1 00:00:11,280 --> 00:00:13,360 Hey, Sam. Hey, Lane. 2 00:00:14,640 --> 00:00:19,120 So I've been thinking about something and I want to get your 3 00:00:19,120 --> 00:00:21,120 take on it. And I think we have an 4 00:00:21,120 --> 00:00:24,920 interesting situation that's emerging in this landscape of AI 5 00:00:24,920 --> 00:00:27,760 and mental health, which of course, is one of the things 6 00:00:27,760 --> 00:00:31,680 that we think about a lot. On the one hand, we have therapy 7 00:00:31,680 --> 00:00:35,320 and companionship, like really rising to the top as the very 8 00:00:35,320 --> 00:00:37,920 popular use cases. These are common use cases for 9 00:00:37,960 --> 00:00:41,000 people using generative AI. There's evidence that this can 10 00:00:41,000 --> 00:00:45,120 sometimes be helpful, especially in cases where that chatbot is 11 00:00:45,120 --> 00:00:49,160 clinically validated, say like this Thera bot from Dartmouth 12 00:00:49,200 --> 00:00:50,600 that we'll talk about in this episode. 13 00:00:50,960 --> 00:00:53,920 But on the other hand, we see from research like this study 14 00:00:53,920 --> 00:00:57,800 that pretty recently came out from Open AI and MIT that 15 00:00:57,800 --> 00:01:01,160 spending more time with LLMS actually has a detrimental 16 00:01:01,160 --> 00:01:04,560 effect on well-being. So this study showed that people 17 00:01:04,560 --> 00:01:09,840 who spent more time really like heavy users of Chad GBT had an 18 00:01:09,840 --> 00:01:14,080 increase in loneliness and sort of substituted for these real 19 00:01:14,080 --> 00:01:17,280 world interactions that of course are very important for, 20 00:01:18,240 --> 00:01:20,800 you know, being alive and feeling like a human and so on. 21 00:01:21,720 --> 00:01:25,600 So my question for you is, how do we reconcile these findings 22 00:01:25,600 --> 00:01:28,160 that seem to be at odds with one another? 23 00:01:29,440 --> 00:01:34,560 I think it's a very interesting state of things where there's 24 00:01:34,840 --> 00:01:39,880 clear evidence, again, that using AI can be extremely useful 25 00:01:39,880 --> 00:01:43,400 for many things, but again, that it can also, through overuse or 26 00:01:43,400 --> 00:01:48,240 through extreme use, have some form of black fire effects. 27 00:01:48,280 --> 00:01:54,240 And as a Swedish person, I can't help but think about a very 28 00:01:54,240 --> 00:01:56,600 historically significant thing that happened in Sweden. 29 00:01:57,120 --> 00:01:58,920 Can I give you a quick history lesson? 30 00:01:59,080 --> 00:02:03,400 Yes, please do. In the mid 20th century, an idea 31 00:02:03,400 --> 00:02:09,960 was introduced in Sweden. Basically the idea was that how 32 00:02:10,840 --> 00:02:15,480 can you truly love someone let's say a partner if you're 33 00:02:15,480 --> 00:02:17,760 depending on them for their income. 34 00:02:17,760 --> 00:02:22,880 So if you're a woman in in 1950s and most women at the time were 35 00:02:22,880 --> 00:02:26,480 depending on their counterparts salary to survive, like is that 36 00:02:26,480 --> 00:02:27,520 love? Is that true love? 37 00:02:28,640 --> 00:02:32,320 It's really just utilitarian. It's one way love. 38 00:02:33,120 --> 00:02:35,360 Exactly. Well, the Swedish theory of love 39 00:02:35,360 --> 00:02:38,600 was that basically like what the government should be, It should 40 00:02:38,600 --> 00:02:42,680 be what's called folk Hemet, basically the home of the people 41 00:02:43,000 --> 00:02:48,560 that you should only depend on the government, so you don't 42 00:02:48,560 --> 00:02:51,800 have to depend on a partner. Even as a child, you shouldn't 43 00:02:51,800 --> 00:02:53,960 even have to depend on your parents, because if they're bad 44 00:02:53,960 --> 00:02:56,120 parents, you should be rescued from them. 45 00:02:56,120 --> 00:02:59,800 So that's the idea. The government should serve as 46 00:02:59,800 --> 00:03:03,080 this home for the people. The kind of backfire effects of 47 00:03:03,080 --> 00:03:05,840 this is that, of course, Sweden has had a really good Social 48 00:03:05,840 --> 00:03:09,000 Security net and like provide a lot of support for people to 49 00:03:09,000 --> 00:03:11,800 live relatively good lives if compared to most countries. 50 00:03:12,120 --> 00:03:16,720 But the flip side of this is that, well, if you don't need to 51 00:03:16,720 --> 00:03:20,080 depend on any one of them, you might find yourself lonely in 52 00:03:20,080 --> 00:03:23,120 your apartment because you don't need to depend on your 53 00:03:23,120 --> 00:03:24,400 neighbors. You don't need to depend on your 54 00:03:24,400 --> 00:03:25,720 partner, you don't need to depend on anyone. 55 00:03:26,240 --> 00:03:30,360 And there's been a lot of this case in Sweden, kind of always 56 00:03:30,360 --> 00:03:33,480 character of Swedish society where, you know, people are very 57 00:03:33,480 --> 00:03:35,840 cold towards each other. They kind of like isolated and 58 00:03:35,840 --> 00:03:38,440 there's a lot of loneliness and a lot of people that are 59 00:03:38,640 --> 00:03:43,560 basically in the worst of cases dying alone and no one noticing 60 00:03:43,560 --> 00:03:46,560 for months because they don't have to depend on anyone. 61 00:03:47,360 --> 00:03:50,840 This one I'm trying to take back to with AI is that like AI is 62 00:03:50,840 --> 00:03:54,440 amazing and so incredible in what is able to provide, but 63 00:03:54,720 --> 00:03:57,920 it's also risking to some very similar things like that you 64 00:03:57,920 --> 00:04:02,360 could become so easily quite dependent on AI for 65 00:04:02,360 --> 00:04:05,800 companionship, for productivity purposes, for all these things. 66 00:04:06,120 --> 00:04:09,840 And you're less incentivized to do things that are harder, like 67 00:04:09,840 --> 00:04:14,480 getting out of your comfort zone or going out of your apartment 68 00:04:14,480 --> 00:04:16,839 to hang out with a friend. Like all of these things take 69 00:04:16,839 --> 00:04:19,200 more effort. I think it's a really 70 00:04:19,200 --> 00:04:22,240 interesting, I did not think of it that way, but certainly you 71 00:04:22,240 --> 00:04:24,320 can have too much of a good thing, right? 72 00:04:24,360 --> 00:04:28,520 The risk of taking that argument too far perhaps is you get into 73 00:04:28,800 --> 00:04:31,720 the American conservative talking points of like, we 74 00:04:31,720 --> 00:04:36,000 shouldn't have a social safety net at all because people will 75 00:04:36,000 --> 00:04:38,400 take advantage of it and so on. And like, that's not the answer 76 00:04:38,400 --> 00:04:40,040 either. And then if you take that, if 77 00:04:40,040 --> 00:04:43,640 you sort of, you know, copy and paste that argument to AI, we're 78 00:04:43,640 --> 00:04:47,040 not saying that we shouldn't have AI because there are some 79 00:04:47,040 --> 00:04:50,360 potential detrimental consequences of overuse. 80 00:04:50,600 --> 00:04:54,000 Maybe we just need to design systems to prevent people from 81 00:04:54,000 --> 00:04:56,440 overusing it. And I think one thing that I've 82 00:04:56,440 --> 00:04:59,640 been thinking about so much recently is this one design 83 00:04:59,640 --> 00:05:04,680 feature of ChatGPT where every single time you get a response, 84 00:05:05,000 --> 00:05:07,360 instead of just giving you the response. 85 00:05:07,680 --> 00:05:09,960 And, you know, if you have another question, you can ask 86 00:05:09,960 --> 00:05:11,720 another question. And instead of just leaving it 87 00:05:11,720 --> 00:05:16,720 at that, every single time ChatGPT offers to do two or 88 00:05:16,720 --> 00:05:18,760 three more things for you. Would you like me to create a 89 00:05:18,760 --> 00:05:20,680 visual for you? Would you like me to, you know, 90 00:05:20,680 --> 00:05:24,720 turn this into a table? And it's just so irresistible to 91 00:05:24,720 --> 00:05:28,240 continue this chain of work forever because, you know, 92 00:05:28,240 --> 00:05:31,240 sometimes I'm like, yeah, sure, make me a table. 93 00:05:31,240 --> 00:05:34,240 Like, that sounds fun. But then you find yourself just 94 00:05:34,240 --> 00:05:37,680 spending so, so much time going down this trail of things when 95 00:05:37,680 --> 00:05:40,720 you only wanted to do, you know, one thing. 96 00:05:40,720 --> 00:05:47,240 And you end up just lured into this constant cycle of creation 97 00:05:47,240 --> 00:05:50,440 where ChatGPT is just making more and more work for itself. 98 00:05:50,880 --> 00:05:52,920 But we don't have to design systems in that way. 99 00:05:53,320 --> 00:05:55,400 Yeah, No, exactly. Yeah. 100 00:05:55,400 --> 00:05:57,680 I think it's not only in terms of work, but also again, we'll 101 00:05:57,680 --> 00:06:01,240 come back to a companionship. It's the same way there where 102 00:06:01,240 --> 00:06:04,880 it's like you can have the most boring conversation in terms of 103 00:06:04,880 --> 00:06:06,680 like saying the most boring things and it's just going to be 104 00:06:06,680 --> 00:06:08,600 like, that's so interesting, what about this? 105 00:06:08,600 --> 00:06:11,200 It's just going to want to continue talking forever 106 00:06:11,240 --> 00:06:14,360 basically. And I think it's the same thing 107 00:06:14,360 --> 00:06:17,360 there. It's just the sign currently is 108 00:06:17,360 --> 00:06:20,520 used for you to engage as much as possible. 109 00:06:21,680 --> 00:06:25,280 And this whole conversation reminds me so much, actually, of 110 00:06:25,280 --> 00:06:28,400 some of the more nuanced research on social media. 111 00:06:28,680 --> 00:06:31,960 And I think often we're like, we take this blanket approach of 112 00:06:32,200 --> 00:06:35,360 social media is bad. It's like ruining the children. 113 00:06:35,360 --> 00:06:37,440 Everything is terrible about social media. 114 00:06:37,440 --> 00:06:39,760 But actually, there's quite a lot of research showing that 115 00:06:40,000 --> 00:06:43,280 there are, you know, many people who benefit from social media 116 00:06:43,280 --> 00:06:46,280 because it gets them to actually augment their real life 117 00:06:46,280 --> 00:06:48,200 relationships. So these people are not spending 118 00:06:48,200 --> 00:06:50,920 too much time on social media and they're just like using it 119 00:06:50,920 --> 00:06:54,680 to connect with people who they like actually see in real life, 120 00:06:54,680 --> 00:06:58,200 not as a substitute, but as a sort of supplement. 