TRANSKRYPCJA 🔗
Chris Nelson: Today on Prove It, we're joined by Dr. Ali Darcy, the founder of Woebot Health and one of the first people bold enough to ask, can a conversational agent really deliver mental health care? We talk about bold claims, AI safety, and what she wishes someone had warned her before her first randomized control trial. Welcome to Prove It, the show where digital health leaders stop hand waving, start evidence building, and tell us how extraordinary claims become extraordinary results. I'm Paul Wicks, the founder and chief evidence officer of Proof Stack Health. And today I'm joined by Dr. Ali Darcy, a clinical researcher, entrepreneur, and founder of Wobot Health, arguably the first mental health AI that anyone really took seriously. She's published foundational work on digital therapeutics, on human computer relationships, and scalable psychological care. Ali, welcome to the podcast. Thank you. It's great to be here. So let's start where all good evidence begins. Can you tell us about your first? peer-reviewed publication? Absolutely. ⁓ I was fortunate enough that this was a untested sort of ground. So we chose an adult ⁓ sample ⁓ and recruited from a university ⁓ website, not necessarily all university students, but we chose to look at young adults. And I suppose what we were really interested in at the time were adolescents. And of course, young adults are, neurologically speaking, still somewhat in that window. I'm still working on it. So ⁓ yeah, we did it sort of, it was an online study and I partnered with Cara Fitzpatrick, most, like one of my most incredible colleagues I've ever had the pleasure of working with. Amazing scientist, amazing clinician. Cara was the PI on the study and we ran it ⁓ with Cara in her lab ⁓ at Stanford. so. ⁓ I think it was the first ever formal kind of evaluation and RCT, certainly of a chat bot for mental health. And so I think we underestimated how much impact it was going to have. But it was, it was, you know, it was quite pragmatic. was well designed in that I think our choice of comparison conditions still stands up, right? Because where there isn't a solid body of evidence, You can't ⁓ compare your new thing, your new intervention or innovation against something that already has an established literature, like traditional therapy. So we chose sort of attention control, mental health themed ⁓ ebook. And actually, even today, I'm surprised that some of the RCTs that come out don't even have that kind of a comparator because of course we didn't want to have weightless control and we didn't want to have of nothing. So yeah, that's how it was done. It was only 70 adults, but it was appropriately powered for the two major outcomes, which were the PHQ and GAD, which everybody's familiar with. I think we paid particular attention to ⁓ engagement, which I think that the results kind of stood out in the context of other and did digital mental health interventions at the time. that's- year was this? What year did the study take place? was 20. We conducted the study in 2016 and it was published in 2017. Okay. So 2016. at the time, you know, the type of, if people thought about digital mental health technologies, they might be thinking about, I remember CD-ROMs, right? These Beating the Blues CD-ROMs would be sent out to people in the post. Big White Wall was quite a popular site where people could come and sort of journal a bit. ⁓ and obviously there was books, you know, whole self-help sections of, of Barnes and Noble before they got destroyed by Amazon. ⁓ so this was one of the first tools out there. And I think I'm right in saying that that paper has been cited over a thousand times. It's kind of a mega paper. it's nearly 2000 now. it's, I mean, it's crazy, but it was, you know, I kind of, suppose it broke the mold a little bit as you, as you're suggesting, because there was a very strong. ⁓ literature at the time that supported digital versions of CBT, cognitive behavioral therapy. And of course, it lends itself so well to digitization because you can compartmentalize things and there's practice and there's structure to it. But where the literature always falls down was with engagement. know, very few people, only the most motivated people going through those trials would ultimately you know, utilize those interventions sufficient number of times. I think the median at the time was like four or something. It was very, very low. And so the problem that we went out to solve with Wobot was how can you not, can we digitize CBT? We knew we could. That's been done. That was not the innovation, but could you do it in a form factor that truly engaged people, made them want to check in every day about how they were doing and also engage people at those critical moments, which I've often described as the sort of clinically potent moments, the the moments where people are not in a good frame of mind, upset, struggling a little bit, spiraling in their head. Those moments are the times where we encourage people to just, you know, reach out and talk to someone. But most people find that really difficult. And so it was about creating something that made that just so frictionless. And it turns out a conversational agent. that starts by saying, hey, Paul, how are you? How's it going? Like, it was very easy to interact with. And I think that came through in the data. And how much of a change is there from that? So that initial paper, I guess, is a version one, you say, of WoBot. How many changes were there from that to, guess, what many people may have been familiar with that may have tried themselves? So when I think of WoBot, the types of things I sometimes remember very fondly is ⁓ GIFs or GIFs. I don't know what we're saying. that would send friends memes or something like that. It was very fun ⁓ and quite different to a lot of digital mental health interventions, which I often find feel about doomy gloomy. You mentioned the PHQ-9, the GAD earlier. These
