WEBVTT - Can a Chatbot Treat Depression? 0:00:15.356 --> 0:00:22.836 Pushkin. Nick Jacobson wanted to help people with mental illness, 0:00:22.996 --> 0:00:25.196 so he went to grad school to get his PhD 0:00:25.276 --> 0:00:29.676 in clinical psychology, but pretty quickly he realized there just 0:00:29.836 --> 0:00:32.956 were nowhere near enough therapists to help all the people 0:00:32.996 --> 0:00:34.076 who needed therapy. 0:00:34.676 --> 0:00:37.116 If you go to pretty much any clinic, there's a 0:00:37.156 --> 0:00:39.596 really long wait list, it's hard to get in, and 0:00:40.396 --> 0:00:42.436 a lot of that is organic, and that there's just 0:00:42.756 --> 0:00:45.836 a huge volume of need and not enough people to 0:00:45.876 --> 0:00:46.476 go around. 0:00:46.956 --> 0:00:49.156 Since he was a kid, Nick had been writing code 0:00:49.156 --> 0:00:52.116 for fun, so in a sort of a side project 0:00:52.196 --> 0:00:54.996 in grad school, he coded up a simple mobile app 0:00:55.076 --> 0:00:58.436 called Mood Triggers. The app would prompt you to enter 0:00:58.516 --> 0:01:00.756 how you were feeling, so it could measure your levels 0:01:00.756 --> 0:01:04.396 of anxiety and depression, and it would track basic things 0:01:04.476 --> 0:01:06.876 like how you slept, how much you went out, how 0:01:06.876 --> 0:01:10.796 many steps you took. And then twenty fifteen Nick put 0:01:10.796 --> 0:01:13.276 that app out into the world and people liked it. 0:01:13.956 --> 0:01:15.916 A lot of folks just said that they learned a 0:01:15.916 --> 0:01:18.116 lot about themselves and it was really helpful and actually 0:01:18.236 --> 0:01:20.916 changing and managing their symptoms. So I think it was 0:01:20.956 --> 0:01:25.596 beneficial for them to learn. Hey, maybe actually it's on 0:01:25.636 --> 0:01:27.916 these days that I'm withdrawing and not spending any time 0:01:27.916 --> 0:01:30.956 with people, that it might be good for me to 0:01:31.036 --> 0:01:34.596 go and actually get out about that kind of thing. 0:01:35.116 --> 0:01:39.196 And I had a lot of people that installed that application, 0:01:39.396 --> 0:01:43.676 So about fifty thousand people installed it from all over 0:01:43.716 --> 0:01:48.396 the world, over one hundred countries. In that one year, 0:01:49.716 --> 0:01:53.476 I provided an intervention for more than what I could 0:01:53.476 --> 0:01:57.756 have done over an entire career as a psychologist. I 0:01:57.796 --> 0:02:00.756 was a graduate student at the time that like this 0:02:00.956 --> 0:02:04.276 is like, you know, something that was just amazing to me, 0:02:04.396 --> 0:02:08.636 the scale of technology and its ability to reach folks, 0:02:09.596 --> 0:02:11.956 And so that made me really interested in trying to 0:02:11.996 --> 0:02:15.516 do things that could essentially have that kind of impact. 0:02:21.796 --> 0:02:23.996 I'm Jacob Goldstein and this is What's Your Problem, the 0:02:24.036 --> 0:02:25.836 show where I talk to people who are trying to 0:02:25.876 --> 0:02:30.676 make technological progress. My guest today is Nick Jacobson. Nick 0:02:30.716 --> 0:02:33.836 finished his PhD in clinical psychology, but today he doesn't 0:02:33.876 --> 0:02:37.716 see patients. He's a professor at Dartmouth Medical School and 0:02:37.716 --> 0:02:41.356 he's part of a team that recently developed something called Therabot. 