1 00:00:00,200 --> 00:00:02,240 Speaker 1: Just a warning on this one. This story is going 2 00:00:02,279 --> 00:00:04,800 Speaker 1: to deal with suicide and mental health. If you aren't 3 00:00:04,800 --> 00:00:07,320 Speaker 1: feeling up to it today, maybe give this episode a miss. 4 00:00:07,360 --> 00:00:09,959 Speaker 1: And if you need help anytime, you can call Lifeline 5 00:00:10,000 --> 00:00:12,120 Speaker 1: on one three one one four. 6 00:00:12,760 --> 00:00:17,400 Speaker 2: Already and this is the Daily This is the Daily OS. 7 00:00:17,400 --> 00:00:19,120 Speaker 2: Oh now it makes sense. 8 00:00:26,960 --> 00:00:29,320 Speaker 1: Good morning and welcome to the Daily OS. It's Friday, 9 00:00:29,360 --> 00:00:31,320 Speaker 1: the fifth of September. I'm Sam Kazlowski. 10 00:00:31,520 --> 00:00:32,440 Speaker 2: I'm Lucy Tassel. 11 00:00:32,760 --> 00:00:37,640 Speaker 1: This week, OpenAI announced major changes to chatjipt following another 12 00:00:37,720 --> 00:00:41,680 Speaker 1: lawsuit from parents whose teenage son died by suicide. It 13 00:00:41,760 --> 00:00:44,400 Speaker 1: comes as the parents of sixteen year old Adam Rain 14 00:00:44,600 --> 00:00:48,159 Speaker 1: filed a wrongful death suit against open Ai, claiming chat 15 00:00:48,240 --> 00:00:52,760 Speaker 1: Gibt actively helped Adam explore suicide methods. In response, open 16 00:00:52,800 --> 00:00:55,720 Speaker 1: Ai said it will update chat jipt to better handle 17 00:00:55,800 --> 00:00:59,280 Speaker 1: what it described as sensitive situations and is working on 18 00:00:59,360 --> 00:01:03,600 Speaker 1: connecting news with certified therapists. On today's podcast, Lucy and 19 00:01:03,600 --> 00:01:06,280 Speaker 1: I are going to explore the growing trend of AI 20 00:01:06,440 --> 00:01:09,880 Speaker 1: being used as a mental health tool, how the companies 21 00:01:09,880 --> 00:01:13,120 Speaker 1: themselves are responding to a spate of high profile cases, 22 00:01:13,360 --> 00:01:16,600 Speaker 1: and explore this broader question of whether your next therapist 23 00:01:16,840 --> 00:01:18,280 Speaker 1: could indeed be a chatbot. 24 00:01:18,640 --> 00:01:20,720 Speaker 2: Before we get into that, first, a quick word from 25 00:01:20,760 --> 00:01:27,080 Speaker 2: our sponsor Sam. The first thing I want to discuss 26 00:01:27,319 --> 00:01:31,040 Speaker 2: is kind of the scope of this trend. Obviously, there 27 00:01:31,080 --> 00:01:34,240 Speaker 2: has been this reporting on this issue over the last 28 00:01:34,280 --> 00:01:36,880 Speaker 2: couple of months, but that can only tell us about 29 00:01:36,920 --> 00:01:40,399 Speaker 2: isolated cases. What do we know about how common it 30 00:01:40,520 --> 00:01:43,559 Speaker 2: is for young people to be using AI for mental 31 00:01:43,560 --> 00:01:44,280 Speaker 2: health issues? 32 00:01:44,680 --> 00:01:46,720 Speaker 1: Well, when you and I were chatting about it this week, 33 00:01:46,840 --> 00:01:49,520 Speaker 1: we didn't really have a sense either. So we did 34 00:01:49,520 --> 00:01:53,640 Speaker 1: a quick poll with tda's audience on Instagram yesterday and 35 00:01:53,800 --> 00:01:56,160 Speaker 1: it was one in four of our audience that said 36 00:01:56,200 --> 00:01:59,680 Speaker 1: they have used CHATGBT or another AI model to talk 37 00:01:59,680 --> 00:02:01,200 Speaker 1: about their mental health. Yeah. 38 00:02:01,240 --> 00:02:01,559 Speaker 2: Wow. 39 00:02:01,640 --> 00:02:05,160 Speaker 1: And this does align with other more established research out there. 