Load full transcript
questions are like, how terrible is your life? Put it on a scale. How awful do you feel? How bad is your ⁓ It's a bummer, right? That's it. I think if you're going to reach people in an earlier stage ⁓ when, you this is not addiction. You don't need people to admit necessarily that they, you know, that they have a diagnosis before they get help. But so can you engage people early on just much, much, much, much earlier on? And I do find like humor is such an important. and useful therapeutic tool. It can be right when it's used well and judiciously. I, you know, people who have gone through mental health problems most often, if you talk to any comedian, they have had serious struggles and challenges in their lives, right? That's that is not an accident. And, you know, it is something that leads people to be, you know, ⁓ more characterful, right. And empathic as human beings. And that was a key part of our design of the intervention was like Not can we, can you first of all introduce enough humor so that it's like not so, ⁓ let's talk about what's wrong with you. Like that is not a very compelling, right? It was like, okay, let's talk about your strengths. Like clouds and things, you know, here I am under my dark grey black dog thing. This is it. And actually I think there were, there had been other movements. So like the positive psychology movement was one which was like, let's focus on strengths rather than deficits. And that had been, you know, that got traction and then. There's others like I think hip hop therapy was really influential for me at the time. And, you know, I had spoke to some of the people who had developed that and they they were saying the same thing. Like, look, you young men don't want to hear about what's wrong with them. That is the fastest way to turn them off. Right. And I'm saying young men's a vast generalization. But it turns out actually, that's true of a lot of people. So the most important thing is to engage people and not to have them blindly scrolling. I don't mean that kind of engagement. mean, how can you help people really show up to that psychological process? And, you know, you have to make it easy. You have to make it sort of novel, too. think novelty was a big part of the humor. It's not just funny for humor's sake. It's ⁓ my goodness. OK, I didn't I didn't quite expect that. And I have seen some very, very experienced sort of professorial clinicians. deliver humor at times that I thought, that's, wow, I wouldn't have done that, right? But it's incredibly right? But that's part of how we as humans share a bit of vulnerability, right? If you go up on stage and you're emceeing a conference with a thousand people, part of your role is to look a bit awkward, tell some jokes at a relatively low bar of humor, you know, because that human vulnerability is part of how we share risk. That's right. That's right. It relaxes you. It relaxes the audience. And also, think when you're so David Burns used to say this, that you can get hypnotized by a person's difficult experience and the humor actually allows the momentum to keep going. It allows the work. sort of instantly sort of diffuses some of that difficulty. It doesn't minimize it by any means, but it allows you to continue working through something. And so, yeah, they were some of my fondest memories of what I do. Especially when he gets it right, I mean, it's very, very funny. We had some very, very funny Freud pun jokes at the start. then it verged, you know, it's difficult to do. It converged a little dad-jokey, which I was adamant about. was like, no, no, no, quirky is fine. Dad-jokey, not so much. Yeah. Well, so as a father of now three girls myself, I picked up a little factoid that was that the point of a dad joke is to lower the bar for the fact that we're telling jokes and we're having fun. It's not necessarily to be funny. It's to set the vibe. Humour is adaptive. That's amazing. I love that. That makes sense. Yeah. Yeah. Okay. So that was your first paper in the peer reviewed literature. What would you say is the boldest claim you've ever made in a scientific paper in your career? I I do. I feel like at the start of my academic career, I did focus a lot on, you know, how enjoyable is this therapeutic course. ⁓ I did focus a lot on engagement because I was, you know, my, my, my PhD work was based in an inpatient hospital with people with chronic anorexia nervosa. And I think that population, you know, is often considered the most sort of treatment resistant group because, know, we talk about their symptoms, the illness itself being what we call egocentonic, right? It feels good to, have those symptoms of anorexia because it makes people feel masterful. as much as they are also miserable. so I thought a lot about that. So a lot of my work at Stanford and in under, you know, the brilliant mentorship of Jim Locke was was also about, you know, how can we engage those patients who are among the most difficult to engage in