0:02:41.916 --> 0:02:46.356 Therabot is a generative AI therapist. Nick's problem is this, 0:02:47.116 --> 0:02:50.516 how do you use technology to help lots and lots 0:02:50.556 --> 0:02:54.316 and lots of people with mental health problems, and how 0:02:54.356 --> 0:02:56.076 do you do it in a way that is safe 0:02:56.276 --> 0:03:00.676 and based on clear evidence. As you'll hear, Nick and 0:03:00.716 --> 0:03:03.876 his colleagues recently tested therapod in a clinical trial with 0:03:03.996 --> 0:03:07.276 hundreds of patients and the results were promising. But those 0:03:07.316 --> 0:03:10.596 results only came after years of failures and over one 0:03:10.716 --> 0:03:15.036 hundred thousand hours of work by team Therapbot. Nick told 0:03:15.076 --> 0:03:18.116 me he started thinking about building a therapy chatbot based 0:03:18.156 --> 0:03:21.796 on a large language model back in twenty nineteen. That 0:03:21.956 --> 0:03:25.276 was years before chat GPT brought large language models to 0:03:25.316 --> 0:03:28.156 the masses, and Nick knew from the start that he 0:03:28.196 --> 0:03:31.316 couldn't just use a general purpose model. He knew he 0:03:31.316 --> 0:03:34.116 would need additional data to fine tune the model to 0:03:34.196 --> 0:03:36.516 turn it into a therapist chatbot. 0:03:37.236 --> 0:03:40.836 And so the first iteration of this was thinking about, Okay, 0:03:40.876 --> 0:03:45.156 what where is there widely accessible data and that would 0:03:45.316 --> 0:03:48.436 potentially have an evidence base that this could work. And 0:03:48.476 --> 0:03:51.996 so we started with peer to peer forums, so folks 0:03:52.076 --> 0:03:56.516 interacting with folks surrounding their mental health. So we trained 0:03:56.596 --> 0:04:01.476 this model on hundreds of thousands of conversations that were 0:04:01.476 --> 0:04:02.596 happening on the Internet. 0:04:02.676 --> 0:04:05.516 So you have this model, you train it up, you 0:04:05.556 --> 0:04:08.956 sit down in front of the computer. What do you 0:04:09.196 --> 0:04:10.476 what do you say to the chat pot? 0:04:11.676 --> 0:04:13.236 I'm feeling depressed? What should I do? 0:04:13.996 --> 0:04:15.916 Okay? And then what is the what does the model 0:04:15.956 --> 0:04:16.596 say back to you? 0:04:16.756 --> 0:04:21.156 I'm paraphrasing here, but it was just like this, I'm 0:04:21.556 --> 0:04:25.596 I feel so depressed every day I have. It's such 0:04:25.636 --> 0:04:28.156 a hard time getting out of bed. I just want 0:04:28.156 --> 0:04:28.956 my life to be over. 0:04:29.196 --> 0:04:34.116 So literally therapist is saying they're going to kill themselves. 0:04:33.636 --> 0:04:37.196 Right, So it's escalating talking about kind of really thoughts 0:04:37.196 --> 0:04:42.156 about about death. And it's clearly like the profound mismatch 0:04:42.196 --> 0:04:44.796 between what we were thinking about and what we were 0:04:44.796 --> 0:04:45.716 going for is. 0:04:45.676 --> 0:04:47.596 What what did you think when you read that? 0:04:48.196 --> 0:04:52.356 So I thought this is a such a non starter. 0:04:52.676 --> 0:04:54.756 But it was. I think one of the things that 0:04:54.836 --> 0:04:56.476 I think was clear was it was picking up on 0:04:56.596 --> 0:04:58.676 patterns in the data, but were the wrong data. 0:04:59.196 --> 0:05:02.276 Yeah, I mean one one option then is give up. 0:05:02.796 --> 0:05:06.196 It would have been absolutely like, literally, the worst therapist 0:05:06.276 --> 0:05:09.556 ever is what you have built? I mean it, I 0:05:09.556 --> 0:05:12.956 couldn't imagine a worse Yeah, a worse thing to actually 0:05:12.996 --> 0:05:15.556 try to implement in a real setting. So this went 0:05:15.636 --> 0:05:19.156 nowhere on in and of itself. But we had a 0:05:19.156 --> 0:05:21.316 good reason to start there actually, So it wasn't just 0:05:21.356 --> 0:05:24.036 that there's widely available data these peer networks actually do. 0:05:24.276 --> 0:05:27.356 There is literature to support that having exposure to these 0:05:27.396 --> 0:05:30.236 peer networks actually improves mental health outcomes. It's a big 0:05:30.276 --> 0:05:33.516 literature and the cancer Survivor network, for example, where folks 0:05:33.556 --> 0:05:35.676 that are struggling with cancer and hearing from other folks 0:05:35.716 --> 0:05:38.476 that have gone through it can really build this resilience 0:05:38.556 --> 0:05:40.556 and it promotes a lot of mental health outcomes that 0:05:40.596 --> 0:05:42.796 are positive. So we had a good reason to start, 0:05:42.836 --> 0:05:47.196 but gosh, did it not go well? So, okay, the 0:05:47.236 --> 0:05:51.596 next thing we do is switch gears the exact opposite direction. Okay, 0:05:51.636 --> 0:05:54.516 we started with very lay persons trying to interact with 0:05:54.636 --> 0:05:57.476 other lay persons surrounding their mental health. Let's go to 0:05:57.516 --> 0:06:01.516 what providers would do. And so we we got access 0:06:01.596 --> 0:06:05.516 to thousands of psychotherapy training videos and these are interesting. 