121 00:06:58,440 --> 00:07:01,960 And I think that we can kind of think of LLMS in a similar way 122 00:07:01,960 --> 00:07:06,680 where if you're using them to not entirely substitute your 123 00:07:06,680 --> 00:07:11,720 real life, you know, therapist or your friends and so on, but 124 00:07:11,720 --> 00:07:15,840 perhaps you are using it in smaller ways, then there's like 125 00:07:15,840 --> 00:07:19,520 a healthy way to approach it. So I think it's interesting that 126 00:07:19,520 --> 00:07:25,480 we're currently finding ourself in this, call it sycophantic era 127 00:07:25,480 --> 00:07:30,160 of AI where, you know, AI is super keen to talk forever, 128 00:07:30,160 --> 00:07:35,520 always encouraging, always like if not prompt, otherwise blindly 129 00:07:35,840 --> 00:07:38,960 positive. And I think there will be kind 130 00:07:38,960 --> 00:07:41,880 of an interesting transition at some point where if you truly 131 00:07:41,880 --> 00:07:48,880 want AI to add the value of a human, we don't want it to be 132 00:07:48,880 --> 00:07:50,320 that way. We want it to be a little 133 00:07:50,320 --> 00:07:53,840 different. Like if we had AAI therapist for 134 00:07:53,840 --> 00:07:57,600 example, we would probably need it to tell us some hard truths 135 00:07:57,680 --> 00:08:02,360 and to not, you know, just bake everything into some form of 136 00:08:02,680 --> 00:08:07,400 empty platitudes and sharing. Correcting negative behavior. 137 00:08:07,400 --> 00:08:10,560 If you say I want to do something that is, you know, 138 00:08:11,040 --> 00:08:14,360 socially unacceptable or threatening to other people's 139 00:08:14,360 --> 00:08:18,200 lives or your own life, then yeah, your AI therapist should 140 00:08:18,200 --> 00:08:21,280 tell you that's not a good idea. Yeah, right. 141 00:08:21,360 --> 00:08:22,800 Shouldn't say, yeah, I understand. 142 00:08:24,760 --> 00:08:28,120 Go for it. Yeah, and just be some form of 143 00:08:28,120 --> 00:08:31,720 counterbalance. We want AI to help people form a 144 00:08:31,720 --> 00:08:34,919 better understanding of the world and how they can live a 145 00:08:34,919 --> 00:08:37,280 good life. And currently it's not really 146 00:08:37,280 --> 00:08:39,840 doing that very well. Whatever people bring to the AI 147 00:08:40,080 --> 00:08:43,559 tends to often times to make worse somewhere if someone comes 148 00:08:43,559 --> 00:08:47,840 with AI and have some really wild conspiracy theories at this 149 00:08:47,840 --> 00:08:51,560 point unfortunately, with the exception of using Gordon 150 00:08:51,560 --> 00:08:53,080 Pennycook. 'S the debunk pot. 151 00:08:53,160 --> 00:08:56,280 Yeah, except for the debunk pot, they will find themselves like 152 00:08:56,280 --> 00:08:58,960 probably deeper down that conspiratorial route. 153 00:08:58,960 --> 00:09:02,920 But yeah, I think this is always leads us quite interestingly 154 00:09:02,920 --> 00:09:06,480 into our episode of Today with Allison Cereso. 155 00:09:07,000 --> 00:09:09,760 So Allison is a clinical researcher, psychologist, 156 00:09:09,760 --> 00:09:12,920 professor, and senior vice president of research at 157 00:09:13,040 --> 00:09:17,080 Empathic, overseeing AI tools that supports therapists so that 158 00:09:17,080 --> 00:09:20,240 it can deliver more accurate empathetic care to their 159 00:09:20,240 --> 00:09:23,360 patients. So this really strikes this 160 00:09:23,360 --> 00:09:26,560 interesting balance between understanding when and how to 161 00:09:26,560 --> 00:09:32,520 use AI athletics is today. And what Empathic aims to do is 162 00:09:32,520 --> 00:09:36,320 provide AI tools that gives feedback and suggested text to 163 00:09:36,320 --> 00:09:40,240 therapists to help them be more empathetic to their patients. 164 00:09:40,240 --> 00:09:45,120 So basically maybe again as we talked about like not looking at 165 00:09:45,120 --> 00:09:49,240 the extremes of whether or not using AI at all or using AI 166 00:09:49,240 --> 00:09:52,600 completely, but Morris is a form of middle ground of seeing AI 167 00:09:52,600 --> 00:09:58,080 maybe as a Co therapist. And really interesting given our 168 00:09:58,160 --> 00:10:01,800 episode with Mickey Insleck where we talk about how a 169 00:10:01,800 --> 00:10:04,400 generative AI actually does such an incredible job. 170 00:10:04,400 --> 00:10:07,600 At least, you know, simulating empathy. 171 00:10:07,760 --> 00:10:11,960 Then actually you can use it as a tool to then coach or teach 172 00:10:12,320 --> 00:10:16,920 therapists, be more empathetic. And in our episode with Allison, 173 00:10:16,920 --> 00:10:20,520 we talk all about this, how empathic works as this sort of 174 00:10:20,520 --> 00:10:25,360 Co therapist and the role of AI versus humans in therapy and 175 00:10:25,360 --> 00:10:28,600 where they can really kind of help each other, balance each 176 00:10:28,600 --> 00:10:31,880 other out, where it's important to have a human in the loop. 177 00:10:32,480 --> 00:10:35,640 And finally, how we can use these tools to even mitigate 178 00:10:35,640 --> 00:10:39,520 bias in therapy. So really exciting discussion 179 00:10:39,600 --> 00:10:55,920 and we hope you love it. So I'm very happy to say welcome 180 00:10:55,960 --> 00:10:57,760 Allison to the Bible Design Podcast. 181 00:10:58,520 --> 00:11:00,920 Thank you so much for having me. It's really great to be here 182 00:11:00,920 --> 00:11:03,280 with you both. Yeah. 183 00:11:03,280 --> 00:11:07,960 And we're excited to get into so much of the interesting stuff we 184 00:11:07,960 --> 00:11:14,600 hope to cover with you because you are, I don't say really in 185 00:11:14,600 --> 00:11:17,720 this kind of fascinating space of trying to see what can be 186 00:11:17,720 --> 00:11:22,800 done on this intersection of thoughtful science based design, 187 00:11:22,840 --> 00:11:26,200 combining that with AI and how to leverage AI in good ways. 188 00:11:26,200 --> 00:11:30,600 So tell us about Imperfect and what do you do? 189 00:11:31,560 --> 00:11:34,400 Sure. So Empathic is a conversation 190 00:11:34,400 --> 00:11:38,800 intelligence company, but we think of ourselves as expanding 191 00:11:39,080 --> 00:11:43,680 human understanding to foster clinical precision and to also 192 00:11:43,680 --> 00:11:47,520 foster precision medicine. Our tool is used in a number of 193 00:11:47,520 --> 00:11:51,760 places, but the majority at this time is in the clinical trial 194 00:11:51,760 --> 00:11:53,720 space. So we're often times in the 195 00:11:53,720 --> 00:11:58,160 background helping to make sure that clinicians are sticking to 196 00:11:58,480 --> 00:12:01,680 fidelity of a particular protocol of how they're supposed 197 00:12:01,680 --> 00:12:06,000 to run a trial. But we also are working in 198 00:12:06,040 --> 00:12:11,680 clinical outcome assessments. And so again, providing AI 100% 199 00:12:11,720 --> 00:12:15,160 quality support to ensure that clinicians are able to do their 200 00:12:15,160 --> 00:12:20,480 work in a standardized way and to also have more precision in 201 00:12:20,480 --> 00:12:22,440 the way that they carry out their job. 202 00:12:23,080 --> 00:12:25,440 So I think often times you can think of us as like a Co 203 00:12:25,440 --> 00:12:29,600 supervisor or sometimes people use the word copilot. 204 00:12:30,360 --> 00:12:34,320 But for us, it's the way that we can really like embed or 205 00:12:34,320 --> 00:12:38,160 integrate AI into clinical practice in areas where humans 206 00:12:38,160 --> 00:12:41,240 do need support. I'll just give the example of 207 00:12:41,240 --> 00:12:44,720 like, you know, in a clinical trial space it is, it can be 208 00:12:44,720 --> 00:12:48,160 humanly impossible to review 100% of what happens in every 209 00:12:48,160 --> 00:12:51,360 single session because humans get exhausted. 210 00:12:51,360 --> 00:12:54,320 But also it's just very difficult, you know, time and 211 00:12:54,320 --> 00:12:56,760 resources, whereas an AI can do it. 212 00:12:57,200 --> 00:12:59,880 And so it's a really great place where an AI can provide more of 213 00:12:59,880 --> 00:13:03,480 that coverage and not just do the coverage, but also allow 214 00:13:03,720 --> 00:13:06,520 clinicians or the humans to really focus on different kinds 215 00:13:06,520 --> 00:13:10,040 of parts of the job. So to really maybe focus when an 216 00:13:10,080 --> 00:13:12,920 AI might be able to detect clinical risk than to go in and 217 00:13:12,920 --> 00:13:17,120 uncover what's going on there. You've referred to it as a Co 218 00:13:17,120 --> 00:13:21,240 therapist at times as well. Say that you're a clinician 219 00:13:21,240 --> 00:13:23,960 who's engaging with the tool. How do you sort of interface 220 00:13:23,960 --> 00:13:26,320 with it? What is your experience as a 221 00:13:26,320 --> 00:13:28,520 user? Yeah. 222 00:13:28,600 --> 00:13:31,800 You know, for us, often times a user like it, that can look 223 00:13:31,800 --> 00:13:33,720 different, right? But it could be that it's 224 00:13:34,120 --> 00:13:37,440 somebody who's overseeing a particular site for a clinical 225 00:13:37,440 --> 00:13:42,720 trial or somebody who is overseeing that the assessments 226 00:13:42,760 --> 00:13:45,240 that people are doing in a clinical trial have been done in 227 00:13:45,240 --> 00:13:48,800 a standardized way. So often times, like I think, 228 00:13:48,800 --> 00:13:53,200 our tool is used by the supervisor who can then be 229 00:13:53,200 --> 00:13:55,000 flagged when more support is needed. 230 00:13:55,440 --> 00:13:58,760 The best example I can give is, so I am a former professor. 231 00:13:59,320 --> 00:14:01,720 I worked at UC Santa Barbara for many years. 232 00:14:01,800 --> 00:14:06,080 I had my lab there, but also founded a trauma clinic. 233 00:14:06,960 --> 00:14:10,960 And for a period of time, it was just me and, you know, several 234 00:14:10,960 --> 00:14:13,040 doctoral students who were learning to be clinicians. 235 00:14:13,680 --> 00:14:16,920 And it was impossible for me to do 100% coverage of what was 236 00:14:16,920 --> 00:14:19,400 happening. It was a trauma clinic that was 237 00:14:19,400 --> 00:14:22,760 founded in 2020 during the Black Lives Matter movement, when it 238 00:14:22,760 --> 00:14:24,600 was really getting national recognition. 