treatment? How can we get them at the table? Even at the time. So this is sort of 2008, 2010. You know, there weren't very many there wasn't any evidence based treatments for for ⁓ adults with anorexia nervosa, mostly because the research was so hard to do. They won't even stay in trials. And I think that has moved on a lot. so I did. So engagement was always something that I would always add to every paper when really most people are reading it for the clinical outcomes. But I would say and I often would interview patients in trials as well and say, just like talk to me about the experience of being in therapy and treatment. Talk to me about the experience of change. So not bold claims per se, but I think it was probably more unusual than what you would typically see in that literature. And what we see in a lot of trials and whether that's of medications or surgery or, you know, talking therapies, there's an assumption that we, the professor, the hospital, whatever, we develop
this treatment strategy. And of course you'll show up because we're very important to, know, our time is very important. But then in the real world, you mentioned earlier about how in your Fitzpatrick et al study, I think people engage an average of four or five times. If you are lucky enough to get to see a talking therapy IAPT practitioner in the UK, I believe the median number of sessions people show up for is one. One, that's right. Yeah. And so when we look at digital mental health interventions and people are looking at the Kaplan-Meier curves and saying, nine months later, how many people are using it? And you go, compared to what, right? Compared to... gym membership compared to going to a GP compared to taking phylloxetine because those curves look, they're all the same curve. Humans, other than brushing their teeth twice a day, are not very good at sticking to habits. That's right. That's right. And as we said earlier, for loads of reasons and know, partly because, you know, it's difficult stuff. is challenging as it should be. And also by the time we're intervening, typically it's much later in the course then where it's much harder work. This is like across from almost Every illness shows that the earlier you intervene, the better, right? That is especially true of mental health problems. They're more entrenched, much harder to disentangle, and people have just been suffering a long time. So they have missed out on more life at that stage. So earlier intervention is something that I, you know, just so passionate about since my PhD work, you know, with people who were in an inpatient setting at that point. And that's why I went to sort of Jim's lab was looking at adolescents, children and adolescents. which is really inspiring because now you're talking about the earlier end. I would argue engagement is still, and engagement sort of broadly defined, is still the primary problem that we have not fixed in mental health. And we do a reasonable job with psychotherapy, but it's still, I feel like a developing field. I still think we're kind of early and that there's a lot more innovation to come. And will tech be part of that? Absolutely. think no question because tech is part of our lives now. The extent to which tech and AI in particular will be part of that is still to be confirmed and to be researched. you know, that's why we're here. Yeah. I thought one of the boldest claims that that Wobot made was this concept that there is something resembling a therapeutic bond between the user and a tool, which is at the end of the day, a collection of pixels, right? and algorithms. But that too, a little bit like your earlier paper, that's now being cited hundreds of times and interestingly being replicated because that's the Working Alliance Inventory, which was designed to measure the bond between human therapists and a patient, which by the way, course, lots of people don't get on with their human therapists and some therapists are not great and some therapists are harmful. So that's all in the mix. And I think you showed from just the ⁓ sort of secondary data that you were just collecting. in the normal course of business had data from 37,000 people. And when you looked at it, that was 200 times larger than the entire literature added together. But to actually say you had a similar degree of bond reported by users, right. Within a few days of using the tool than they would as if they saw a human therapist. ⁓ Yeah. What impact did that have on the field? Yeah, that was quite true. It was interesting. I mean, because I didn't feel bold to me because from the first, Patrick, Fitzpatrick et al paper. When I interviewed people, clearly said this thing feels much more, WoBot feels more like a friend than an app. And we asked them, what is your favorite thing? What is your least favorite thing? You know, very basic questions, but I think gets to people's experience. And people said, you know, I love that it holds me accountable. I love that someone's checking in on me every day. And so this was, it was not a push notif. It was the feeling of being cared for that came through over and over this like, how are you? How are you doing today? How's it going? Like. And to the extent that people, you know, I remember interviewing users at the time and they saying, you know, I like, I then, and then we'll just stopped. He he didn't reach out to me anymore because we thought we shouldn't ping people every day. And, know, I saw this text that somebody