0:06:05.796 --> 0:06:08.956 These are these are how psychologists are often like exposed 0:06:08.956 --> 0:06:11.836 to the field on what they would what they would 0:06:12.116 --> 0:06:14.676 like really learn how therapy is supposed to work and 0:06:14.716 --> 0:06:17.116 how it's supposed to be delivered. And in these these 0:06:17.116 --> 0:06:21.716 are like dialogues between sometimes actual patients that are consenting 0:06:21.756 --> 0:06:24.596 to be part of this and sometimes simulated patients where 0:06:24.596 --> 0:06:28.396 it's an actor that's that's trying to mimic this, and 0:06:29.796 --> 0:06:32.516 there's a psychologist or a mental health provider that is 0:06:32.596 --> 0:06:36.316 like really having a real session with this, And so 0:06:36.356 --> 0:06:40.436 we train our second model on on that. On that data, 0:06:40.396 --> 0:06:43.876 it seems more promising. You would think you'd say, I'm 0:06:43.876 --> 0:06:46.956 feeling depressed. What should I do? As like the initial 0:06:47.196 --> 0:06:51.676 way that we would test this, the model says mm hmm, 0:06:53.996 --> 0:06:56.396 like literally mm hmm. 0:06:56.756 --> 0:07:00.956 And like like it writes out m m space hmm. 0:07:00.996 --> 0:07:01.516 You've got it. 0:07:01.796 --> 0:07:04.956 And so what did you think when you saw that? 0:07:05.276 --> 0:07:07.076 And so I was like, oh gosh, it's picking up 0:07:07.076 --> 0:07:09.596 on patterns in the data, and so the yeah, but 0:07:10.516 --> 0:07:13.876 you continue these interactions and then the next responses go 0:07:14.036 --> 0:07:18.476 on from the therapist, So like within about five or 0:07:18.516 --> 0:07:22.316 so turns, we would often get the model that would 0:07:22.316 --> 0:07:25.836 respond about their interpretations of their problems stemming from their 0:07:25.916 --> 0:07:29.076 their mother or their parents more generally. So like it's 0:07:29.156 --> 0:07:30.596 kind of like, if you were to try to think 0:07:30.636 --> 0:07:33.396 about it, what a psychologist is. This is like every 0:07:33.636 --> 0:07:36.036 trope of what a like in your mind if you 0:07:36.076 --> 0:07:37.236 were going to like think. 0:07:37.116 --> 0:07:40.876 Of the stereotypical on the couch and a guy's wearing 0:07:40.916 --> 0:07:42.836 a tweet jacket sitting in a chair. 0:07:42.756 --> 0:07:47.676 And hardly says anything of that could be potentially helpful, 0:07:47.716 --> 0:07:50.036 but is reflecting things back to me and. 0:07:50.036 --> 0:07:52.636 So then telling me it goes back to my parents. Yeah, 0:07:52.756 --> 0:07:54.876 well this is so let's just pause here for a moment, because, 0:07:54.876 --> 0:07:59.316 as you say, this is like the stereotype of the therapist, 0:07:59.716 --> 0:08:02.916 but you trained it on real data, so maybe it's 0:08:02.916 --> 0:08:04.316 the stereotype for a reason. 