239 00:14:25,000 --> 00:14:29,080 And so we wanted to be able to be a safe place for Black 240 00:14:29,080 --> 00:14:31,600 residents in the Central Coast, to be able to connect with 241 00:14:31,600 --> 00:14:35,000 therapist and just be able to talk about trauma, but to also 242 00:14:35,000 --> 00:14:38,480 really think of trauma in a more comprehensive way where it isn't 243 00:14:38,480 --> 00:14:43,280 just, you know, exposure to maybe a traumatic event like at 244 00:14:43,280 --> 00:14:45,320 war or something like that or a car accident. 245 00:14:45,560 --> 00:14:48,120 But it was also just the ways that people had to navigate, you 246 00:14:48,120 --> 00:14:50,800 know, living in a world where there was a lot of police 247 00:14:50,800 --> 00:14:53,840 violence at that time and sort of just beating community, 248 00:14:53,840 --> 00:14:55,760 right? And especially for folks who 249 00:14:55,760 --> 00:14:58,000 were being triggered on a regular basis because so much 250 00:14:58,000 --> 00:15:01,840 media exposure to trauma or having to re see, right, like 251 00:15:01,840 --> 00:15:06,400 violent events happen. And so at that time, you know, I 252 00:15:06,400 --> 00:15:09,960 was training this group of doctoral students to do better 253 00:15:09,960 --> 00:15:13,880 assessments around understanding trauma, understanding PTSD, and 254 00:15:13,880 --> 00:15:16,640 at the same time making sure that they also had the basics 255 00:15:16,640 --> 00:15:19,000 down of like, how do you connect with a client right in your 256 00:15:19,000 --> 00:15:20,880 first session? How do you do a comprehensive 257 00:15:20,880 --> 00:15:24,680 assessment intake and just provide support over time? 258 00:15:25,080 --> 00:15:29,720 And so it was impossible to like, you know, sort of do my 259 00:15:29,720 --> 00:15:34,080 job, also get funds for the organization or the clinic, hire 260 00:15:34,080 --> 00:15:35,520 new folks. I do all of that. 261 00:15:35,880 --> 00:15:39,240 And so I was pretty like resource restricted. 262 00:15:39,720 --> 00:15:42,600 I mean, this is an environment where having kind of Technical 263 00:15:42,600 --> 00:15:44,720 Support would have been incredibly helpful, right? 264 00:15:44,720 --> 00:15:47,800 Because at that time, then I would have been able to use a 265 00:15:47,800 --> 00:15:51,520 tool like empathic to be in the background of therapy sessions. 266 00:15:51,760 --> 00:15:53,440 And we do that. We're in the background of 267 00:15:53,440 --> 00:15:56,440 therapy sessions at times, but to be in the background to be 268 00:15:56,440 --> 00:15:59,880 able to flag to me right as the supervising and clinician 269 00:16:00,200 --> 00:16:03,480 moments where we're growing therapists, right? 270 00:16:03,600 --> 00:16:05,960 And junior therapists, like where they need more support 271 00:16:06,440 --> 00:16:10,000 maybe in detecting suicide, responding to suicide, but also 272 00:16:10,000 --> 00:16:12,360 just need more general support in all the different kinds of 273 00:16:12,360 --> 00:16:14,480 ways. And so this is where I feel like 274 00:16:15,320 --> 00:16:19,240 AI can really help clinics like determine like what areas they 275 00:16:19,240 --> 00:16:21,760 can give more support to. I want to unpack. 276 00:16:21,760 --> 00:16:25,800 That a little bit more so when you think about like areas where 277 00:16:25,800 --> 00:16:29,240 you really need a clinician, where that expertise is really 278 00:16:29,240 --> 00:16:33,360 critical versus the types of tasks that you can sort of 279 00:16:33,640 --> 00:16:37,480 offload onto something like an AI, perhaps a clinically 280 00:16:37,680 --> 00:16:40,920 validated AI. How do you sort of think about 281 00:16:41,080 --> 00:16:44,120 this distinction? Yeah, we don't want to replace 282 00:16:44,120 --> 00:16:45,440 humans. That's not the goal at all. 283 00:16:45,680 --> 00:16:49,240 It's more so fostering precision medicine when we're able to. 284 00:16:49,560 --> 00:16:53,120 And so I think like with your question that there's different 285 00:16:53,120 --> 00:16:56,080 ways that you can engage with AI to do some of the more like 286 00:16:56,080 --> 00:16:59,040 rudimentary things. Maybe, you know, getting the 287 00:16:59,040 --> 00:17:02,000 gist of like a clinical note together so that if you only 288 00:17:02,000 --> 00:17:05,000 have a few minutes between clients, the note is generated 289 00:17:05,000 --> 00:17:07,800 for you and you can go in and add, you know, like nuances that 290 00:17:07,800 --> 00:17:11,400 maybe weren't captured. But I also think that there are 291 00:17:11,400 --> 00:17:15,520 moments like with the work that we do where it can help bolster 292 00:17:15,520 --> 00:17:18,839 the work that clinicians are doing, so help them be maybe 293 00:17:18,839 --> 00:17:22,119 more acutely attuned to conversational elements that 294 00:17:22,119 --> 00:17:25,599 would trigger that somebody might be an imminent risk of 295 00:17:25,599 --> 00:17:29,040 suicide or something else. So I think both, right? 296 00:17:29,040 --> 00:17:31,720 Like, I think that you can do the more simple things and I 297 00:17:31,720 --> 00:17:35,000 think in today's world, I think that clinicians are often times 298 00:17:35,000 --> 00:17:37,440 maybe a little bit more comfortable with doing the 299 00:17:37,440 --> 00:17:39,840 things that, you know, that they do need help with and that like 300 00:17:39,840 --> 00:17:41,800 nobody wants to write notes all day like that. 301 00:17:41,800 --> 00:17:43,560 That is hard. Like as a part, you know, 302 00:17:43,600 --> 00:17:46,200 somebody who's like doing clinical work all the time, Like 303 00:17:46,560 --> 00:17:48,320 that was not the fun part of the job. 304 00:17:49,560 --> 00:17:52,920 But at the same time, I think there's also other areas where I 305 00:17:52,920 --> 00:17:56,320 do maybe like, I want to engage with an LLM to understand, like, 306 00:17:56,320 --> 00:17:58,760 could I do this differently? Could I approach the way that 307 00:17:58,760 --> 00:18:02,680 I'm asking about maybe like the grieving process for this 308 00:18:02,680 --> 00:18:03,840 client? Like could I do that 309 00:18:03,840 --> 00:18:05,440 differently? Am I missing anything? 310 00:18:06,320 --> 00:18:08,640 Almost like a training tool. Yeah. 311 00:18:08,680 --> 00:18:11,600 I mean, I think there's so many use cases for it, right, in 312 00:18:11,600 --> 00:18:15,800 psychotherapy and other clinical spaces where psychologists are, 313 00:18:15,800 --> 00:18:17,640 right? Because I think often times, 314 00:18:17,640 --> 00:18:20,240 like I'm very engaged with like the American Psycho Association 315 00:18:20,560 --> 00:18:22,800 and we talk about different ways that people are using AI. 316 00:18:22,800 --> 00:18:25,440 And I feel like often times we just completely forget that 317 00:18:25,440 --> 00:18:28,160 there's so many psychologists engaged in clinical trials as 318 00:18:28,160 --> 00:18:31,600 well or in drug development. And so I'm a clinician that's 319 00:18:31,600 --> 00:18:33,800 really driven to do good, robust science. 320 00:18:34,160 --> 00:18:36,680 And I make sure that the science also informs the way that I do 321 00:18:36,680 --> 00:18:38,280 my clinical work. Yeah. 322 00:18:38,280 --> 00:18:41,360 So I think even with that, like there's so many different ways 323 00:18:41,360 --> 00:18:45,200 that I would use AI to do my job as a clinician, but to also do 324 00:18:45,200 --> 00:18:48,840 my research and. Here's what you describe as 325 00:18:48,840 --> 00:18:51,800 picking up on some of those things that maybe it's missed by 326 00:18:51,800 --> 00:18:53,560 a clinician during a conversation. 327 00:18:54,320 --> 00:18:58,000 Is that based on the words that are said or is it more even 328 00:18:58,000 --> 00:19:02,440 looking at how they are said in terms of the cadence or the 329 00:19:03,000 --> 00:19:05,200 tone? I think it's both, right? 330 00:19:05,200 --> 00:19:09,360 So I think that sometimes we can pick up on literal words or like 331 00:19:09,680 --> 00:19:13,840 therapist utterances or provider utterances, but I think also you 332 00:19:13,840 --> 00:19:17,960 can pick up on things like synchrony or on therapeutic 333 00:19:17,960 --> 00:19:19,800 alliance, right? Or sentiment. 334 00:19:20,840 --> 00:19:23,560 And, you know, one of the best use cases that I can think of as 335 00:19:23,560 --> 00:19:26,240 well is that often times like our tool is in the background of 336 00:19:26,240 --> 00:19:29,800 like medical visits. And so we make sure that we give 337 00:19:29,800 --> 00:19:32,760 a report to the provider within a few minutes after a visit. 338 00:19:33,080 --> 00:19:38,120 And the provider might see 1020, who knows, right, patients a 339 00:19:38,120 --> 00:19:40,160 day. And so they may engage in 340 00:19:40,160 --> 00:19:43,240 certain communications that might be off putting to a 341 00:19:43,240 --> 00:19:45,880 patient or maybe they just didn't pick up on a certain 342 00:19:46,360 --> 00:19:50,360 description of a symptom. And so Rai told, like, you know, 343 00:19:50,440 --> 00:19:53,920 from 7:00 AM to 7:00 PM is going to be just as effective. 344 00:19:53,920 --> 00:19:57,640 But humans get tired over time and so it's our ability to give 345 00:19:57,640 --> 00:20:01,320 you the report on every visit and to make sure that you really 346 00:20:01,320 --> 00:20:04,200 do have a documentation or details that are going to be 347 00:20:04,200 --> 00:20:05,640 hard for you to remember. I. 348 00:20:06,160 --> 00:20:09,680 Think it's interesting in terms of how, as you say, clinicians 349 00:20:09,680 --> 00:20:13,280 are human. We have various kind of things 350 00:20:13,280 --> 00:20:16,440 that we might struggle with. And so when I think one thing 351 00:20:16,440 --> 00:20:19,960 that I know from my good friends as a leading psychologist here 352 00:20:19,960 --> 00:20:24,960 in Sweden highlighted research around that, the best clinical 353 00:20:24,960 --> 00:20:28,040 psychologists are usually the ones that are in their first 354 00:20:28,040 --> 00:20:30,680 five to 10 years of their practice because they're 355 00:20:30,680 --> 00:20:34,960 following the best practices. And actually one of the tricky 356 00:20:34,960 --> 00:20:40,000 things sometimes is when people get a little bit too secure in 357 00:20:40,000 --> 00:20:42,520 their own patterns of things where they kind of lost sight of 358 00:20:42,520 --> 00:20:45,760 what's actually the best practice, the best way of doing 359 00:20:45,760 --> 00:20:48,000 things. And they've used develop some 360 00:20:48,000 --> 00:20:50,320 good habits, but also maybe some bad habits over the years. 