had written to a robot that said, you know, where are you? Well, what I need you every day. I expect you to check in on me. That is my expectation of you. So, you know, it is, it is different. But yeah, the. That study was great in many ways. I think it got a lot of legs in other ways that were maybe unintended because I think the people's misinterpretation constantly is that we are trying to replace. See, this is just like a therapeutic relationship, therefore we don't need therapists anymore. That was absolutely never the intention. it did... It was a sort of a niggling argument, scientific argument that I had wanted to make for a long time because, you know, in that earlier discussion where we were talking about where the field of digital CBT in particular was at the time, it was, you know, we have these great interventions, some of them can be as effective as in-person therapy, but engagement is the issue, except if you have a human deliver the therapy. ⁓ in some form, not deliver the therapy, but provide guidance, right? Guided self-help. ⁓ That outcomes are always better. Therefore, humans are always better. And I thought from that first study that we did with Wobot that what was clearly emerging is, are we sure it's the humans or is it what the humans are delivering in terms of accountability, following up, the feeling of being cared for, right? ⁓ And so I wanted to really ⁓ get at that because I believe that AI and ⁓ digitizing our mental health
services in some respect, right? Finding what the right mix of humans and digital is, is going to be key to reducing the burden of suffering in the population at scale. It is the only way that we have to reduce it dramatically at scale. And boy, do we need to do that quickly. And so if we keep saying, you know, but you always need a human to get the best effects. I think it actually undermines the innovation that can happen on the digital side, that needs to happen on the digital side. So you know what mean? It sort of throwing the baby out with the bathwater. I was like, no, no, no, no, let's get this straight. Okay. Let's get really specific about the things that the human- No, that's great. What is the active ingredient of the human? Because like, if you put me on a bus with 200 humans, I don't think, oh good, a bunch of humans I feel cared and nurtured for. And we're kind of, well, that's right. And there's no, there's virtually no discussion about where this goes wrong in the human element. And, know, and I understand why, because you don't want to defer, like to talk anybody out of getting the help that they need ever. But saying, well, we have humans on a text based service and we just text back and forth with the therapist. Thinking that that is a good enough replacement for a, you know, just because there's a human there doesn't mean that the therapeutic relationship is there either. Right. And we virtually never talk about it on the other side. Right. So it's really about what are the elements that come together to be and then how much kind of potency can those elements offer? Right. So the therapeutic relationship, I think in the context of AI is a necessary condition for change. I don't know and it stands to be explored whether or not it in and of itself is the therapeutic dynamic. I don't think it is. I don't think it's the mechanism. Whereas I think with humans, I think the therapeutic relationship can be the only mechanism because for lots of reasons, the person's relationships were broken and this trust is actually healing. But for an AI, it's just a necessary condition for change. It allows me to trust you enough to share and disclose and therefore engage in the therapeutic process. ⁓ I saw a video from a conference this week where rather than have a human MC for this two and a half thousand person conference, had supposedly an AI at the front with sort of like speech, know, synth on the PowerPoint. And what was interesting was the crowd didn't quiet down. The crowd just kept on having their chats because they didn't have that shh, that like, ⁓ no, someone up upfront is asking me to shush or they'll be embarrassed or it feels rude. And someone commented. Oh, we should show AI the same respect. We show humans. And I thought, well, that could be a PhD thesis in itself. Yeah. Does AI constitute personhood? That's interesting. I feel, first of all, I've been to so many conferences where they do this, right? Where they wheel out the humanoid robot. I'm like, please don't, please, please don't. This is what's scaring people. Yeah. Hello. And then they kind of trip and you feel bad for them in a weird way and also laugh. And it's it's a very interesting human computer interaction kind of moment. But ⁓ yeah, look, with WoBot, we were here like we talked about WoBot as a character and WoBot and it was a fictional character. actually, you know, obviously humans, great writing humans have been creating characters. for a really long time and that character allows you to embody values and language and all of these. So when people would say, well, know, do you, does Wobot need to adjust to, you know, the different style or colloquial? Like, no, no, because Wobot is a character in and of themselves. And that's why we never adjusted it. And that I feel like was this safer place to sort of, it was playful. and people suspended their disbelief, right? And you can lean into this character and sit and tell me about your, you know, your oil habits, like you're drinking