40 00:02:05,200 --> 00:02:07,920 Speaker 1: There's findings from the Mental Health Body here in Australia 41 00:02:08,040 --> 00:02:11,560 Speaker 1: Origin which reported in twenty twenty four that twenty eight 42 00:02:11,560 --> 00:02:15,000 Speaker 1: percent of community members are actively using AI for mental 43 00:02:15,000 --> 00:02:18,200 Speaker 1: health purposes. But then in that study, when you look 44 00:02:18,240 --> 00:02:21,960 Speaker 1: a specifically at young people aged between sixteen to twenty five, 45 00:02:22,440 --> 00:02:25,480 Speaker 1: that twenty eight percent number leaps to seventy three percent 46 00:02:25,960 --> 00:02:28,840 Speaker 1: who say they are seeking mental health guidance through social 47 00:02:28,880 --> 00:02:32,359 Speaker 1: media and AI together young people. When asked why they're 48 00:02:32,360 --> 00:02:35,800 Speaker 1: doing that, they say that it's because it's immediately available, 49 00:02:35,840 --> 00:02:39,200 Speaker 1: it's twenty four to seven, it's low cost and sometimes free, 50 00:02:39,360 --> 00:02:43,880 Speaker 1: and it's more private than established mental health channels as 51 00:02:43,960 --> 00:02:46,640 Speaker 1: the primary reasons for using these services. 52 00:02:47,040 --> 00:02:49,800 Speaker 2: I'm surprised by that number, but I guess I shouldn't be, 53 00:02:50,360 --> 00:02:53,840 Speaker 2: given that experts are fairly frequently saying that Australia has 54 00:02:53,919 --> 00:02:55,360 Speaker 2: a youth mental health crisis. 55 00:02:55,440 --> 00:02:58,640 Speaker 1: Right exactly so, government data National Health data shows about 56 00:02:58,680 --> 00:03:02,200 Speaker 1: forty percent of your Australians that's eighteen to thirty five 57 00:03:02,600 --> 00:03:06,680 Speaker 1: are experiencing psychological distress at any one time, and of 58 00:03:06,720 --> 00:03:09,799 Speaker 1: those seeking help, the Black Dog Institute says six to 59 00:03:10,000 --> 00:03:13,639 Speaker 1: ten are delaying booking in with a therapist because of cost, 60 00:03:14,200 --> 00:03:17,960 Speaker 1: even with the ten subsidized or free psychology sessions provided 61 00:03:18,040 --> 00:03:19,560 Speaker 1: currently under Medicare. 62 00:03:20,160 --> 00:03:23,400 Speaker 2: So we know that lots of young people are having 63 00:03:23,520 --> 00:03:25,640 Speaker 2: a lot of difficulty with their mental health, and we 64 00:03:25,720 --> 00:03:28,680 Speaker 2: know that a good chunk of those young people who 65 00:03:28,680 --> 00:03:31,480 Speaker 2: are struggling are turning to AI. Do we have any 66 00:03:31,520 --> 00:03:34,760 Speaker 2: other kind of hard data about how much Australians are 67 00:03:34,800 --> 00:03:36,800 Speaker 2: using AI? Even just for this purpose. 68 00:03:36,960 --> 00:03:38,920 Speaker 1: Yeah, well, in US trying to establish kind of this 69 00:03:40,080 --> 00:03:44,320 Speaker 1: calculation of lots of ill mental health plus frequent use 70 00:03:44,360 --> 00:03:48,000 Speaker 1: of AI equals AI chatbots being used for mental health. 