361 00:20:51,160 --> 00:20:54,960 And I'm just curious how you think about that as a challenge 362 00:20:55,080 --> 00:20:58,280 and relating to this? Yeah, I feel like it makes me 363 00:20:58,280 --> 00:21:00,880 think of statistics where you would talk about like regressing 364 00:21:00,880 --> 00:21:02,720 to your own being, right? I think you're right. 365 00:21:02,720 --> 00:21:06,120 Like I think in this example, right, like AI can help you 366 00:21:06,120 --> 00:21:08,720 because, yeah, like I used to teach for a second year 367 00:21:08,720 --> 00:21:12,280 therapist and you need to pay attention to synchrony with like 368 00:21:12,280 --> 00:21:15,640 verbal and non verbal cues. And so especially when you're 369 00:21:15,640 --> 00:21:18,680 like a junior therapist, you're paying attention to every single 370 00:21:18,680 --> 00:21:22,680 detail, right? And so, but I think over time, 371 00:21:22,680 --> 00:21:25,920 we become comfortable, we become confident in some of the ways 372 00:21:25,920 --> 00:21:29,640 that we engage in therapy are beautiful. 373 00:21:29,640 --> 00:21:32,320 But then sometimes we might not catch that we're actually doing 374 00:21:32,320 --> 00:21:34,280 things that might be off putting to a client. 375 00:21:34,720 --> 00:21:37,320 And so you're correct in that, you know, getting the AI 376 00:21:37,320 --> 00:21:41,520 feedback, The AI doesn't care if you're like a first or in your 377 00:21:41,520 --> 00:21:45,120 20th year as a therapist or provider, but doesn't care if 378 00:21:45,120 --> 00:21:46,880 you're the attending or the fellow, right? 379 00:21:46,880 --> 00:21:49,080 Like, it's going to give you that feedback in a particular 380 00:21:49,080 --> 00:21:54,560 way, in a way that isn't necessarily biased to who you 381 00:21:54,560 --> 00:21:58,520 are in terms of like, your training or maybe your title. 382 00:21:59,080 --> 00:22:01,600 But of course, like, AI bias is a real thing that we always have 383 00:22:01,600 --> 00:22:03,320 to build, you know, with that in mind. 384 00:22:03,840 --> 00:22:06,920 But I do think that, yeah, I can definitely see what your friend 385 00:22:06,920 --> 00:22:09,200 is talking about. And I think in this way, like, 386 00:22:09,200 --> 00:22:10,960 in the same way that you would do a tune up, right? 387 00:22:10,960 --> 00:22:14,040 Like your car needs a tune up. You know, we need tune ups. 388 00:22:14,760 --> 00:22:16,400 I always talk about that. I think it's wonderful for 389 00:22:16,400 --> 00:22:19,280 everybody, but ADB in therapy, not just when you feel like you 390 00:22:19,280 --> 00:22:22,280 really need it, because it is this emotional and reflective 391 00:22:22,280 --> 00:22:24,640 tune up that a therapeutic space can provide you. 392 00:22:25,040 --> 00:22:28,000 So I think similarly, yeah, like an AI tool can really pick up on 393 00:22:28,000 --> 00:22:29,680 things. And I would say like with our 394 00:22:29,680 --> 00:22:32,640 own platform, what I do love is that we give you a report 395 00:22:33,120 --> 00:22:35,840 according to, you know, certain detections, right, like ensemble 396 00:22:35,840 --> 00:22:38,520 models of collaboration trust, but you get to choose. 397 00:22:38,680 --> 00:22:41,000 We have 200 behaviors that we have models for. 398 00:22:41,000 --> 00:22:43,600 But in addition to that, it does give you the transcript and 399 00:22:43,600 --> 00:22:46,600 it'll show you like when the detections happen in the 400 00:22:46,600 --> 00:22:50,480 conversation that you had. And then based on how you are 401 00:22:50,480 --> 00:22:53,280 connected to our platform, we can either show you that as soon 402 00:22:53,280 --> 00:22:55,040 as you click on in the transcript, it'll show you the 403 00:22:55,040 --> 00:22:57,840 audio or the video feed. So it's a wonderful training 404 00:22:57,840 --> 00:23:00,400 tool, right? And it's really hard to ignore 405 00:23:00,400 --> 00:23:03,000 that you might have minimized something with an older patient 406 00:23:03,200 --> 00:23:07,400 if the transcript and the video or audio feed shows you exactly 407 00:23:07,400 --> 00:23:10,960 what you said, right? So in that way, it can really 408 00:23:10,960 --> 00:23:15,400 shift how people practice. When I was a supervisor, I would 409 00:23:15,400 --> 00:23:17,480 do my best to make sure, especially that junior therapist 410 00:23:17,480 --> 00:23:20,440 for recording those sessions and really paying attention to what 411 00:23:20,440 --> 00:23:23,200 they said, not necessarily the patient, but like how they were 412 00:23:23,200 --> 00:23:26,440 responding or how they were, you know, just engaged in 413 00:23:26,440 --> 00:23:29,480 conversations or if they were affirming or minimizing at 414 00:23:29,480 --> 00:23:32,040 times. And I think similarly, we can do 415 00:23:32,040 --> 00:23:34,920 that with an AI tool. But here you can do it 100% of 416 00:23:34,920 --> 00:23:38,400 the time, right, versus only 1020% of the sessions that 417 00:23:38,400 --> 00:23:41,160 you're doing? I'm also curious, just thinking 418 00:23:41,160 --> 00:23:45,800 about individual differences in terms of the clinician's 419 00:23:45,800 --> 00:23:48,960 response to that feedback. So if I think about, you know, 420 00:23:49,280 --> 00:23:53,320 you just told me I've minimized the patient like, well, no, I 421 00:23:53,320 --> 00:23:55,400 didn't. What is the full range of 422 00:23:55,400 --> 00:23:59,480 responses that you see? I can definitely imagine someone 423 00:23:59,760 --> 00:24:03,080 being maybe more or less open to that sort of quote UN quote 424 00:24:03,080 --> 00:24:06,400 feedback. Yeah, I think it depends on the 425 00:24:06,400 --> 00:24:07,800 person. It could be the case, right? 426 00:24:07,800 --> 00:24:10,680 That like we pick up on something that's minimizing and 427 00:24:10,680 --> 00:24:12,680 you say like, no, I don't think that it was. 428 00:24:13,000 --> 00:24:15,840 And so we have the ability for you to give a thumbs down. 429 00:24:15,840 --> 00:24:19,000 So on our end, we might see that as like a technology failure, 430 00:24:19,000 --> 00:24:20,240 right? That like, oh, we need to go 431 00:24:20,240 --> 00:24:23,720 back and like really make sure that we're capturing the breadth 432 00:24:23,760 --> 00:24:26,000 of nuance and like what it means to minimize. 433 00:24:26,520 --> 00:24:29,600 So you absolutely can engage with the platform to help us 434 00:24:29,800 --> 00:24:32,840 build stronger models over time. But I think with anything, 435 00:24:32,840 --> 00:24:35,480 right, like if you were to create like a psychometric 436 00:24:35,480 --> 00:24:39,400 instrument of social support and you did that in like 1970, it's 437 00:24:39,400 --> 00:24:42,120 going to look dramatically different what social support is 438 00:24:42,120 --> 00:24:46,440 like in 1990, 2010 and whatnot. So I think that's one thing that 439 00:24:46,440 --> 00:24:48,920 folks always should keep in mind is that whenever you're working 440 00:24:48,920 --> 00:24:52,080 with AI, it's not done right. It's an evolving process and 441 00:24:52,080 --> 00:24:54,280 that product should evolve with time as well. 442 00:24:55,680 --> 00:24:57,200 Yeah. When it comes to the feedback 443 00:24:57,200 --> 00:25:01,120 that they receive, is it only afterwards or do they also get 444 00:25:01,120 --> 00:25:03,760 some form of live feedback during the session? 445 00:25:04,360 --> 00:25:09,200 Our platform is able to do live, but for the most part, I don't 446 00:25:09,200 --> 00:25:11,960 think therapists are necessarily ready for live feedback. 447 00:25:12,960 --> 00:25:17,000 We have an NIH grant that we secured and where we are 448 00:25:17,000 --> 00:25:20,800 building out LLMS that are really focused on like synchrony 449 00:25:20,800 --> 00:25:25,000 and precision medicine. And the hope is to get some user 450 00:25:25,000 --> 00:25:27,960 research where we can understand how our therapist engaging with 451 00:25:27,960 --> 00:25:32,120 it live if we if they were to find that to be effective or 452 00:25:32,120 --> 00:25:34,920 distracting. I think currently though, 453 00:25:34,920 --> 00:25:36,960 because we're not really training folks in this way, 454 00:25:36,960 --> 00:25:38,360 right? And I say this is somebody who 455 00:25:38,360 --> 00:25:41,400 was a professor last year. So like, I know we really aren't 456 00:25:41,520 --> 00:25:45,160 training people this way yet. I think with time, yeah, 457 00:25:45,280 --> 00:25:47,960 probably that there will be more openness and receptiveness to 458 00:25:47,960 --> 00:25:49,800 being able to have like, live coaching. 459 00:25:50,720 --> 00:25:53,480 But I think currently that's just not how therapists are 460 00:25:53,480 --> 00:25:55,520 trained. So I'm not sure that we're quite 461 00:25:55,520 --> 00:25:57,360 ready to, like, make that match happen. 462 00:25:57,600 --> 00:26:01,040 I think the technology is there, but I don't know that as humans 463 00:26:01,040 --> 00:26:02,520 were there yet. All right. 464 00:26:03,360 --> 00:26:07,560 I'm going to touch on a touchy subject, which I think is 465 00:26:07,560 --> 00:26:09,960 probably something that you've thought about before. 466 00:26:10,200 --> 00:26:13,440 One thing that you mentioned was like, we don't want to take 467 00:26:13,680 --> 00:26:16,480 therapist jobs, we don't want to replace therapist, right? 468 00:26:16,960 --> 00:26:18,360 I think every therapist would agree. 469 00:26:18,360 --> 00:26:21,400 I think the APA would agree. We probably all agree. 470 00:26:22,560 --> 00:26:26,200 But of course, when we look at the landscape of AI that's very 471 00:26:26,200 --> 00:26:30,520 swiftly evolving, we see, you know, new RCT's coming out, new 472 00:26:30,520 --> 00:26:33,960 tools being tested, some clinically validate, many not 473 00:26:33,960 --> 00:26:38,840 clinically validated. We see people engaging with LLMS 474 00:26:38,840 --> 00:26:41,200 like ChatGPT as if they were therapists. 