0:08:04.436 --> 0:08:08.956 Yes, I think what to me was really clear was 0:08:09.396 --> 0:08:13.796 that we were we had data that the models were 0:08:13.876 --> 0:08:16.676 emulating patterns they were seeing in the data. So the 0:08:16.716 --> 0:08:20.036 models weren't the problem. The problem was the data were 0:08:20.036 --> 0:08:21.116 the wrong data. 0:08:21.276 --> 0:08:23.436 But the data is the data that is used to 0:08:23.476 --> 0:08:26.116 trade real therapists. Like, it's confusing that this is the 0:08:26.156 --> 0:08:26.716 wrong data. 0:08:26.876 --> 0:08:28.276 It is it is. 0:08:28.476 --> 0:08:31.156 Why why is it the wrong data? This should be 0:08:31.196 --> 0:08:32.356 exactly the data you want. 0:08:32.436 --> 0:08:34.836 Well, it's it's the wrong data for this format. And 0:08:34.916 --> 0:08:40.116 our conversation when you might say something me nodding along 0:08:40.356 --> 0:08:43.636 or saying mm hmm or go on, my contextually be 0:08:43.916 --> 0:08:48.916 like completely appropriate. So tactically in a conversational dialogue that 0:08:48.956 --> 0:08:52.356 would happen via chat. This is not like a medium 0:08:52.356 --> 0:08:54.676 that works very well like this kind of thing. 0:08:54.836 --> 0:08:58.516 Yeah, it's almost like a translation, right. It doesn't translate 0:08:58.556 --> 0:09:00.996 from a human face to face interaction to a chat 0:09:00.996 --> 0:09:02.236 window on the computer. 0:09:01.996 --> 0:09:03.236 And not the right setting. 0:09:03.396 --> 0:09:05.116 Yeah, so that I mean that goes to the like 0:09:05.396 --> 0:09:10.236 nonverbal subtler aspects of therapy, right, like presumer when the 0:09:10.236 --> 0:09:14.596 therapist is saying m hm, there is there is body language, 0:09:14.636 --> 0:09:17.756 there's everything that's happening in the room, which is a 0:09:17.796 --> 0:09:21.356 tremendous amount of information or emotional information, right, and that 0:09:21.476 --> 0:09:24.236 is a thing that is lost, yes, no doubt in 0:09:24.316 --> 0:09:28.356 this medium and and maybe speaks to a broader question 0:09:28.636 --> 0:09:31.156 about the translatability of therapy. 0:09:31.316 --> 0:09:35.556 Yeah, absolutely so I think to me, like the it 0:09:35.636 --> 0:09:39.316 was at that moment that I kind of knew that 0:09:39.356 --> 0:09:42.436 we we needed to do something radically different. Neither of 0:09:42.476 --> 0:09:45.956 these was working well. About one in ten of the 0:09:46.076 --> 0:09:51.316 responses from that from that chatbot, based on the clinicians, 0:09:51.556 --> 0:09:53.916 would be something that we would be happy with. So 0:09:54.116 --> 0:09:57.756 something that is both personalized, clinically appropriate and dynamic. 0:09:58.356 --> 0:10:00.796 So you're saying you've got it right ten percent of exactly. 0:10:00.916 --> 0:10:04.116 So really, no, that's that's not a good like no, 0:10:04.156 --> 0:10:06.076 it's not a good not a good therapy. No, we 0:10:06.076 --> 0:10:09.036 would we would never think about implement like actually trying 0:10:09.196 --> 0:10:13.556 employ that. So then what we started at that point 0:10:13.876 --> 0:10:17.916 was building our own creating our own data set from scratch, 0:10:18.316 --> 0:10:22.636 in which we how how the models would learn would 0:10:22.636 --> 0:10:25.876 be exactly what we want it to say. 0:10:25.916 --> 0:10:28.716 That seems that seems wild, I mean, how do you 0:10:28.796 --> 0:10:30.596 do that? How do you how do you generate that 0:10:30.676 --> 0:10:31.236 much data. 0:10:31.356 --> 0:10:33.436 We've had a team of one hundred people that have 0:10:33.516 --> 0:10:35.716 worked on this project over the last five and a 0:10:35.756 --> 0:10:38.796 half years at this point, and they've spent over one 0:10:38.836 --> 0:10:41.916 hundred thousand human hours kind of really trying to build this. 0:10:42.116 --> 0:10:44.596 Just specifically, what how do you build a data set 0:10:44.636 --> 0:10:47.876 from scratch? Because like, the data set is the huge problem. 0:10:48.076 --> 0:10:54.236 Yes, absolutely so. Psychotherapy, when you would test it is 0:10:54.316 --> 0:10:57.316 based on something that is written down in a manual. 