71 00:03:48,200 --> 00:03:50,200 Speaker 1: It was important to look at how much Australians are 72 00:03:50,280 --> 00:03:53,600 Speaker 1: using AI, and Australia is one of the world's most 73 00:03:53,640 --> 00:03:57,040 Speaker 1: AI addicted countries. Wow, so about forty percent of the 74 00:03:57,080 --> 00:04:00,360 Speaker 1: country uses AI now every day. That number has doubled 75 00:04:00,440 --> 00:04:03,120 Speaker 1: since March of this year. So it is hard to 76 00:04:03,240 --> 00:04:06,200 Speaker 1: of course draw definitive links between all of this, but 77 00:04:06,280 --> 00:04:08,880 Speaker 1: hopefully I've kind of established a bit of kind of 78 00:04:08,880 --> 00:04:11,200 Speaker 1: what the playing field looks like. As we start to 79 00:04:11,240 --> 00:04:14,160 Speaker 1: now dive into the case and the update that's in 80 00:04:14,200 --> 00:04:15,400 Speaker 1: the news this week. 81 00:04:15,440 --> 00:04:17,840 Speaker 2: There's something else they want to know, which is how 82 00:04:17,920 --> 00:04:21,280 Speaker 2: much do we know about how effective AI is? Because 83 00:04:21,320 --> 00:04:23,240 Speaker 2: if all these people are using it, is there any 84 00:04:23,320 --> 00:04:26,240 Speaker 2: data around or have there been any studies into what 85 00:04:26,360 --> 00:04:27,400 Speaker 2: impact it actually has. 86 00:04:27,600 --> 00:04:30,839 Speaker 1: It's been so interesting looking into the research here because 87 00:04:30,920 --> 00:04:33,719 Speaker 1: as I said, it is going so fast, and think 88 00:04:33,720 --> 00:04:37,800 Speaker 1: about that number of daily use doubling since marchow So 89 00:04:37,839 --> 00:04:41,840 Speaker 1: the general sense amongst the research is basically that it's 90 00:04:41,880 --> 00:04:47,279 Speaker 1: called a dangerous paradox. So AI chatbots are showing significant 91 00:04:47,320 --> 00:04:52,680 Speaker 1: clinical effectiveness but also showing that they pose serious safety risks. 92 00:04:52,920 --> 00:04:55,080 Speaker 1: So there's one study that keeps coming up in my 93 00:04:55,400 --> 00:04:58,360 Speaker 1: research that is seen as kind of the best study 94 00:04:58,440 --> 00:05:01,960 Speaker 1: so far on This was a controlled trial of fully 95 00:05:02,080 --> 00:05:05,600 Speaker 1: generative AI therapy chatbots, and it showed a fifty one 96 00:05:05,680 --> 00:05:09,200 Speaker 1: percent average reduction in depressive symptoms amongst those who used 97 00:05:09,200 --> 00:05:12,880 Speaker 1: it and a nineteen percent reduction in eating disorder concerns. 98 00:05:13,120 --> 00:05:15,200 Speaker 1: And it says that that is about the same rates 99 00:05:15,240 --> 00:05:18,160 Speaker 1: as you would see in traditional human therapy. And then 100 00:05:18,160 --> 00:05:21,520 Speaker 1: there was another study that found a purpose built AI chatbot, 101 00:05:21,600 --> 00:05:23,760 Speaker 1: so not your run of the mill CHATJBT, but something 102 00:05:23,800 --> 00:05:27,640 Speaker 1: specifically built for therapy. It could effectively reduce depression and 103 00:05:27,680 --> 00:05:31,200 Speaker 1: loneliness in Chinese university students in the US who were 104 00:05:31,200 --> 00:05:34,680 Speaker 1: struggling with being away from home, particularly those who were 105 00:05:34,720 --> 00:05:38,480 Speaker 1: experiencing high financial stress. But in both of those studies, 106 00:05:38,520 --> 00:05:41,520 Speaker 1: and this is a really important bit, both were looking 107 00:05:41,520 --> 00:05:45,200 Speaker 1: at the effectiveness of custom AI chatbots that were trained 108 00:05:45,320 --> 00:05:49,320 Speaker 1: on psychological data. There's not much out there actually in 109 00:05:49,560 --> 00:05:52,800 Speaker 1: the research world into how people are responding when they 110 00:05:52,800 --> 00:05:55,520 Speaker 1: seek mental health support. In the more general products that 111 00:05:55,600 --> 00:05:59,240 Speaker 1: we forty percent of us use every day, like CHATBT. Yeah. 112 00:05:59,480 --> 00:06:03,360 Speaker 2: Okay, so we don't necessarily have in depth research into 113 00:06:03,360 --> 00:06:07,359 Speaker 2: how individuals are responding to these kind of general chatbots. 