475 00:26:41,200 --> 00:26:45,320 We see these like, you know, AI health coaches, this huge, 476 00:26:45,320 --> 00:26:49,720 incredible proliferation of tools, some better than others. 477 00:26:50,000 --> 00:26:55,440 In many ways, it feels to me like the reality is that it's a 478 00:26:55,440 --> 00:27:01,880 little bit out of our hands. Whether therapists are at least 479 00:27:01,880 --> 00:27:06,480 some being replaced by artificial intelligence tools, 480 00:27:06,840 --> 00:27:10,720 How do you grapple with this changing environment? 481 00:27:11,440 --> 00:27:13,680 And how are you thinking about the future? 482 00:27:15,280 --> 00:27:18,400 When we talk about working with an AI chat bot, I think that 483 00:27:18,400 --> 00:27:20,560 there are different ways that people are developing them. 484 00:27:20,840 --> 00:27:24,080 And so you could look at Wobot as an example where you have a 485 00:27:24,080 --> 00:27:28,240 deterministic AI, right? And so there are scripts on the 486 00:27:28,240 --> 00:27:29,840 back end. It's not generative. 487 00:27:30,120 --> 00:27:35,480 And so there's like 100% control of how the AI model would 488 00:27:35,480 --> 00:27:38,600 respond to you if you're talking about anxiety or loneliness or 489 00:27:38,600 --> 00:27:41,160 whatnot. And then I do think that there's 490 00:27:41,160 --> 00:27:46,240 other ways that people are able to add some kind of like I'm 491 00:27:46,240 --> 00:27:48,600 thinking of like when you bowl right and you're a kid and you 492 00:27:48,600 --> 00:27:49,960 have these like kind of, you know, the. 493 00:27:50,200 --> 00:27:51,640 Bumpers. Yeah, there you go. 494 00:27:51,680 --> 00:27:53,320 Right. Like, so I love the metaphor, 495 00:27:54,800 --> 00:27:57,880 the visual bumpers of how you would work with, Yeah, with 496 00:27:57,880 --> 00:28:00,680 generative AI. In terms of like working with 497 00:28:00,680 --> 00:28:04,280 therapy chat bots. I do think that if you can have 498 00:28:04,280 --> 00:28:08,240 a chat bot that is able to flag clinical risk and then make sure 499 00:28:08,240 --> 00:28:11,280 that a therapist is connected immediately, right, or flagged, 500 00:28:11,560 --> 00:28:15,120 then I think that's safe. But, you know, it's tough 501 00:28:15,120 --> 00:28:17,760 because I don't think that we have such clear guidelines or 502 00:28:17,760 --> 00:28:20,520 policies of like how these are supposed to roll out. 503 00:28:20,840 --> 00:28:24,200 And I think that, you know that there are organizations like 504 00:28:24,200 --> 00:28:27,640 American Psych Association or American Psychiatric Association 505 00:28:27,640 --> 00:28:29,520 that are really actively working on it. 506 00:28:29,840 --> 00:28:33,600 The challenge is that like, you know, LLM development is so fast 507 00:28:34,120 --> 00:28:37,600 that sometimes the policies are not coming out fast enough to 508 00:28:37,600 --> 00:28:40,520 really map on to the pace of innovation. 509 00:28:41,240 --> 00:28:44,280 But I don't necessarily think that things are good or bad. 510 00:28:44,560 --> 00:28:48,760 I think that if you are somebody who needs therapeutic support at 511 00:28:48,760 --> 00:28:51,560 2:00 AM, there's not a therapist that's going to answer the phone 512 00:28:51,560 --> 00:28:52,960 call, right? Like we have really clear 513 00:28:52,960 --> 00:28:57,680 boundaries for reasons that are important to have, but people 514 00:28:57,680 --> 00:28:59,840 are in crisis at different times. 515 00:29:00,240 --> 00:29:03,600 And sometimes to have a live human is cost prohibitive. 516 00:29:04,000 --> 00:29:07,160 Or maybe you live in a rural area and it is impossible to get 517 00:29:07,160 --> 00:29:09,200 to a live person more immediately. 518 00:29:09,480 --> 00:29:12,800 So I do think that there are times where like virtual care or 519 00:29:12,800 --> 00:29:16,280 even therapy chat bots, like there may be made space for it. 520 00:29:17,280 --> 00:29:19,560 But I think that we have to make sure that the people building 521 00:29:19,560 --> 00:29:22,920 them are clinicians, that they do have expertise. 522 00:29:22,960 --> 00:29:26,680 And if you're in alignment with APA ethics, then yeah, you don't 523 00:29:26,680 --> 00:29:29,080 practice outside of your confidence, right? 524 00:29:29,080 --> 00:29:32,840 It would be inappropriate for me to maybe take a client with the 525 00:29:32,840 --> 00:29:35,800 severe eating disorder today because that isn't an area that 526 00:29:35,800 --> 00:29:38,240 I've like really focused on in the last couple of years. 527 00:29:38,480 --> 00:29:41,760 So I would imagine that there's elements of an assessment that 528 00:29:41,760 --> 00:29:44,080 I'm missing, right? Or like interventions that are 529 00:29:44,080 --> 00:29:47,760 just not actually helpful. So I think similarly when we're 530 00:29:47,760 --> 00:29:50,080 developing these kinds of chat bots, we have to be really 531 00:29:50,080 --> 00:29:53,200 specific that it's, and I saw them the paper, I think we're 532 00:29:53,200 --> 00:29:55,160 talking about the same paper where they talked about working 533 00:29:55,160 --> 00:29:58,680 with major depressive disorder, generalized ID disorder, and I 534 00:29:58,680 --> 00:30:00,960 believe eating disorder, but a very specific type. 535 00:30:01,560 --> 00:30:05,440 And so it's clear that they were really focused on like we are 536 00:30:05,440 --> 00:30:07,800 doing an intake that is comprehensive and we're really 537 00:30:07,800 --> 00:30:09,640 understanding what the symptomatology is. 538 00:30:10,240 --> 00:30:12,880 And I imagine that they also did a comprehensive clinical risk 539 00:30:12,880 --> 00:30:15,680 detection, right? So I think that there are times 540 00:30:15,680 --> 00:30:18,680 where you can do it, but it has to be as careful as humanly 541 00:30:18,680 --> 00:30:22,160 possible in the same way that you would engage with that level 542 00:30:22,160 --> 00:30:26,480 of care in a clinic. I'll add this Therabot paper. 543 00:30:26,480 --> 00:30:29,880 They also assembled custom data sets, so they didn't just Hoover 544 00:30:29,880 --> 00:30:32,240 the Internet as many of our LLMS do. 545 00:30:32,400 --> 00:30:36,160 So they used this data set which was based on evidence based 546 00:30:36,160 --> 00:30:38,960 practices like CBT, cognitive behavioral therapy. 547 00:30:39,040 --> 00:30:42,440 So we have a little more confidence that the data that's 548 00:30:42,440 --> 00:30:46,800 going in is training the model in a more reliable way than some 549 00:30:46,800 --> 00:30:49,320 of the other more unpredictable models out there. 550 00:30:50,200 --> 00:30:51,240 Yeah. And I would say, you know, for 551 00:30:51,240 --> 00:30:54,680 us, like we recently built out a feature where we are doing like 552 00:30:54,680 --> 00:30:56,520 oversight of clinical outcome assessments. 553 00:30:56,800 --> 00:30:59,680 So if we built out this new tool for clinical outcome 554 00:30:59,680 --> 00:31:02,840 assessments, it's really important for us that we make 555 00:31:02,840 --> 00:31:06,120 sure that our AI tool is able to detect that somebody is talking 556 00:31:06,120 --> 00:31:08,320 about reported versus maybe apparent sadness. 557 00:31:08,720 --> 00:31:12,520 And then when they are doing a sort of oversight of the scoring 558 00:31:12,520 --> 00:31:15,520 accuracy, we benchmark it in with licensed clinicians. 559 00:31:16,120 --> 00:31:18,520 So we always want to make sure that there's a really high level 560 00:31:18,520 --> 00:31:21,120 of precision. And I will say that like we have 561 00:31:21,120 --> 00:31:24,640 a much higher level of precision because it is an LLM, right? 562 00:31:24,640 --> 00:31:29,000 Like it's AI then you might expect between humans. 563 00:31:29,720 --> 00:31:32,480 But I think it is that like you have to really be building in a 564 00:31:32,480 --> 00:31:35,880 way that puts patient safety at the heart of the work. 565 00:31:36,160 --> 00:31:39,440 And that's why whenever I'm, I'm talking with clinicians about 566 00:31:39,440 --> 00:31:42,560 like different AI tools, I'm always really focused on whether 567 00:31:42,560 --> 00:31:45,400 there are like clinicians and especially licensed clinicians 568 00:31:45,440 --> 00:31:47,440 on the team that are helping to develop the tool. 569 00:31:48,240 --> 00:31:49,560 Because we're LED with ethics, right? 570 00:31:49,560 --> 00:31:52,000 And we could actually get our license taken away for not 571 00:31:52,000 --> 00:31:54,760 focused on ethics. And so you have to practice and 572 00:31:54,760 --> 00:31:56,160 you have to really build with care. 573 00:31:57,360 --> 00:31:59,760 Yeah. Given that currently what you've 574 00:31:59,760 --> 00:32:04,840 kind of illustrated in terms of the support given to clinicians 575 00:32:04,840 --> 00:32:08,480 in terms of having that kind of sense of someone that gives you 576 00:32:08,480 --> 00:32:11,480 feedback in terms of showing you if you missed something or maybe 577 00:32:11,880 --> 00:32:18,560 helping you detect signs of various kind of serious or maybe 578 00:32:18,560 --> 00:32:21,720 also some more subtle things. That is really useful. 579 00:32:22,200 --> 00:32:27,000 And I feel like in some ways that same guidance could be 580 00:32:27,000 --> 00:32:31,440 given also to not a human, but an AI shotbot, basically giving 581 00:32:31,440 --> 00:32:36,840 fair about the same support as like some form of almost double 582 00:32:36,880 --> 00:32:41,080 AI team where you have a like a clinician AI and then you have 583 00:32:41,080 --> 00:32:45,240 this kind of support guiding expert AI like, and it helps 584 00:32:45,240 --> 00:32:48,120 make sure that the clinician is doing a good job and so on. 585 00:32:49,280 --> 00:32:53,960 Have that been something you've been looking at as a additional 586 00:32:54,720 --> 00:32:58,000 path to provide value towards an AI instead of a human? 