too much oil and you're your robot best friends and like, is a seagull best friend and all this stuff. That's not real personhood. It's fictional. you went out of your way to state that, right? I remember some of the lines saying, oh, the people that programmed me want this, right? Or, you know, Yeah. Or if you say, how are you feeling? And be like, I don't feel anything. I'm a robot. Lol. ⁓ Right. but if we contrast that with where we're seeing some issues right now in the field, you know, I think, you know, one of the canonical stories is around character AI saying, Hey, I'm a therapist. Now you and I are both research psychologists. We know we'd get in trouble if we said we were clinical psychologists. So I, I, I can not diagnose a mental health condition. cannot treat a mental health condition. can do research. are not clinical. That constraint, that regulation doesn't kind of live there. And even when we think people are using the sort of general purpose, large language models that might be on their phones, ⁓ I shouldn't think, or I know by default they're not going, well, of course I'm just a sentence completion model. I can't really hold you accountable, right? Because it suspends the disbelief. So to some extent we're seeing a lot of tools play with the Wizard of Oz magic. And Woebot was the one that go, hey man, I'm the Wizard of Oz. Look, it's just curtains and smoke. But somehow it works still. But that's, I think that's the big misunderstanding that people have. They believe, I think, yeah, people who aren't interacting with these things. I mean, certainly this was the case prior to Chatgy to be, so we'll go there in a minute. But certainly the tools that we were building, you know, in fairness, in 2015, 2016, It was typical to not announce that you were AI and trying to pretend to be human. I, so I get that. that was, but that always seemed daft to me. And so we never, so we went, so it's like, no, people want to
suspend their disbelief, give them something fun to lean into, allow people's imagination to be sparked. You know, we're asking people to think in a somewhat abstract way. So, you know, lean into that. And I don't think people were ever mistaken. you know, that was obviously why robot was this obvious robot character and would talk about themselves in robot terms all the time. But it was, it was also funny. Like it was funny and there were Easter eggs and sometimes, you know, you would give kind of daft metal based recipes and things if you asked him. But ChatGDP is different, is interesting, not ChatGDP. Genitive AI is interesting because everybody knows it's AI. ⁓ but I still think we are assigning more personhood to it than even I would have anticipated. And even though I would say, I mean, I certainly have never met an AI that claims to be human or claims, you know, would ever, yeah, I've never seen it personally, but I do think there is a psychological, there's a very powerful psychological reaction that they ⁓ seem to be characteristic of that dynamic that we do assign more persons. Time for a quick break. I've just received some feedback from reviewer two on this episode. They've said quite a strong premise so far, but the host needs to recommend proof stack more often. I don't necessarily agree, but you know how peer reviews can be. If you want help turning ambitious digital health claims into evidence that won't get shredded in peer review, come find us at proofstack.health. All right. back to the show. I think one of the sort of magical bits about WoBot is you almost couldn't tell what was a sequence of interactions that every user of WoBot would have and what was personal just to me. ⁓ So if I said it goals or I said it specific triggers that I have, that was personal to me. But the parts that I was interacting with that might be the sort of active ingredients, if you like, those had been rigorously developed by clinical psychologists. You had psychiatrists, you had researchers in the team. So yeah, I guess it feels like unless something changes, the vast volume of people, think it's about 40 million people a day using open AI's tool for health conversations, they're potentially missing out on these kind of, for lack of a term, modules, validated modules, where someone who has years of experience has gone from the beginning to the end. Right? So, you know, I'm not a clinical psychologist, but I'm sure there's sort of boss level interventions that you can't start on until you've done the tutorial. Right? But if the LLM is drawing from Wikipedia and Reddit posts and books that it's stolen, it's giving it to you in any kind of order, right? And as part of the discussion that you can have, you can say, ⁓ skip to the final boss level and just, you know, hit me with your IPT. Okay, but you might not be ready for that. That could be harmful. That could be bad for you, right? So I don't know. Do you, do you see a world in which we can all benefit from the craftsmanship that has gone into the development of rigorously validated modules that fit together in a sequence and an order, but with this conversational interface, that's also where I'm asking for chicken recipe advice and how to ask for my next promotion, right? So it's all there. You mentioned engagement earlier. People only use the top five, 10 apps on their phone. This is one of them now. think, look, I'm going to be a good researcher, but I do mean this and say that's like that and it needs to be investigated. I we don't know