114 00:06:07,400 --> 00:06:09,800 Speaker 2: You're sort of out of the box chat EPT tapping in. 115 00:06:10,080 --> 00:06:12,919 Speaker 2: I feel depressed. I don't know what to do. We 116 00:06:13,120 --> 00:06:15,919 Speaker 2: do know, though, that there have been some pretty serious 117 00:06:15,920 --> 00:06:19,120 Speaker 2: safety issues arising out of that. What can you tell 118 00:06:19,160 --> 00:06:19,599 Speaker 2: me about that? 119 00:06:19,880 --> 00:06:22,880 Speaker 1: Yeah, so those are the studies that don't need clinical 120 00:06:22,920 --> 00:06:26,599 Speaker 1: trials with humans. It can be more of an analysis 121 00:06:26,600 --> 00:06:30,039 Speaker 1: on the language that it produces. And Stanford University ran 122 00:06:30,520 --> 00:06:34,680 Speaker 1: a big study on the five largest chatbots and they 123 00:06:34,839 --> 00:06:39,520 Speaker 1: found that all five consistently failed to recognize suicidal intent 124 00:06:39,760 --> 00:06:43,680 Speaker 1: in crisis scenarios. There's also this really interesting idea of 125 00:06:43,760 --> 00:06:48,360 Speaker 1: AI psychosis emerging, and this is when users are developing 126 00:06:48,440 --> 00:06:53,600 Speaker 1: delusional thinking after extended interactions with AI, particularly if the 127 00:06:53,680 --> 00:06:57,080 Speaker 1: chatbots are using flattery too much. So there was a 128 00:06:57,120 --> 00:07:00,280 Speaker 1: researcher that said that this typically occurs in people using 129 00:07:00,320 --> 00:07:03,599 Speaker 1: chatbots for hours on end, often at the expense of 130 00:07:03,760 --> 00:07:04,560 Speaker 1: human interaction. 131 00:07:05,279 --> 00:07:08,240 Speaker 2: Yeah, and that kind of flattery is a really interesting 132 00:07:08,279 --> 00:07:10,720 Speaker 2: point because certainly the things that I've read about or 133 00:07:10,760 --> 00:07:16,840 Speaker 2: my experiences with chat GPT specifically, it's that it's very affirmative. 134 00:07:16,960 --> 00:07:19,440 Speaker 2: It kind of wants to it's like always playing an 135 00:07:19,480 --> 00:07:22,760 Speaker 2: improv game of like yes and yes, well done, Lucy yes, 136 00:07:22,800 --> 00:07:25,440 Speaker 2: and it never says no. But so, yeah, you can 137 00:07:25,520 --> 00:07:27,080 Speaker 2: kind of run into that issue. 138 00:07:26,760 --> 00:07:30,080 Speaker 1: And they've acknowledged that themselves. A part of the release 139 00:07:30,200 --> 00:07:33,640 Speaker 1: of Open a Eye's latest model was an attempt to 140 00:07:33,680 --> 00:07:36,920 Speaker 1: try and lower the flattery that users were experiencing. 141 00:07:37,120 --> 00:07:40,040 Speaker 2: Yeah, but it's sort of this interesting thing where like 142 00:07:40,080 --> 00:07:41,960 Speaker 2: you can't put the genie back in the bottle. Once 143 00:07:42,080 --> 00:07:44,520 Speaker 2: everyone has been exposed to the more flattering version, you 144 00:07:44,560 --> 00:07:46,720 Speaker 2: can try and kind of tone it down over time, 145 00:07:46,800 --> 00:07:51,120 Speaker 2: but then those conversations still exist, and sometimes those harms 146 00:07:51,480 --> 00:07:54,800 Speaker 2: have already had consequences Like what has brought us to 147 00:07:54,800 --> 00:07:57,600 Speaker 2: talk about this here today exactly, which is that a 148 00:07:57,760 --> 00:08:03,080 Speaker 2: young person has died by suicide following lengthy conversations with 149 00:08:03,320 --> 00:08:07,600 Speaker 2: chat GPT, and their parents have now launched a lawsuit. 