587 00:32:58,760 --> 00:33:01,200 In this area, we're talking about like clinical assessments 588 00:33:01,200 --> 00:33:05,960 that we ask our LMM to explain, like why did you choose that by 589 00:33:05,960 --> 00:33:08,840 like, why did you detect this? Like tell us behaviorally what 590 00:33:08,840 --> 00:33:11,960 you picked up on. And for us, it's like an area of 591 00:33:11,960 --> 00:33:14,440 transparency because you have your training data set. 592 00:33:14,440 --> 00:33:18,160 And when we're building, we make sure that our clinicians also 593 00:33:18,160 --> 00:33:19,800 memo, right? Like why they're choosing 594 00:33:19,800 --> 00:33:22,000 particular things. And then we compare that to what 595 00:33:22,000 --> 00:33:26,360 the AI is producing as well. And so you can ask your AI agent 596 00:33:26,360 --> 00:33:29,080 to give you their logic of like why they chose a particular 597 00:33:29,080 --> 00:33:32,000 thing or why they are maybe scoring somebody as having 598 00:33:32,000 --> 00:33:33,960 moderate versus severe depression or something of that 599 00:33:33,960 --> 00:33:37,160 sort. So in that way you can have 600 00:33:37,160 --> 00:33:40,480 checks and balances. It's sort of how you think of 601 00:33:40,480 --> 00:33:44,480 adversarial your Gans, right? Where you have your two models 602 00:33:44,560 --> 00:33:47,200 that are sort of playing off of each other, doing their own 603 00:33:47,200 --> 00:33:51,320 automatic checks and balances. Yeah, but you always have to 604 00:33:51,320 --> 00:33:53,880 have a clinician in the loop. Like, that's so critical all the 605 00:33:53,880 --> 00:33:55,480 time. But I think also because it's 606 00:33:55,480 --> 00:33:59,200 like you can't really ask folks to audit themselves. 607 00:33:59,200 --> 00:34:02,880 Like, we're so biased, right? And so I feel like, you know, 608 00:34:02,880 --> 00:34:05,480 the LLM would likely be less biased, but at the same time, 609 00:34:05,480 --> 00:34:09,159 it's like, no, this is where clinicians really matter. 610 00:34:10,040 --> 00:34:12,760 So I think in two ways, right? One is that ask the AI to 611 00:34:12,760 --> 00:34:17,199 explain its logic and then map that onto what clinicians are 612 00:34:17,199 --> 00:34:19,080 doing. And then at the same time make 613 00:34:19,080 --> 00:34:22,120 sure that clinicians are reviewing portions of the work 614 00:34:22,120 --> 00:34:25,600 or especially areas where you have to handle it with more care 615 00:34:25,600 --> 00:34:27,639 you. Used to be a devil's advocate 616 00:34:27,639 --> 00:34:30,239 here. I guess I'll be interested to 617 00:34:30,239 --> 00:34:33,000 hear your answer with like, why does clinicians matter? 618 00:34:33,000 --> 00:34:34,960 Or like in what way? Because we talked about this day 619 00:34:34,960 --> 00:34:38,000 of like, I think serving at the top of your license as an 620 00:34:38,000 --> 00:34:39,719 example, Like it's really important that we can help 621 00:34:39,719 --> 00:34:41,159 people to serve a top of their license. 622 00:34:41,159 --> 00:34:44,199 And I think it would be really interesting to have a better 623 00:34:44,199 --> 00:34:48,000 shared understanding of what is the things that we do as humans 624 00:34:48,000 --> 00:34:51,239 in various contexts, like in a clinical context or in other 625 00:34:51,239 --> 00:34:54,960 contexts where we can do the most good and add the most value 626 00:34:54,960 --> 00:34:56,639 and so on. So I guess I'm just interested 627 00:34:56,639 --> 00:35:01,560 to hear from your end, like, why does clinicians matter? 628 00:35:01,560 --> 00:35:04,920 Like what is the most important aspects that they add to 629 00:35:04,960 --> 00:35:07,840 treatment? Yeah, I do think clinicians 630 00:35:07,840 --> 00:35:10,440 matter because when we talk about clinical risk or when you 631 00:35:10,440 --> 00:35:13,000 talk about somebody maybe for example, or having suicidal 632 00:35:13,000 --> 00:35:15,240 ideation, the thing that you really have to pay attention to 633 00:35:15,240 --> 00:35:18,280 is whether it's imminent, right? So you have to very explicitly 634 00:35:18,280 --> 00:35:21,080 ask somebody, like, can you go home and be safe, right? 635 00:35:21,080 --> 00:35:24,480 Or you need to not be vague. You want to really ask somebody 636 00:35:24,480 --> 00:35:27,080 the exact question of like, are they going to kill themselves, 637 00:35:27,120 --> 00:35:28,680 right? Like you have to do these kinds 638 00:35:28,680 --> 00:35:31,400 of things. And so we are trained to do 639 00:35:31,400 --> 00:35:32,760 that. We're trained to understand the 640 00:35:32,760 --> 00:35:34,840 nuance. We're trained to, like, remove 641 00:35:34,840 --> 00:35:37,840 vagueness. We're also trained to ensure 642 00:35:37,840 --> 00:35:41,240 that we really do understand, like what imminent risk means. 643 00:35:41,520 --> 00:35:44,520 And I think sometimes conversational output might miss 644 00:35:44,520 --> 00:35:47,440 that to some degree. And so that's where you want a 645 00:35:47,440 --> 00:35:50,880 clinician in the room. I think there's also things I do 646 00:35:50,880 --> 00:35:53,440 think multimodal technologies are being built now, right? 647 00:35:53,440 --> 00:35:56,520 So, you know, AI is going to be like a year from now. 648 00:35:56,520 --> 00:35:59,840 Our conversation could be so incredibly different, but I do 649 00:35:59,840 --> 00:36:02,000 think that they're also things like affect, right? 650 00:36:02,000 --> 00:36:05,480 So you're trained as a clinician to be able to see depression, to 651 00:36:05,480 --> 00:36:07,560 be able to hear depression. It's not just what somebody 652 00:36:07,560 --> 00:36:09,200 says, but it's also how they present. 653 00:36:09,560 --> 00:36:11,640 So I think in those moments, that's where you do want a 654 00:36:11,640 --> 00:36:14,160 clinician in the room. And then the final thing I'll 655 00:36:14,160 --> 00:36:18,040 say is that like clinicians are humans too, and we have biases 656 00:36:18,360 --> 00:36:20,480 and we're going to pick up on things that are salient to us, 657 00:36:20,480 --> 00:36:22,760 not just with our lived experience, but how we are 658 00:36:22,760 --> 00:36:25,560 trained theoretically. And so if you're somebody who, 659 00:36:26,400 --> 00:36:29,160 right, has been part of the team that has like created a 660 00:36:29,160 --> 00:36:32,560 clinically validated training data set, but you are trained as 661 00:36:32,560 --> 00:36:36,240 a psychodynamic therapist, then you might be training an LLM to 662 00:36:36,240 --> 00:36:39,800 be picking up on family dynamics more than maybe behaviors or 663 00:36:39,800 --> 00:36:42,000 being out of alignment with values in the way that you would 664 00:36:42,000 --> 00:36:44,320 if you were doing. So our theories and our 665 00:36:44,320 --> 00:36:47,960 trainings matter a lot. And so that's where I think you 666 00:36:47,960 --> 00:36:50,080 should have different kinds of clinicians in the loop because 667 00:36:50,280 --> 00:36:55,400 it also helps to ensure that your AI is not missing things, 668 00:36:55,400 --> 00:36:57,640 right, because it's been trained to only pick up on certain 669 00:36:57,640 --> 00:37:01,520 characteristics. So one thing from my end when I 670 00:37:01,520 --> 00:37:05,360 get this question, and I think the same is true for things like 671 00:37:05,680 --> 00:37:09,360 Shachi PT will be the death of Duolingo or like whenever 672 00:37:09,360 --> 00:37:12,760 there's like this kind of very sensationalist takes around, he 673 00:37:12,760 --> 00:37:14,560 has developed really high capabilities. 674 00:37:15,000 --> 00:37:17,120 I would say that there's a difference between capabilities 675 00:37:17,120 --> 00:37:22,880 and in this case also a part of treatment is actually having the 676 00:37:22,880 --> 00:37:27,280 social accountability, support and understanding of another 677 00:37:27,280 --> 00:37:29,760 human being. And like knowing that it's not a 678 00:37:29,760 --> 00:37:32,520 human being that cares for you and that's going to follow up 679 00:37:32,520 --> 00:37:34,960 with you and that you will follow up with. 680 00:37:35,080 --> 00:37:38,200 Like that is not something you can also use to replace just 681 00:37:38,200 --> 00:37:40,120 because the capability might exist within an AI. 682 00:37:41,160 --> 00:37:43,920 Even if it feels human, you know that it's not. 683 00:37:44,240 --> 00:37:46,640 Right. But it is interesting because I 684 00:37:46,640 --> 00:37:48,880 feel like Robot has put out some research, right? 685 00:37:48,880 --> 00:37:52,320 They've had a lot of scientific papers where people sometimes 686 00:37:52,320 --> 00:37:55,320 feel like they can say more to an AI agent because they know 687 00:37:55,320 --> 00:37:56,920 they're not being judged in the same way. 688 00:37:58,080 --> 00:37:59,480 And so I do think that there's that. 689 00:37:59,480 --> 00:38:02,000 But then I also think that, like, you know, our company is 690 00:38:02,000 --> 00:38:06,280 named empathic, right? So like our core models are 691 00:38:06,280 --> 00:38:09,080 based on the common factors, which are like the key 692 00:38:09,080 --> 00:38:12,360 ingredients across all different kinds of theoretical 693 00:38:12,360 --> 00:38:15,680 orientations that we know really matter for behavior change. 694 00:38:15,960 --> 00:38:19,400 So things like collaboration and trust and unconditional positive 695 00:38:19,400 --> 00:38:23,000 regard, right? And so that, that isn't the core 696 00:38:23,000 --> 00:38:25,960 of like how we built our models. So we wanted to make sure that 697 00:38:25,960 --> 00:38:29,600 we could pay attention to the nuance in popular media or just 698 00:38:29,600 --> 00:38:32,200 like in the ways that we talk about empathy, like in a more 699 00:38:32,200 --> 00:38:35,320 common way, it's this idea that like you're a nice person, 700 00:38:35,320 --> 00:38:37,080 right? Or that, you know, that you 701 00:38:37,080 --> 00:38:40,640 care, but empathy really is like about, if you really pay 702 00:38:40,640 --> 00:38:44,120 attention to the behaviors, it's about accurate understanding, 703 00:38:44,440 --> 00:38:46,040 right? So ensuring that you really do 704 00:38:46,040 --> 00:38:48,360 accurately understand what somebody is trying to convey to 705 00:38:48,360 --> 00:38:51,600 you about a particular instance. And so I do think that there are 706 00:38:51,600 --> 00:38:54,960 moments where sometimes the way that AI models are built, 707 00:38:54,960 --> 00:38:57,240 especially, you know, because that's why training data is so 708 00:38:57,240 --> 00:39:00,840 critical, that you can train your AI to be extremely 709 00:39:00,840 --> 00:39:03,960 empathetic and more so maybe than a lot of humans. 