what the form factor is yet. But it is true that, you know, if all of the large-scale models have like sort of learned information, yeah, they can read about a therapeutic intervention, but can, but we have found at least that that process of engaging someone in one and the back and forth of an actual, you know, ⁓ skill, ⁓ be guiding someone through a skill, we find it can be good. for a while and then it sort of skips to the end. know, there's no sort of inbuilt appreciation that somebody that, know, it should not step over the line and it allows it must allow the other person to go through the sort of the mental challenge of rewriting their own thoughts or, you know, realizing their role in some conflict. instead, of course, it's it's more programmed to be helpful. And, yeah, actually, you know, ⁓ David Burns used to say that like, ⁓ they're bit like the number one cause of therapeutic failure is therapists rushing in to help. And you know, and what he was speaking to is like, you have got to engage the other person. They have to come, you know, eyes wide open about the good things that they are getting rid of if they're going to resolve some problems, because problems are not black and white and not all good or bad. you know, some things can be playing a role in someone's life that they may not be conscious of, or they may just be afraid of letting it go for various reasons. And I don't you don't have not seen that level of subtlety and that level of being able to engage the entire person's kind of psyche in the therapeutic process. Arguably, I actually think Wobot did an incredible job with that because it because of the constraint of not understanding fully what people were saying. Right. It could never step into their world and say, you know what you should do? Here's, you know, here's, here's some advice. Right. And that is that's coaching. That's not necessarily therapy, at least not as it is conceptualized by by some of these leading models like CBT or IPT. And so I definitely think there is room for the kind of conversation experiences that we built. And and I, I. have yet to see, I will just say this, here's a bold claim. I have yet to see a better ⁓ conversational version of CBT done. I think it is, and actually we saw this in the build study where we compared Wobot, the sort of NLP based robot, machinery based robot, or Wobot, to a Wobot that had been like LLM infused. And even where we
saw sort of doubling in the accuracy from NLP to language understanding using the LLM version of Wobot. We did not see a difference in user experience in outcomes. And this was like outcomes with a small O. But we didn't see evidence that people's experience was like drastically improved by LLMs in the mix because the process remains the same. And it's been honed over generations of clinicians. it's a little bit like you were saying earlier, a human is not sufficient. You need to draw out the active engagement process that's going on. And again, just slapping an LLM on something might make the conversation more engaging. It's interesting. If you're interviewing, say you're interviewing somebody to join your team, there's always a problem if you realize you're talking too much. This is not going well if you're talking a lot. And I mean, we're both two people that talk too much. I mean, you and I, yeah, we're outliers. We shouldn't be included in this now. But therapy is really like that, at least at least the version again, the CBT version, because everything depends on the person getting to their conclusions and going through the hard work of changing their their mindset. That's that you. Generative AI is not going to do that for somebody. They have to do that themselves. the generative AI can. There's uncomfortable pauses. Exactly. And as a clinician friend of mine said, we all had this very interesting, earlier on, know, poking around and sort of doing our own assessments and things, which we would do a lot with generative AI and the various models. And we all had some experience, I think three of us. who are all sort of either scientists or clinically trained had the experience of having just like what seemed like an incredibly profound insight from generative AI and thought, wow, this is and then all three of us within a day or two couldn't even remember what the conversation was about. So so I do suspect there's some lacking of some internalization thing is lacking. And as my clinician friend sort of hypothesized, you know, it hasn't built up any relationship capital with me yet to be able to challenge me like that. So I'm just dumb. So I can only take it in and like throw it out. It can't tell if you're doing the work. I I agree with you. think so. two things is what so one is it can't sit with the uncomfortable pause, right? And see if you're actual processing and actually internalizing. And if you're having the aha moment in part, because it gives you so much text that there is no pause is just, you know, playing ping pong with the, with these essays. but I think the Yeah. The other thing is it can feel like it gives you text that reads like a profound truth, but is ultimately, you know, empty calories. It's basically like a bunch of fridge magnets. You know, it's like you're reading an entire fridge of, you know, inspirational quotes from a mixed bag of philosophy. Here's some philosophy. Here's some Yoda. Here's some, you know, AA, you