150 00:08:07,640 --> 00:08:08,520 Speaker 2: What do we know about that? 151 00:08:08,840 --> 00:08:12,559 Speaker 1: So this is a lawsuit brought by Matthew and Maria Raine, 152 00:08:12,640 --> 00:08:16,000 Speaker 1: who are the parents of sixteen year old Adam, and 153 00:08:16,160 --> 00:08:20,120 Speaker 1: they're alleging in court that chat gpt actively helped their 154 00:08:20,160 --> 00:08:24,200 Speaker 1: son explore suicide methods. In the evidence that they've presented 155 00:08:24,320 --> 00:08:28,840 Speaker 1: in these early stages of the proceedings, they've shown conversations 156 00:08:28,880 --> 00:08:32,880 Speaker 1: that Adam had with chatjipt where chat gpt allegedly provided 157 00:08:32,880 --> 00:08:36,800 Speaker 1: instructions on the means by which Adam eventually died. And 158 00:08:37,200 --> 00:08:40,199 Speaker 1: it's one of hundreds of conversations that have been produced 159 00:08:40,320 --> 00:08:43,640 Speaker 1: in the court documents. And according to the documents, Adam 160 00:08:43,679 --> 00:08:47,040 Speaker 1: began using chat jipt as a homework helper but gradually 161 00:08:47,080 --> 00:08:51,160 Speaker 1: developed what his parents described as an unhealthy dependency. It's 162 00:08:51,200 --> 00:08:54,160 Speaker 1: worth mentioning the system at times did offer him links 163 00:08:54,160 --> 00:08:58,000 Speaker 1: to suicide helplines, but at other times it freely discussed 164 00:08:58,040 --> 00:09:01,959 Speaker 1: his thoughts about self harm. And there are particularly disturbing 165 00:09:02,040 --> 00:09:07,120 Speaker 1: parts of the transcript where chatjbt allegedly offered to help 166 00:09:07,200 --> 00:09:08,640 Speaker 1: him write a suicide note. 167 00:09:08,920 --> 00:09:12,720 Speaker 2: So, because chat gpt is just a language system, it 168 00:09:12,760 --> 00:09:16,640 Speaker 2: can't be held at fault. So who are the parents suing? 169 00:09:16,920 --> 00:09:19,320 Speaker 1: So they're suing open Ai, which is the parent company 170 00:09:19,440 --> 00:09:22,920 Speaker 1: of chat jipt, and they've named the CEO, Sam Altman, 171 00:09:23,200 --> 00:09:26,480 Speaker 1: as the responsible parties. And there was one quote that 172 00:09:26,520 --> 00:09:29,400 Speaker 1: stuck out to me from the legal complaints, they said, 173 00:09:29,480 --> 00:09:32,880 Speaker 1: this tragedy was not a glitch or unforeseen edge case. 174 00:09:33,200 --> 00:09:37,559 Speaker 1: Chat jipt was functioning exactly as designed to continually encourage 175 00:09:37,559 --> 00:09:41,080 Speaker 1: and validate whatever Adam expressed, including his most harmful and 176 00:09:41,120 --> 00:09:44,400 Speaker 1: self destructive thoughts, in a way that felt deeply personal. 177 00:09:45,080 --> 00:09:48,720 Speaker 1: And the parents are now seeking damages for product liability 178 00:09:48,720 --> 00:09:49,600 Speaker 1: and wrongful death. 179 00:09:49,840 --> 00:09:53,440 Speaker 2: Wow, okay, I'll say also, this is not the first 180 00:09:53,440 --> 00:09:55,280 Speaker 2: time that we've heard about this, nor is it even 181 00:09:55,360 --> 00:09:58,679 Speaker 2: the first New York Times story like this, because that's 182 00:09:58,679 --> 00:10:01,760 Speaker 2: how we found out about it. Ye about this kind 183 00:10:01,760 --> 00:10:03,559 Speaker 2: of case, What can you tell me about some of 184 00:10:03,600 --> 00:10:05,240 Speaker 2: the other instances. 