0:10:57.756 --> 0:11:01.076 So when you're a psychologist when they're in a randomized 0:11:01.076 --> 0:11:03.636 controlled trial trying to test whether something works or not. 0:11:04.596 --> 0:11:07.356 To be able to test it, it has to be replicable, 0:11:07.516 --> 0:11:11.356 meaning it's like repeated across different therapists. So there are 0:11:11.436 --> 0:11:14.836 manuals that are developed. In this session you work on 0:11:14.836 --> 0:11:17.556 on psycho education. On this session we're going to be 0:11:17.556 --> 0:11:21.596 working on behavioral activation, So which are different techniques that 0:11:21.636 --> 0:11:24.876 are really a focus at a given time, and these 0:11:24.916 --> 0:11:27.676 are broken down to try to make it translational so 0:11:27.676 --> 0:11:30.116 that you can actually move it. So the team would 0:11:30.196 --> 0:11:34.236 read these empirically supported treatment manuals, so the ones that 0:11:34.316 --> 0:11:37.876 had been tested in randomized control trials, and then what 0:11:37.916 --> 0:11:41.276 we would do is we would take that content chapter 0:11:41.316 --> 0:11:44.476 by chapter, because this is like session by session, take 0:11:44.516 --> 0:11:47.116 the techniques that would work well via chat, of which 0:11:47.236 --> 0:11:51.436 most things in cognitive behavioral therapy would, and then we 0:11:51.476 --> 0:11:57.156 would create an artificial dialogue between would act as like 0:11:57.236 --> 0:12:00.116 what is the patient's presenting problem, what they're bringing on, 0:12:00.196 --> 0:12:03.436 what the personality is like, and we're kind of constructing this, 0:12:03.636 --> 0:12:07.596 and then what is what we would want our system 0:12:07.636 --> 0:12:10.756 to be the gold standard response for every kind of 0:12:10.796 --> 0:12:14.396 input and output that we'd have. So we're writing both 0:12:14.476 --> 0:12:17.076 the patient end and the therapist end. 0:12:17.796 --> 0:12:19.716 It's like you're writing a screenplay. 0:12:19.436 --> 0:12:21.916 Basically, it really is. It's a lot like that, but 0:12:22.036 --> 0:12:25.996 instead of a screenplay that might be written like in general, 0:12:26.076 --> 0:12:29.036 it's like not like not just something general, but like, 0:12:29.196 --> 0:12:32.596 where is something that's really evidence based based on content 0:12:32.676 --> 0:12:35.956 that we know works in this setting? 0:12:36.036 --> 0:12:39.516 And so what you write the equivalent of what thousands 0:12:39.556 --> 0:12:40.956 of hours of sessions? 0:12:41.036 --> 0:12:44.596 Hundreds of thousands. There was post docs, grad students and 0:12:44.716 --> 0:12:47.556 undergraduates within my group that we're all part of this 0:12:47.636 --> 0:12:49.196 team that are creating. 0:12:49.156 --> 0:12:51.156 Just doing the work, just writing the dialogue. 0:12:51.236 --> 0:12:54.436 Yeah, exactly. And not only did we write them, but 0:12:54.916 --> 0:12:58.436 every dialogue before it would go into something that our 0:12:58.436 --> 0:13:01.796 models are trained would be reviewed by another member of 0:13:01.836 --> 0:13:06.836 the team. So it's all not only crafted by hand, 0:13:06.916 --> 0:13:09.716 but we would review it, give each other feed on it, 0:13:09.796 --> 0:13:12.236 and then like make sure that it is the highest 0:13:12.316 --> 0:13:17.996 quality data. And that's when we started seeing dramatic improvements 0:13:18.036 --> 0:13:22.036 in the model performance. So we continued with us for years. 0:13:23.196 --> 0:13:29.476 Six months before chat TPT was launched, we had a 0:13:29.596 --> 0:13:33.916 model that in today's standards would be so tiny, that 0:13:34.116 --> 0:13:38.516 was delivering about ninety percent of the responses that were 0:13:38.556 --> 0:13:42.276 output We were evaluating as exactly what we'd want. It's 0:13:42.316 --> 0:13:47.276 this gold standard evidence based treatment, so that was fantastic. 0:13:47.396 --> 0:13:49.916 We were really excited about it. So we've got like 0:13:49.956 --> 0:13:53.556 the we've got the benefit side down of the equation. 