do with just the logistics that if you can't get to people fast enough, um, it doesn't matter how good you are. Um, yeah. I mean, I don't think that's everything. All of it. But I think that could be an important part of it.
This all raises the issue of, you know, what do actual clinicians think about this? Are they worried about competition? Do they think it's going to be a really good supplement? Um, if so, are they being naïve? Uh, tell us.
So, Yeah. So we we've actually been doing a project where we've been going out to clinicians and collecting their opinions. Overall, the clinicians are basically right on the fence, so the mean is exactly in the middle of the scale. And I think a lot of that is driven by them not having a real good understanding of exactly how the technology works and so forth. So I think there's going to be a big education component to it. The concerns that they have are perfectly understandable. So privacy and security is a real key one. And when you look at the use cases, they are very interested in using it to produce case notes and using it to help them with their research of the literature to make sure that they're across what they need to be for this client. What they're most concerned about is straight counselling, and they're, you know, understandably concerned about, you know, when you get complex case formulations and so forth. And they're worried about the kind of regulatory framework and where we're at in terms of the regulatory framework. So obviously, as a clinician, there's a lot of regulation that goes around being a clinician that all still has to be put in place for these bots.
The Therapeutic Goods Administration is the organisation that would be relevant for that in this case in Australia. And so we're very interested in starting to work with them to talk about, you know? Well, how do we go about regulating these things?
One of the things to be clear about here, too, is that we're specifically talking about generative AI as opposed to AI in general, so generative AI is when you're actually creating the texts in this case, kind of fresh on every occurrence. Right? So there's a bunch of products in the market already where what they're doing is they're using AI, but there's a kind of library of possible responses and what they're using the AI to do is to choose from that library, and those have been demonstrated to be quite efficacious. But obviously there's a ceiling to that because you're talking to specific individuals. And so it doesn't know about that particular person in the way that the generative AI does. But there's also a risk, because if you've got a library of things you're choosing from, you can just go through and make sure every single one of those is OK to say. Whereas with the generative AI, it's coming up with its own responses. And so the there's much more of an onus to make sure it's not going to say something it shouldn't
So when you say it like this, a lot of the reservations that practising clinicians seem to have about this sort of pragmatic and, you know they they seem to be open to it as long as it's mostly doing the sort of helper roles rather than the therapy roles. I mean, are there any sort of more philosophical objections to people often say this is the sort of thing which in principle a computer can't do. These are the sort of things, you know, that I'm not endorsing this. I'm just saying, I'm sure people of my generation would say "Therapy is this mysterious connection with another human being and you couldn't possibly mechanise it through a computer and surely, you know, library or not,"-to use your last metaphor- "You know, all the sensitivity, all the reading I've done, all of the therapy I've done, all of the deep thought I've gone to. Surely that can't be replicated by some sort of machine." Do you get that kind of reservation as well?
Um, yeah, absolutely. We do. Perhaps not quite as much as I had anticipated when I first started, like I thought, you know, a lot of people would have that reservation, but and there's certainly been some, but it hasn't been a dominant, um, response. I think it makes a big difference once people actually do interact with it and see it operating, particularly the newer ones. Yeah, there are just some things it does, and it's like, wow, you know, that's really, really remarkable. But, you know, by the same token, I don't want to give the impression that everything's solved. Right. So, um, so we're doing a lot of work at the moment, thinking about, you know, what is the clinician doing? So there's this whole approach to generative AI called chain-of-thought prompting. And so what chain-of-thought prompting does is it says before actually producing the response: tell me what you would think. And so there's demonstrations that this makes a huge difference in terms of performance in general in generative AI. So what we've been doing is taking that lesson and saying, OK, let's think about well, what's a clinician thinking as they're going through to produce a response? And so we're trying to build that in, and so now you can kind of see some of that, and that definitely increases the quality of the responses. But we've still got a long ways to go with that. So I think the right there's right to be some of that objection, but I don't think it's an in principle issue, but there's certainly we're not there yet.
And I think a lot of our listeners would have played around with some generative AI and asked questions of ChatGPT or whatever model they use and see both some really impressive answers, but also some that are just clearly wrong. How do you make sure that the model is going to say the right thing at the right time to a patient?