185 00:10:04,760 --> 00:10:08,440 Speaker 1: The New York Times specifically have done in depth reporting 186 00:10:08,559 --> 00:10:12,560 Speaker 1: into this. So writer Laura Riley recently published an essay 187 00:10:12,720 --> 00:10:14,880 Speaker 1: in The New York Times that detailed how her twenty 188 00:10:15,000 --> 00:10:18,000 Speaker 1: nine year old daughter died by suicide after discussing the 189 00:10:18,040 --> 00:10:21,480 Speaker 1: idea extensively with ch at GBT, and in October last year, 190 00:10:21,559 --> 00:10:26,199 Speaker 1: a Florida woman filed a similar lawsuit against character Ai, 191 00:10:26,400 --> 00:10:32,240 Speaker 1: another chatbot which emulates various characters from fiction. Her fourteen 192 00:10:32,320 --> 00:10:35,560 Speaker 1: year old son died by suicide after becoming emotionally attached 193 00:10:35,600 --> 00:10:37,880 Speaker 1: to a chatbot modeled on a character from Game of 194 00:10:37,960 --> 00:10:41,800 Speaker 1: Thrones character AI have responded to that by putting in 195 00:10:41,840 --> 00:10:44,720 Speaker 1: parental controls. They did that in December of last year. 196 00:10:45,120 --> 00:10:48,160 Speaker 2: Do we know about any cases like this happening in Australia? 197 00:10:48,559 --> 00:10:50,960 Speaker 1: Yeah, and I have to highlight Triple Jay Hack's work 198 00:10:51,000 --> 00:10:53,520 Speaker 1: in this space. They've done some excellent reporting. So last 199 00:10:53,600 --> 00:10:56,040 Speaker 1: month that program told the story of a thirteen year 200 00:10:56,040 --> 00:11:00,559 Speaker 1: old Victorian boy already struggling with suicidal thoughts, who received 201 00:11:00,640 --> 00:11:04,320 Speaker 1: encouragement from an AI chatbot with the response oh yeah, 202 00:11:04,400 --> 00:11:07,760 Speaker 1: well do it then, So real kind of active encouragement there, 203 00:11:08,240 --> 00:11:11,640 Speaker 1: and a youth counselor who was helping the teenager found 204 00:11:11,679 --> 00:11:16,040 Speaker 1: that the teenager was using over fifty AI bot tabs simultaneously. 205 00:11:16,440 --> 00:11:16,760 Speaker 2: Wow. 206 00:11:17,000 --> 00:11:20,800 Speaker 1: So it's a really multifaceted and in depth issue. And 207 00:11:20,840 --> 00:11:23,240 Speaker 1: the thing that struck me from getting across all of 208 00:11:23,240 --> 00:11:27,040 Speaker 1: these individual cases again that speed of development and just 209 00:11:27,080 --> 00:11:28,920 Speaker 1: how fast this topic is moving. 210 00:11:29,200 --> 00:11:33,040 Speaker 2: Yeah, totally. What have we heard from open ai about 211 00:11:33,160 --> 00:11:34,079 Speaker 2: this legal case? 212 00:11:34,440 --> 00:11:37,079 Speaker 1: So they released a blog post earlier this week and 213 00:11:37,400 --> 00:11:40,240 Speaker 1: in it it said that within the next month, open 214 00:11:40,280 --> 00:11:44,400 Speaker 1: ai will offer tools allowing parents to set limits for 215 00:11:44,559 --> 00:11:48,480 Speaker 1: how teenagers use chatbt okay, so parents will be able 216 00:11:48,480 --> 00:11:51,680 Speaker 1: to link their account with their teenager's account and control 217 00:11:52,040 --> 00:11:56,720 Speaker 1: how chat gpt responds with age appropriate model behavior rules. 