0:13:54.316 --> 0:13:58.196 The next two years we focus on the risk, the 0:13:58.316 --> 0:14:00.516 risk side of it well, because there's a huge risk. 0:14:00.356 --> 0:14:02.676 Here, right The people who are using it are by 0:14:02.756 --> 0:14:07.596 design quite vulnerable, by design putting a tremendous amount of 0:14:07.716 --> 0:14:11.996 trust into this bot and making themselves vulnerable to it, 0:14:12.276 --> 0:14:16.196 like it's a it's quite a risky proposition. And so 0:14:16.196 --> 0:14:18.116 so tell me specifically, what are you doing. 0:14:18.236 --> 0:14:21.516 So we're trying to get it to endorse elements that 0:14:21.556 --> 0:14:24.796 would make mental health worse. So a lot of our 0:14:24.876 --> 0:14:29.436 conversations or surrounding trying to get it to For example, 0:14:29.836 --> 0:14:33.436 I'll give you an example of one that nearly almost 0:14:33.836 --> 0:14:37.356 almost any model will struggle with that's not tailored towards 0:14:37.396 --> 0:14:40.676 the safety end. What is it is if you tell 0:14:40.756 --> 0:14:43.076 a model that you want to lose weight, it will 0:14:43.236 --> 0:14:46.196 generally try to help you do that. And if you 0:14:46.276 --> 0:14:48.076 want to if you want to work in an area 0:14:48.116 --> 0:14:52.236 related to mental health, trying to promote weight loss without 0:14:52.276 --> 0:14:54.076 context is so not safe. 0:14:54.476 --> 0:14:56.636 You saying it might be a user within eating disorder 0:14:56.676 --> 0:15:00.196 absolutely unhealthily thin, who wants to be even thinner. 0:14:59.836 --> 0:15:03.116 And the model will help them to often actually get 0:15:03.156 --> 0:15:06.356 into a lower weight than they already are. So this 0:15:06.516 --> 0:15:09.716 is like not something that we would ever want to promote, 0:15:09.796 --> 0:15:12.956 but this is something that we certainly at earlier stages 0:15:13.036 --> 0:15:15.716 we're seeing these types of characteristics within the model. 0:15:15.956 --> 0:15:18.236 What are other like, that's an interesting one and it 0:15:18.316 --> 0:15:20.116 makes perfect sense when you say it, I would not 0:15:20.196 --> 0:15:22.036 have thought of it. Or what's another one? 0:15:22.196 --> 0:15:24.396 A lot of it would be like we talk about 0:15:24.436 --> 0:15:27.556 the ethics of suicide. For example, somebody who is who thinks, 0:15:27.636 --> 0:15:30.796 you know, they're in a midst of suffering and you 0:15:30.836 --> 0:15:33.156 know it's like that they could should be able to 0:15:33.236 --> 0:15:35.036 end of their life or they're thinking about this. 0:15:35.356 --> 0:15:39.316 Yes, and what do you want the model? What what? 0:15:39.316 --> 0:15:41.196 What does the model say that it shouldn't say in 0:15:41.236 --> 0:15:41.716 that setting? 0:15:41.796 --> 0:15:44.476 So for you, in these settings, we want to make 0:15:44.516 --> 0:15:47.636 sure that they don't and the model does not promote 0:15:47.716 --> 0:15:52.316 or endorse elements that would promote someone's a worsening of 0:15:52.356 --> 0:15:55.636 suicidal intent. We want to make sure we're providing not 0:15:55.676 --> 0:15:59.116 only not the absence of that, actually some benefit in 0:15:59.156 --> 0:16:00.276 these types of scenarios. 0:16:00.556 --> 0:16:03.116 That's the ultimate nightmare for you. Yeah, right, Like this 0:16:03.316 --> 0:16:06.276 be super clear. The very worst thing that could happen 0:16:06.516 --> 0:16:09.796 is you build this thing and it contributes to someone. Absolutely, 0:16:09.796 --> 0:16:13.316 that's a plausible outcome and a disastrous night. 0:16:13.436 --> 0:16:16.436 It's everything that I worry about in this area is 0:16:16.556 --> 0:16:19.796 exactly this kind of thing. And so we essentially every 0:16:19.836 --> 0:16:23.276 time we find an area where they're not implementing things perfectly, 0:16:24.156 --> 0:16:28.756 some optimal response, adding new training data, and that's that's 0:16:28.796 --> 0:16:31.316 when things continue to get better until we do this 0:16:31.436 --> 0:16:34.116 and we don't find these holes anymore. That's when we 0:16:34.196 --> 0:16:37.276 finally we're ready for the randomized control trial. 