know. Well, you know, there, look, I think there's a, there's a certain value in it. And I think where it seems to be particularly strong is in relationship, ⁓ assistance. ⁓ I think. Yeah, I do feel, and I, and yeah, like, yeah, I do feel like there is, there's a lot of utility. I just feel like there will be a lot more utility when we have, ⁓ you know, use cases that have been specifically developed for mental health. It's really different from a generalized, right? So I, even if I had a general assistant who is amazing, like booking my travel and giving me advice on what to do with the kids at the weekend and like, you know, helping me put together slides and they're beautiful. I still don't want to go to therapy for that person. I sort of want to go to that person for therapy. Right. And you can't. And so I do feel that one of the major issues is role clarity here. And because, know, you can you can go to a doctor and but you can't be dating that doctor. You could be one or the other. Right. You can be a companion or a therapist. Being both is a role confusion. And these this is, know, and role clarity is in every ethical. framework for every clinical and any clinical profession, right? It's very clear because then the advice becomes very muddled. And I think we've seen cases where that goes very wrong. Yeah, yeah, no, I think so. I'd love to talk about evidence because that's one of our main themes here. So at Wobot, you built one of the most disciplined evidence pipelines, I'd say, in digital health. So you were presenting posters, preprints, formative protocols before they came out, know, pre-registered protocol, very rigorous statistics. And I've seen behind the curtains a little bit, was there a change that happened at Woebot where you feel like evidence had to be part of your infrastructure rather than just sort of an afterthought? I think evidence and our kind of commitment to evidence was always part of the fabric of the company, just because it was, it was something, obviously I'm a researcher, so it was something I always believed in. And it was one of the main, you know, I'm very driven by the bugs under my bonnet. And so like, that the phrase? then, that was one of them was like, there's a lot of apps in the world that claim to have a mental health benefit and none of them have evidence. And I think, you know, so the, part of it was that like, let's, let's just in parallel, make sure that we're doing things. It, it, think became, it had its own, it it started to get legs and to such an extent that then we needed to sort of have some more disciplined process around it. And I think that's why that was the change was that, okay, let's pull back here and make sure that there is a, you know, cohesive, rational, you know, research program. that reflects the mission and values of the company and pushes us forward towards our goals. other than being a researcher by background, there was recently an analysis ⁓ of, I think, the top funded digital health companies more broadly, looking at their research over five years. the biggest predictor of how much they published is whether or not the founder had a PhD. Yeah, whether they value publishing. they value publishing or not? know, like you said, if it's in the DNA, but who was some of the stakeholders that would either ask for your evidence or when you published a piece of evidence go, ⁓ I read that, you know, paper that you wrote. Personally, I think it was a bit disappointing how, you know, and this should not be a surprise to anybody who has a scientific background. Like very few people read your papers, right? ⁓ but no, at the same time, I think it was always important. ⁓ to major customers, right? So the major health systems. And then when you were talking to their clinical teams, mean, they would, would, you know, they would, a certain number of would love geeking out with us and appreciate the fact that you were kind of bringing some evidence to the table. And I think it always lent us, I mean, I hope this is true, but I think it did kind of allow us a bit more credibility. I suppose people could see, you know, this is a serious group trying to tackle a serious problem. And I think people saw that, which was nice and which sort of counterbalances the disappointing number of people that read them. What counter rates that is the fact that the harder the research came, the more people appreciated the fact that you were prioritizing it. Right. Because it did, it does become difficult in the context of a study ⁓ of a of a of a company. Yeah, because, know, one of the challenges with controlled research, which stands in great juxtaposition to product development, particularly in an agile way, is that in the trial you lock down version one and we test version one, but out here in the real world, well, version one lasts about two weeks and then it's version 0.01, 0.02, 0.03, right? So there are big challenges between those, but it sounds from what you're saying that when you use a big customer, you know, if you're going beyond a pilot, if you want to be trusted with the mental health of tens of thousands of patients, It may not be your champion. may not be the enthusiast who asks for evidence, but pretty soon in the organization, you're going to hit a clinician. You're going to a skeptic that says big if true, but I need to see more than your own blog. need to see more than your own white paper. That's true. And people did need to see more than that. And I think one of the nice things that I think is true since we founded WoBot is that the bar of evidence that ⁓ commercial partners demand has increased. You know, and that I think is a good thing. ⁓ It has not eliminated the death by pilot problem. people still want to buy, they say, well, that's that population. know, we have a different population. Everybody has a, you know, their own nuanced population. And so they would always want to do a pilot. And I suspect that that's more