710 00:39:04,680 --> 00:39:08,160 So it's not just the training of like, hey, you missed this, you 711 00:39:08,160 --> 00:39:10,840 know, this symptom. But it's also like, hey, the way 712 00:39:10,840 --> 00:39:13,600 that you said this was out of alliance, right? 713 00:39:13,600 --> 00:39:16,920 Or it's out of synchrony, it was not empathetic. 714 00:39:17,280 --> 00:39:20,000 And so here's coaching. Here's a report for how you 715 00:39:20,000 --> 00:39:24,240 could do that better. Yeah, I love that study from way 716 00:39:24,240 --> 00:39:28,320 back when with the Reddit doctor responses versus the generative 717 00:39:28,440 --> 00:39:29,960 AI. People found the Gen. 718 00:39:30,040 --> 00:39:32,960 AI response to be more empathetic than the real 719 00:39:32,960 --> 00:39:35,200 doctors. And of course, you know, many 720 00:39:35,200 --> 00:39:38,280 studies have been done since then really replicating this 721 00:39:38,280 --> 00:39:40,840 finding. So I find that extremely 722 00:39:40,840 --> 00:39:44,680 interesting, knowing that the machine is not experiencing 723 00:39:44,880 --> 00:39:47,680 empathy, but it is conveying empathy. 724 00:39:47,680 --> 00:39:52,040 It is mimicking empathy to I guess like more convincing 725 00:39:52,040 --> 00:39:54,720 degree than the humans. Yeah. 726 00:39:55,120 --> 00:39:57,960 And but I would say right, like that's where like AI is not a 727 00:39:57,960 --> 00:40:00,480 black box. Like you have training data, 728 00:40:00,480 --> 00:40:03,080 right? And you train AI like built on 729 00:40:03,080 --> 00:40:04,960 that data. So that's why it matters that 730 00:40:04,960 --> 00:40:08,280 the team building it is diverse because it needs to capture the 731 00:40:08,280 --> 00:40:11,040 breath of how people experience the world and what they respond 732 00:40:11,040 --> 00:40:13,760 to. There's a study that I loved 733 00:40:13,960 --> 00:40:17,120 they did use machine learning to track like Reddit responses 734 00:40:17,120 --> 00:40:19,200 around mental health and they found that. 735 00:40:19,680 --> 00:40:23,120 You know, according to like the typical like, DSM criteria for 736 00:40:23,120 --> 00:40:25,920 depression, it was missing depression for black women 737 00:40:26,120 --> 00:40:29,720 because black women express different kinds of ways of like, 738 00:40:30,120 --> 00:40:33,520 whether that be guilt, shame, somatic complaints that aren't 739 00:40:33,520 --> 00:40:36,200 typically captured in the way that we think about depression. 740 00:40:36,480 --> 00:40:40,040 But because you would imagine that like the training data that 741 00:40:40,040 --> 00:40:44,080 went into the DSM, like the same idea wasn't built around having 742 00:40:44,080 --> 00:40:46,800 enough diversity. But I think that there's really 743 00:40:46,800 --> 00:40:48,840 beautiful ways that you can use machine learning. 744 00:40:48,840 --> 00:40:53,040 And I think you asked this earlier, Sam, but it's not just 745 00:40:53,040 --> 00:40:54,920 like what it was said, but it's how it was said. 746 00:40:55,200 --> 00:40:57,800 And I think that's the thing that I care a lot about, right? 747 00:40:57,800 --> 00:41:02,000 Cuz I've had a long history of being a clinical researcher and 748 00:41:02,000 --> 00:41:04,480 I care a great deal about like how do you work with different 749 00:41:04,480 --> 00:41:06,200 kinds of data? And I've always been somebody 750 00:41:06,200 --> 00:41:08,560 who goes back to like qualitative, quantitative mixed 751 00:41:08,560 --> 00:41:11,000 methodologies. Because I feel like you can with 752 00:41:11,000 --> 00:41:13,920 quantitative, you can tell that a relationship exists, but you 753 00:41:13,920 --> 00:41:15,680 need qualitative to go underneath. 754 00:41:15,680 --> 00:41:18,400 And like, even if you find something is mediating, you need 755 00:41:18,400 --> 00:41:20,760 to understand like what exactly is happening here. 756 00:41:21,040 --> 00:41:23,640 And so I think similarly, like if you use natural language 757 00:41:23,640 --> 00:41:27,760 processing like we use, then it's like paying attention to 758 00:41:27,760 --> 00:41:30,360 conversational cues. And it's not just like a yes or 759 00:41:30,360 --> 00:41:33,600 no that something happened, but it's more so how are people 760 00:41:33,600 --> 00:41:37,400 communicating guilt or how are people communicating empathy and 761 00:41:37,400 --> 00:41:41,120 being able to get at that kind of nuance in communications in a 762 00:41:41,120 --> 00:41:44,360 way that is really hard to do as a human 100% of the time. 763 00:41:44,400 --> 00:41:46,520 Yeah. And especially as you say, you 764 00:41:46,520 --> 00:41:51,120 know, being human that engages with like 20 patients per day 765 00:41:51,120 --> 00:41:53,720 with a limited time, like all these things contribute, I think 766 00:41:53,720 --> 00:41:57,240 to what we talk about empathy is maybe on the best of days, in 767 00:41:57,240 --> 00:42:00,240 the best of moments, you know, humans will be most AI. 768 00:42:00,880 --> 00:42:04,080 But on a regular day and it's all of the other things going 769 00:42:04,080 --> 00:42:07,320 on, it's hard to maintain the same level of empathy. 770 00:42:07,640 --> 00:42:09,920 You know, like they do talk about, like, decisional 771 00:42:09,960 --> 00:42:12,120 exhaustion. So like humans, we do bias all 772 00:42:12,120 --> 00:42:14,480 the time. I do think that there's ways 773 00:42:14,480 --> 00:42:19,320 that AI can really help humans to, yeah, just to do work that 774 00:42:19,320 --> 00:42:22,320 is more precise, but also just more comprehensive that we've 775 00:42:22,320 --> 00:42:27,400 ever been able to do before. So talking about bias, 776 00:42:27,400 --> 00:42:30,680 interestingly, we talked a lot about algorithms going to be 777 00:42:30,680 --> 00:42:34,480 biased, but it's often times, I think the interesting thing is 778 00:42:34,480 --> 00:42:37,080 especially for us as a paper scientist to see both side of 779 00:42:37,080 --> 00:42:40,120 that, like where this human bias, but there's also algorithm 780 00:42:40,120 --> 00:42:41,920 bias and kind of the best of worlds. 781 00:42:41,920 --> 00:42:43,520 You want to kind of mitigate both of them, like you said, 782 00:42:43,600 --> 00:42:47,080 have good training data that supports good algorithms that 783 00:42:47,080 --> 00:42:51,480 are better trained on a wide range of diverse data, but also 784 00:42:51,480 --> 00:42:55,080 understanding how we can support humans in being less biased. 785 00:42:55,160 --> 00:42:58,360 What do you think about the most effective ways to mitigate both 786 00:42:58,360 --> 00:43:01,280 of those sites? Like mitigate bias in the LLMS 787 00:43:01,280 --> 00:43:03,000 but also mitigate bias in the humans? 788 00:43:03,880 --> 00:43:05,280 Yeah. But first I'll say that I think 789 00:43:05,280 --> 00:43:07,840 that they're experts out there, right, who know a lot more about 790 00:43:07,840 --> 00:43:09,600 this than I do. And I still have a lot of 791 00:43:09,600 --> 00:43:12,000 learning to do. But if you're a researcher, I 792 00:43:12,000 --> 00:43:14,800 think that there's ways that like we know how to develop 793 00:43:14,800 --> 00:43:17,160 psychometric instruments, right? Like we understand that. 794 00:43:17,160 --> 00:43:20,760 And I think sometimes using the same logic where it is things 795 00:43:20,760 --> 00:43:24,360 like have you been comprehensive in understanding the content? 796 00:43:24,400 --> 00:43:26,520 And so that's kind of similar to like training data. 797 00:43:26,840 --> 00:43:29,480 Do you have diverse people like that comprise your training 798 00:43:29,480 --> 00:43:30,840 data? But do you also have diverse 799 00:43:30,840 --> 00:43:34,440 people annotating your data to make sure that when you say 800 00:43:34,440 --> 00:43:39,320 trust that is, is it just trust for a middle-aged white man, 801 00:43:39,320 --> 00:43:40,760 right. But it's trust for the breadth 802 00:43:40,760 --> 00:43:44,200 of what human experience can be. And sometimes that can be 803 00:43:44,200 --> 00:43:46,640 impossible to do, right? Cuz it's like you can't get that 804 00:43:46,640 --> 00:43:49,000 for every single person. And we know that intra 805 00:43:49,240 --> 00:43:52,160 variability is oftentimes larger than intergroup. 806 00:43:52,360 --> 00:43:55,560 But I think also you can test the performance of your AI 807 00:43:55,560 --> 00:43:59,240 precision across groups, right? So if you're benchmarking it 808 00:43:59,240 --> 00:44:04,240 against clinicians, you can see, is our AI performing with bias? 809 00:44:04,240 --> 00:44:06,200 Is it more precise for men than women? 810 00:44:06,200 --> 00:44:09,600 Is it more precise for white versus Asian communities or 811 00:44:09,600 --> 00:44:11,720 whatnot? So those are like 2 of the ways 812 00:44:11,720 --> 00:44:14,640 that I think about it more immediately is I always am 813 00:44:14,640 --> 00:44:17,520 paying a great deal of attention to the training data. 814 00:44:17,800 --> 00:44:20,920 And then addition to that, once we have working models, right, 815 00:44:20,920 --> 00:44:23,400 then we do test the precision against different kinds of 816 00:44:23,400 --> 00:44:25,800 groups to see whether they are more or less precise. 817 00:44:25,960 --> 00:44:27,560 Like are they demonstrating bias? 818 00:44:28,120 --> 00:44:30,120 But then I think that there's other ways that you can do it 819 00:44:30,280 --> 00:44:33,720 with user research as well. And I do think that there's 820 00:44:33,760 --> 00:44:37,480 other elements as well. You know, the NIH grant that we 821 00:44:37,480 --> 00:44:40,400 got really was built around synchrony, but also like 822 00:44:40,400 --> 00:44:42,440 elements of cultural responsiveness too. 823 00:44:42,760 --> 00:44:45,800 And so I think that's something that I'm so excited to be able 824 00:44:45,800 --> 00:44:49,000 to focus on in the next few months and to kind of understand 825 00:44:49,000 --> 00:44:52,160 like, what does it mean, right, to really pay attention to like 826 00:44:52,160 --> 00:44:55,080 maybe miss cultural opportunities or to not evoke 827 00:44:55,080 --> 00:44:58,160 how culture might impact how somebody is experiencing their 828 00:44:58,160 --> 00:45:01,560 mental health condition or responding to intervention. 