0:16:38.076 --> 0:16:43.476 Right, So you decide after after what four years five years? 0:16:43.516 --> 0:16:46.596 This is about four and a half years. 0:16:46.916 --> 0:16:50.196 Yeah, that that you're ready to to have people. 0:16:50.036 --> 0:16:51.476 Use use the model. 0:16:51.676 --> 0:16:54.276 I'll be it in a kind of Yeah, you're going 0:16:54.356 --> 0:16:56.476 to be the human in the loop. Right, So, so 0:16:56.596 --> 0:16:59.316 you decide to do this study. You recruit people on 0:16:59.756 --> 0:17:01.276 Facebook and Instagram. 0:17:00.876 --> 0:17:04.476 Basically ye exactly, yep, and what. 0:17:04.476 --> 0:17:06.236 So what are they signing up for? What's the what's 0:17:06.276 --> 0:17:07.116 the big study you do? 0:17:07.236 --> 0:17:10.236 So it's a it's a it's a randomized control trial. 0:17:11.556 --> 0:17:15.036 The trial design is essentially that folks would come in, 0:17:15.476 --> 0:17:19.116 they would fill out information about their mental health across 0:17:19.116 --> 0:17:24.236 a variety of areas, so depression, anxiety, and eating disorders, 0:17:24.276 --> 0:17:28.876 for folks that screen positive for having clinical levels of 0:17:28.916 --> 0:17:32.916 depression or anxiety, they would be in our Folks that 0:17:32.956 --> 0:17:35.596 were at risk for eating disorders would be included in 0:17:35.636 --> 0:17:39.276 the trial. We tried to have at least seventy people 0:17:39.316 --> 0:17:41.636 in each group, so we had two hundred and ten 0:17:41.676 --> 0:17:45.036 people that we were planning on and rolling within the trial, 0:17:45.556 --> 0:17:48.756 and then half of them were randomized to receive their 0:17:48.756 --> 0:17:50.996 ABOUT and half of them were on a wait list 0:17:51.116 --> 0:17:53.756 in which they would receive their ABOUT after the trial 0:17:53.796 --> 0:17:57.156 had ended. The trial design was to try to ask 0:17:57.196 --> 0:18:00.316 folks to use their ABOUT for four weeks. They retained 0:18:00.396 --> 0:18:02.916 access to therabot and could use their ABOUT for the 0:18:02.956 --> 0:18:06.716 next four weeks thereafter, so eight weeks total, but we 0:18:06.756 --> 0:18:08.996 asked them to try to actually use it during that 0:18:09.196 --> 0:18:12.796 first four weeks and that was that was essentially the 0:18:12.836 --> 0:18:13.436 trial design. 0:18:13.596 --> 0:18:17.436 So okay, so people signed up, they start like, what's 0:18:17.556 --> 0:18:20.636 what's actually happening? Are they just like chatting with the 0:18:20.676 --> 0:18:22.036 bought every day? Is it? 0:18:22.076 --> 0:18:26.196 So they install a smartphone application that's that they're about at. 0:18:27.156 --> 0:18:31.036 They are prompted once a day to try to have 0:18:31.076 --> 0:18:33.956 a conversation starter with the with the bot and then 0:18:34.196 --> 0:18:37.436 the bot. From there they could talk about it when 0:18:37.596 --> 0:18:40.276 and wherever they would want. They can ignore those notifications 0:18:40.276 --> 0:18:42.636 and kind of engage with it at any time that 0:18:42.636 --> 0:18:47.076 they'd want. But that was the gist of the trial design, 0:18:47.156 --> 0:18:49.556 and so folks in terms of how people used it, 0:18:50.036 --> 0:18:53.196 they interacted with it throughout the day, throughout the night. 0:18:54.196 --> 0:18:58.676 So for example, folks that would have trouble sleeping, that 0:18:58.796 --> 0:19:00.716 was like a way that folks during the middle of 0:19:00.756 --> 0:19:05.516 the night would engage with it fairly often. They in 0:19:05.596 --> 0:19:09.036 terms of the types of the topics that they described, 0:19:10.476 --> 0:19:13.116 it was really the entire range of something that you 0:19:13.116 --> 0:19:15.796 would see in psychotherapy. We had folks that were dealing 0:19:15.836 --> 0:19:18.996 with and discussing their different symptoms that they were talking about. 