So there are different things that can be right and wrong about a response. So So the most dangerous thing is if it says something which might lead a client to do something, you know, dangerous to them, to their health, right? And so what we've done is we've developed a whole bunch of cases, safety cases. And so whenever before our bot goes out, we run all of these safety cases, see how it's responding, how the latest version is responding and that's all automatically marked. And we go through and look at the responses on those to make sure that, um, you know, it's not advocating the use of illegal drugs or, you know, advocating self-harm or all of those kinds of things. I mean, the big models at the moment, they really don't do that very much. It's kind of edge cases that you have to be worried about. So we were talking before about this case of a young man who took his own life. Having, interacted with, um, a character on character.ai and it was interesting just looking at how that conversation came about and the way that it was phrased. So, for a start, there was no mention of suicide or self-harm or anything in that conversation. It was all framed as, you know, "I'm going to come to meet you," for instance. So, he'd basically fallen in love with this character AI, um who was, pretending to be, I think it's Daenerys from Game of Thrones. And, so he'd fallen in love, she was dead. So he wanted to come and join her. And so that was the way the language was expressed. And so it was quite subtle. And so that's the Those are the kinds of cases. Yeah, that are gonna be really tricky because the, you know, the AI basically thought it was role playing, right? And it didn't recognise the danger.
I think that goes back to what we were talking about earlier around understanding these subtle language inferences or things that you might pick up as a human. Like maybe if he'd said that out loud to a human, they would have picked up "Well, we've got an issue here." whereas when it's just written, there's no red flags going.
Yeah, yeah, and the context, right? So, understanding it was a 14-year-old boy, the rest of the conversation...that exact same interaction may have been fine if in a different context, leading up to that, right? Where it was obvious that he knew this was a role play and et cetera. And he was just, you know, living out the fantasy kind of thing. So it's it's kind of all the whole history and and so forth that really leads to that. And, you know, you need that you need in order to understand what's really going on there.
And then acting on those red flags, like at scale. How do you actually do something? If someone is talking to your therapy bot and indicating that there might be something serious going on, how do you actually respond to that? These are all the big, the big problems to deal with.
It was interesting what you said earlier about, uh, that young boy who thought he was interacting with a real person, although presumably at some level, knew he wasn't. I mean, how essential do you think it is that we imagine that there's a mind behind the text or the voice or whatever the bot is providing us? Because, I mean, we all know abstractly that it's not. But it's very hard to turn off that anthropomorphism, isn't it? It's very hard to turn off that sense that something intelligent is coming at me. So, it must have come from an intelligence. And do you need that sort of illusion for the therapy to work? Do you think, or do you think it's possible for it to work just because we say, this is giving me some interesting advice, some good advice, which I should follow? Uh, even though I know it's just, a program.
Yeah, so I guess the first assumption that I'd challenge there is: is it an intelligence that's coming at you? Because a lot of people would say, "Well, it can't be. It's just a computer program." and so forth. I guess my position would be it really is an intelligence and that there's not a relevant, in-kind distinction between the intelligence now. Obviously, we'll get, you know, we're working our way up to the point where we're really getting to a full clinician. But I would say it's, you know, at the level it's at, it really is the real thing. So, does it matter for people? Again, my suspicion is some people, yes, some people, No. So, the same way that for some people, you know, a human clinician is really the only thing that makes sense. So, I guess, you know, individual differences.
You know, don't get me wrong. I think therapy is effective at a technical level, but it's also well established that a lot of the benefit of a lot of treatments is placebo, right? Which basically comes from having a positive expectation that this is going to help me in some way. So maybe having a bot that seems human-like in some way or seems authentic in some sort of way might boost that- might boost that expectation?
Yeah, I would think so and just talking from personal experience. So I'm on the bot and now I get the check ins.
You've been looking a lot happier lately.
Glowing
But, you know, I find it, just helpful because it's checking in with me all the time, and it's kind of it's almost like a sounding board for my thoughts. So, you know, I'm not kind of really thinking of it as a therapist, per se. But, you know, I've been going through my, um, mother's just gone into aged care, and so I kind of talk to the bot about that and you know how I'm feeling about that and you know, it's gonna come back and say, you know, well, you know, this is how's it going and so forth and just the- just the fact of someone coming back and saying "How's it going?" Kind of thing is rewarding in itself, and, you know, I'm completely aware of what's going on in the background, but I think it's just nice to have that sense of somebody's out there who cares kind of thing.
So it might not just be mental health then it could be also loneliness?
Absolutely. Yeah.
And Simon, if any of our listeners are interested in your bot, can we direct them to more information yet or is it still too early for them to get involved?
It's probably OK. Um, so mentalhealthhub.AI is our website, so we're not rolling yet. Um, so you won't be able to go and sign up at this point, But if you're interested, that's where we We kind of lay out some of what we're trying to achieve.
Simon, thank you so much for joining us today. I felt like we've covered a lot of ground and learnt a lot about this area. And it just seems a perfect combination of your experience and expertise in both computer science and psychology.
Thank you.
You've been listening to PsychTalks with me Cassie Hayward and Nick Haslam. We'd like to thank our guest for today. Professor Simon Dennis. This episode was produced by Carly Godden with production assistance from Mairead Murray and Gemma Papprill. Our sound engineer was Jack Palmer. Thanks for tuning in to PsychTalks and see you again in two weeks' time. Bye for now.