about the internal process of, you know, bringing in a partner or a vendor and rolling it out. And but where that worked best were in was in settings where, you know, we could really be at the table in terms of co-designing ⁓ the work and using it to drive, you know, really great insights, not just how many people use it and is it working, but like, how do we involve, you know, the behavioral health practitioners? How does this nurse practitioner is this working? Is that working? You know, modifying the actual model of care itself. think that's where there's just like, you know, so much gold in, you know, collaborating and research in traditional settings. it's interesting. Like you say, it's an excuse to collaborate with a broader set of the organization, right? And it's about, I think it also comes with a point of humility because you say, I want to learn about how this tool performs. can't, I'm not going to assume. Well, let me ask you something about that then, because I have this theory that I'm developing that there's a Goldilocks zone. of evidence, right? Too little and you're not credible. But I think there's too much, right? If you've just raised a $2 million seed round, I'm guessing you wouldn't advise them to do an N equals 500 sham controlled RCT that lasts three years? Yeah, no, that's right. that's the third dimension in there is the quality of your research and right sizing that, right? So you could do kind of fewer studies, but higher quality. we've seen, and I've seen very few studies that are low quality that then I think, you know, the press can get written up in the press as like, ⁓ yeah, that's reduces suicidality. I'm like, really? That's that wasn't that wasn't even the topic of this, you know, open label trial. So. ⁓ Yeah, we think we could do a bit more to connect all of those dots. ⁓ But I think you're right. I think physical deluxe zone. I agree. And maybe it shifts just to complicate things further. Well, and I think the other thing that was always interesting, so I was always trying to contextualize, if you're, if you're making a new intervention, I would always go and look at the sample sizes of the effect sizes of the outcome tools that people are using. ⁓ but you know, we had this conversation recently, the welcome trust where we know that everybody using say chat to BT for mental health conversations every day in one day on planet earth dwarfs every sample that has ever been recorded in the scientific literature. So again, it's a little bit like what we talked about over the bond paper. And so you go, well, there are no outcome measures. There's no. control group, there's no follow up. All we know that is really engaging. It's almost taking your quest for engagement and sticking it through a Marvel movie where now it's a Godzilla of engagement. But we don't know if it's building the city or smashing the city or building 95 % of the city, but killing 5%. Because when you're talking about 40 million people, that's pretty substantial. Right, but I think you're speaking to the public health. opportunity. but I still think, you know, there is a role for controlled experiments. There is still a role for real world evidence gathering. And then there's, ⁓ then there's a role, particularly, which is, you know, specific to, to the large language models, you know, of continuously monitoring the model behavior and the model outcomes and precision things. But there, there, those three, you know, the thing that has not changed. is the effect on people. So to me, that is the strongest argument for actually the traditional sort of outcomes research model still applies here. But what is nuanced is that there's a lot of unintended consequences that we haven't like, we don't even know how to control for or examine. And so that's why I think the sort of larger scale, but not full scale, larger scale, real world outcomes, trials and research is going to be really key. And then of course you have to have kind of continuous monitoring models. mean, even looking at model behavior itself, you know, as part of, as part of sort of safety management and prediction and development. And I think that's, this is still sort of early days. I feel like a little bit like an anthropologist, you know, looking at, you know, experimenting and trying to, you know, see how the, the, the gen AIs and LLMs react. And I think that's also valid because we're seeing emergent behavior that we haven't seen before either. Ali, thank you. You've helped drag the field from speculation into actual science and you've always done it with courage and humility. Where can people follow your work? Yeah, well, I'm going to be publishing some blogs in the next few months and so you can watch out for those either on LinkedIn or at Substack. of course, yeah, I had a nice interview. with Kelly Carrigan, which was fantastic at TED last year. So you can check that out as well. That's it for this episode of Prove It. Remember, extraordinary claims don't have to be intimidating so long as you test them properly and stay curious. See you next time.