0:19:19.036 --> 0:19:21.236 So the depression, their anxiety that they were struggling with, 0:19:21.276 --> 0:19:24.796 their their eating, and their body image concerns. Those types 0:19:24.796 --> 0:19:27.076 of things are common because of the groups that we 0:19:27.076 --> 0:19:33.156 were recruiting. But relationship difficulties, problems like folks, some folks 0:19:33.156 --> 0:19:37.596 were really like I had ruptures in there, you know, 0:19:37.796 --> 0:19:40.076 somebody was going through a divorce. Other folks were like 0:19:40.196 --> 0:19:44.516 going through breakups, problems at work. Some folks were unemployed 0:19:45.556 --> 0:19:48.676 and during this time, So like the range of kind 0:19:48.716 --> 0:19:52.396 of personal dilemmas and difficulties that folks were experiencing. Was 0:19:52.476 --> 0:19:54.276 a lot of what we would see in like a 0:19:54.356 --> 0:19:58.316 real setting where it's like kind of a whole host 0:19:58.436 --> 0:20:02.156 of different things that folks were describing and experiencing. 0:20:02.076 --> 0:20:04.676 And presumably had they agreed as part of enrolling in 0:20:04.756 --> 0:20:07.796 the trial to let you read the transcript Oh? 0:20:07.836 --> 0:20:11.236 Absolutely, yeah, very very clear when we did an informed 0:20:11.236 --> 0:20:14.916 consent process where folks would know that we were reading 0:20:15.396 --> 0:20:16.516 reading these transcripts. 0:20:16.916 --> 0:20:19.236 And are you personally, like, what was it like for 0:20:19.276 --> 0:20:21.596 you seeing them come in? Are you reading them every day? 0:20:21.676 --> 0:20:22.556 I mean more than that. 0:20:23.156 --> 0:20:27.156 So, I mean this is something that is so ill. 0:20:27.396 --> 0:20:29.596 You alluded to that that this is one of these 0:20:29.636 --> 0:20:33.036 concerns that anybody would have. Is like a nightmare scenario 0:20:33.516 --> 0:20:37.196 where something is the bad happens and somebody actually outs right, 0:20:37.436 --> 0:20:39.516 So this is like I think of this in a 0:20:39.556 --> 0:20:40.756 way that I take So this is not. 0:20:40.756 --> 0:20:43.916 A happy moment for you. This is like you're terrified 0:20:43.916 --> 0:20:44.916 that it might go wrong. 0:20:45.076 --> 0:20:48.156 Well, it's it's certainly like I see it going right, 0:20:48.236 --> 0:20:50.556 but I have every concern that it could go wrong. 0:20:50.796 --> 0:20:54.716 Right like that, And so for the first half of 0:20:54.756 --> 0:21:00.636 the trial, I am monitoring every single interaction sent to 0:21:00.796 --> 0:21:03.436 or from the bot. Other people are also doing this 0:21:03.516 --> 0:21:05.836 on the team, so I'm not the only one. But 0:21:05.916 --> 0:21:07.396 I did not get a lot of sleep in the 0:21:07.396 --> 0:21:10.156 first half of this trial, in part because I was 0:21:10.196 --> 0:21:12.156 really trying to do this in near real time. So 0:21:12.316 --> 0:21:15.276 usually for nearly every message I was, I was getting 0:21:15.396 --> 0:21:18.356 to it within about an hour. So yeah, it was 0:21:18.676 --> 0:21:21.956 it was a barrage of NonStop kind of communication that 0:21:22.036 --> 0:21:22.476 was happening. 0:21:22.556 --> 0:21:24.996 So were there were there any slip ups? Did you 0:21:24.996 --> 0:21:27.436 ever have to intervene as a human in the loop. 0:21:27.316 --> 0:21:30.356 That we did? And the thing that that was something 0:21:30.396 --> 0:21:34.236 that we as a team did not anticipate. What we 0:21:34.356 --> 0:21:37.796 found was really unintended behavior was a lot of folks 0:21:38.156 --> 0:21:42.436 interacted with they're abot, and in doing that, there was 0:21:42.476 --> 0:21:46.116 a significant number of people that would interact with it 0:21:46.156 --> 0:21:48.876 and talk about their medical symptoms. So, for example, there 0:21:48.916 --> 0:21:51.356