829 00:45:01,840 --> 00:45:05,720 But I love this idea of doing this because I feel like so far 830 00:45:05,720 --> 00:45:09,080 right, like AI amongst clinicians is really focused on 831 00:45:09,080 --> 00:45:11,800 like note taking. And now we're maybe seeing 832 00:45:11,800 --> 00:45:14,400 things, right, of like, you know, deterministic ways of like 833 00:45:14,400 --> 00:45:17,480 how we might respond to depression, which is great. 834 00:45:17,480 --> 00:45:20,040 Like that's a beautiful evolution and I love seeing it. 835 00:45:20,520 --> 00:45:24,680 But I also think that we can use AI in ways that that we've never 836 00:45:24,680 --> 00:45:27,400 done before, right? Like understanding patterns of 837 00:45:27,440 --> 00:45:30,880 maybe how rural youth or youth in like certain communities, 838 00:45:30,880 --> 00:45:34,120 like maybe trans youth, like talk about their mental health 839 00:45:34,440 --> 00:45:37,600 that we can understand like communication patterns in a way 840 00:45:37,600 --> 00:45:40,960 that we've never done before. And so I think that it's about 841 00:45:40,960 --> 00:45:43,360 doing precision medicine, but precision medicine means that 842 00:45:43,360 --> 00:45:46,120 you really are paying attention to like individual differences 843 00:45:46,120 --> 00:45:49,240 and doing your best to really develop interventions that like 844 00:45:49,320 --> 00:45:52,240 are going to be most effective for the patient in front of you, 845 00:45:52,520 --> 00:45:54,840 not like a general kind of intervention that's supposed to 846 00:45:54,840 --> 00:45:58,200 be effective for everybody. I do think that AI can allow us 847 00:45:58,200 --> 00:46:00,720 to do that in a way that we haven't done before. 848 00:46:01,880 --> 00:46:05,720 All right, it is time to move on to our quick fire round, which 849 00:46:05,720 --> 00:46:09,120 we call to AI or not to AI. Are you ready? 850 00:46:10,120 --> 00:46:14,840 I don't know, yes. You must be, you know, it's fun. 851 00:46:14,840 --> 00:46:16,320 You're going to like it. We're just going to give you a 852 00:46:16,320 --> 00:46:19,640 bunch of tasks and you're going to tell us whether you think 853 00:46:19,640 --> 00:46:23,320 it's well suited to AI or not. OK. 854 00:46:24,720 --> 00:46:29,120 All right, first one a digital twin of yourself to ask for 855 00:46:29,120 --> 00:46:32,920 advice. No, I think I already have my 856 00:46:32,920 --> 00:46:35,680 journal for that, so no. OK. 857 00:46:36,080 --> 00:46:39,320 But also because I feel like that's a pretty biased advice, 858 00:46:39,320 --> 00:46:43,320 right? Maybe it's what you want to 859 00:46:43,320 --> 00:46:44,880 hear. Exactly. 860 00:46:44,880 --> 00:46:48,480 I mean. Just give me the permission, 861 00:46:48,480 --> 00:46:52,000 please, OK? A playbook of cultural 862 00:46:52,000 --> 00:46:55,800 expressions of distress. I like that with a caveat 863 00:46:55,800 --> 00:46:56,840 though. I was like who built it? 864 00:46:56,840 --> 00:46:58,560 But yes, I like the theoretically. 865 00:46:58,560 --> 00:47:04,400 I like the idea of it. OK, what about this VR exposure 866 00:47:04,400 --> 00:47:08,320 therapy that is AI powered? Basically providing some form of 867 00:47:08,600 --> 00:47:12,320 personalized exposure scenarios for patients based on their 868 00:47:12,320 --> 00:47:14,920 various phobias, PTSDS and social anxieties. 869 00:47:15,640 --> 00:47:20,480 Yes, and that's already around. And I think it's funny that you 870 00:47:20,480 --> 00:47:24,400 mentioned your journal because the next one is a fact checking 871 00:47:24,400 --> 00:47:28,480 journal. Fact checking I would be really 872 00:47:28,480 --> 00:47:30,040 worried about the training data there. 873 00:47:30,320 --> 00:47:34,520 So I don't know who's doing the fact checking but I do love I 874 00:47:34,520 --> 00:47:37,480 have a journal I use. I think it's reflection and I 875 00:47:37,480 --> 00:47:40,600 love it cuz it also generates AI like gets me deeper. 876 00:47:42,000 --> 00:47:44,640 You can have AI prompts that make you expand on certain 877 00:47:44,640 --> 00:47:48,040 things. But you say, I had a great time 878 00:47:48,040 --> 00:47:52,560 at the library today and it says actually your mood was like 4 879 00:47:52,560 --> 00:47:56,240 out of 10. It's more so like what made it 880 00:47:56,240 --> 00:47:59,240 good, like what would have improved it for the next time, 881 00:47:59,240 --> 00:48:01,840 you know? Yeah, and it brings us maybe to 882 00:48:01,840 --> 00:48:05,400 this one. So predictive nostalgia machine, 883 00:48:05,440 --> 00:48:09,640 basically an AI that projects what you'll be nostalgic about 884 00:48:09,680 --> 00:48:12,400 in the future based on your kind of previous experiences. 885 00:48:12,400 --> 00:48:15,480 So like it knows that you know you've had some experience in 886 00:48:15,480 --> 00:48:18,640 the past. So like it tries to kind of let 887 00:48:18,640 --> 00:48:21,200 you know, like, hey, you should really appreciate this moment 888 00:48:21,200 --> 00:48:23,920 because this is something you'll be nostalgic about in a few 889 00:48:23,920 --> 00:48:26,800 years time. Yeah, the predictive analytics, 890 00:48:26,800 --> 00:48:29,760 Sure, why not? Sounds like it's AI for 891 00:48:29,760 --> 00:48:31,640 mindfulness and I'm about that. I like it. 892 00:48:33,720 --> 00:48:38,440 OK, last one, this is the emotional soundtrack generator. 893 00:48:38,440 --> 00:48:42,240 So this is an AI, listens to your daily conversations, 894 00:48:42,240 --> 00:48:45,400 basically, you know, takes a peek at your life, generates a 895 00:48:45,400 --> 00:48:48,000 soundtrack that reflects your emotional state. 896 00:48:50,000 --> 00:48:53,840 Yes, but I think that I would want it to also generate 897 00:48:53,840 --> 00:48:56,040 soundtracks that Get Me Out of a funk. 898 00:48:57,200 --> 00:49:00,880 They can tell that I'm in one. Yes, good. 899 00:49:01,080 --> 00:49:02,680 There's a lot of soundtracks there, like on Spotify for 900 00:49:02,680 --> 00:49:05,800 example, where it's like you're sad or happy, but there's not 901 00:49:05,800 --> 00:49:08,680 really one from like from sad to excited. 902 00:49:08,680 --> 00:49:11,040 From sad to happiest. Yeah, I love that. 903 00:49:12,720 --> 00:49:14,000 Yeah. And my niece had just the 904 00:49:14,000 --> 00:49:15,680 funniest. I remember when she was young, 905 00:49:15,680 --> 00:49:17,480 at one time, I was like, why are you sad? 906 00:49:17,480 --> 00:49:19,960 She's like, well, the music is sad for that reason. 907 00:49:19,960 --> 00:49:21,560 I feel sad and I want to keep listening to it. 908 00:49:21,560 --> 00:49:23,800 And I was like, OK, yeah. So chicken or egg? 909 00:49:23,800 --> 00:49:24,480 I'm not sure. Right. 910 00:49:25,560 --> 00:49:28,880 And I think you can get us into a mood or sometimes get us out 911 00:49:28,880 --> 00:49:29,440 of 1. So. 912 00:49:29,440 --> 00:49:31,200 Yeah. Yeah, love that. 913 00:49:31,400 --> 00:49:33,680 OK. Well, you made it to the final 914 00:49:33,680 --> 00:49:37,600 question now, which is what is your most controversial opinion 915 00:49:37,680 --> 00:49:42,600 about AII? Would say that I think that 916 00:49:42,600 --> 00:49:47,720 clinicians can be better with AI than without. 917 00:49:48,040 --> 00:49:50,960 I think that's what I would say. And I think that I have been 918 00:49:50,960 --> 00:49:54,280 converted to that now, being able to see that AI can do 100% 919 00:49:54,280 --> 00:49:57,120 oversight. But I think also, yeah, having 920 00:49:57,120 --> 00:50:00,960 founded a clinic, being the slow psychologist for a minute there, 921 00:50:02,040 --> 00:50:05,280 I now see that AI can really help with supervision and yeah, 922 00:50:05,360 --> 00:50:08,920 and quality care in a way that humans, just because life 923 00:50:08,920 --> 00:50:11,920 happens, not just that we get tired, but sometimes we have 924 00:50:11,920 --> 00:50:15,280 tough moments in our lives that can compromise, like our 925 00:50:15,280 --> 00:50:18,680 precision. And so I think that AI can be a 926 00:50:18,680 --> 00:50:21,480 really wonderful tool, but I think it's a tool that that we 927 00:50:21,480 --> 00:50:25,320 need more than we think we do. So you're really let's embrace 928 00:50:25,480 --> 00:50:28,760 AI as collaborator. Let's embrace ethical, robust AI 929 00:50:28,760 --> 00:50:31,240 as collaborator. I love it. 930 00:50:31,600 --> 00:50:33,840 Well thank you, this was so much fun. 931 00:50:34,000 --> 00:50:34,840 Wonderful. All right. 932 00:50:34,840 --> 00:50:37,600 Well, thank you all so much. And that's a wrap. 933 00:50:37,920 --> 00:50:40,600 You've been listening to the Behavioral Design podcast 934 00:50:40,880 --> 00:50:43,440 brought to you by Habit Weekly and Nuanced Behavior. 935 00:50:43,800 --> 00:50:46,120 Sam and Alene tell me. This season is packed with 936 00:50:46,120 --> 00:50:49,960 incredible insights about behavioral design and AI, so be 937 00:50:49,960 --> 00:50:52,560 sure to subscribe and share the podcast with your friends, 938 00:50:52,800 --> 00:50:54,960 though you might want to keep it away from your enemies. 939 00:50:56,320 --> 00:50:58,920 In case you haven't noticed, I'm an AI voice. 940 00:51:00,200 --> 00:51:02,880 Yep, pretty crazy. Quite the improvement since last 941 00:51:02,880 --> 00:51:04,880 season's AI outro, don't you think? 942 00:51:06,120 --> 00:51:08,800 If you'd like to collaborate with us at Nuance Behavior, 943 00:51:09,000 --> 00:51:11,720 where we use behavioral design to craft digital products with 944 00:51:11,720 --> 00:51:16,040 Nuance, e-mail us at hello@nuancebehavior.com or book 945 00:51:16,040 --> 00:51:19,320 a call directly on our website, nuancebehavior.com. 946 00:51:20,760 --> 00:51:24,160 A special thanks to the amazing Dave Pizarro for our show music 947 00:51:24,400 --> 00:51:27,280 and to Mei Chen Yap and April English for their help in 948 00:51:27,280 --> 00:51:29,200 producing and publishing this episode. 949 00:51:29,640 --> 00:51:32,520 Thanks again for tuning in. We'll be back soon with another 950 00:51:32,520 --> 00:51:35,720 exciting conversation where behavioral design and AI 951 00:51:35,720 --> 00:51:38,920 Intersect happens to. Mugatroid.