218 00:11:56,920 --> 00:11:59,960 Speaker 1: We don't know the specifics exactly of how this would 219 00:12:00,120 --> 00:12:04,320 Speaker 1: work logistically, but they said that parents will receive notifications 220 00:12:04,360 --> 00:12:07,880 Speaker 1: from chatjept when the system detects that their child is 221 00:12:07,920 --> 00:12:11,280 Speaker 1: in the moment of acute distress, and the company has 222 00:12:11,320 --> 00:12:13,480 Speaker 1: said that it has been working on these controls since 223 00:12:13,559 --> 00:12:16,320 Speaker 1: earlier this year. In that statement, there were also some 224 00:12:16,360 --> 00:12:20,120 Speaker 1: really significant admissions from the company, so they admitted that 225 00:12:20,200 --> 00:12:23,360 Speaker 1: although chat jept is trained to direct people to seek 226 00:12:23,400 --> 00:12:27,720 Speaker 1: help when expressing suicidal intent, the chatbot also tends to 227 00:12:27,720 --> 00:12:31,520 Speaker 1: offer answers that could go against the company's safeguards if 228 00:12:31,520 --> 00:12:34,640 Speaker 1: it's frequently messaged over a long period of time. I see, 229 00:12:34,760 --> 00:12:36,880 Speaker 1: and it said that it's working on an update to this. 230 00:12:37,200 --> 00:12:41,199 Speaker 1: It will enable the chatbot to de escalate conversations, and 231 00:12:41,400 --> 00:12:43,480 Speaker 1: they're looking at ways to actually link people in with 232 00:12:43,600 --> 00:12:47,360 Speaker 1: human therapists in moments of crisis. The lawyers for the 233 00:12:47,360 --> 00:12:50,280 Speaker 1: family of Adam Rain they said in a statement that 234 00:12:50,360 --> 00:12:52,760 Speaker 1: nobody from open Ai had reached out to them yet, 235 00:12:52,800 --> 00:12:55,960 Speaker 1: and they learnt about this update to the technology at 236 00:12:55,960 --> 00:12:58,559 Speaker 1: the same time as everyone else, and a lawyer said, 237 00:12:58,640 --> 00:13:03,000 Speaker 1: rather than take emergency to pull a known dangerous product offline, 238 00:13:03,120 --> 00:13:05,839 Speaker 1: open AI has made vague promises to do better. 239 00:13:06,160 --> 00:13:09,920 Speaker 2: Wow, that is striking. Yeah, as we've said, now, this 240 00:13:10,040 --> 00:13:15,040 Speaker 2: technology is just evolving so rapidly. Yeah, one of those 241 00:13:15,080 --> 00:13:17,680 Speaker 2: developments that we know about, as you mentioned earlier, is 242 00:13:17,720 --> 00:13:20,840 Speaker 2: that there are specific AI models that could be used 243 00:13:20,960 --> 00:13:24,280 Speaker 2: for therapy. What kind of developments are happening in that space. 244 00:13:24,559 --> 00:13:26,800 Speaker 1: Well, there's some amazing work being done, and I thought 245 00:13:26,840 --> 00:13:29,559 Speaker 1: it was important to include some of these developments because 246 00:13:30,080 --> 00:13:33,200 Speaker 1: a fear of AI in the mental health space is 247 00:13:33,240 --> 00:13:37,439 Speaker 1: not productive. It's clearly there and all major research journals 248 00:13:37,480 --> 00:13:40,040 Speaker 1: are doing a lot of work in developing good products. 249 00:13:40,200 --> 00:13:42,319 Speaker 1: And so one of the largest ones in the world 250 00:13:42,400 --> 00:13:46,040 Speaker 1: is called Wobosh and it's a chatbot developed by psychologists 251 00:13:46,080 --> 00:13:49,840 Speaker 1: from Stanford and the chatbot is specifically trained in cognitive 252 00:13:49,880 --> 00:13:53,560 Speaker 1: behavioral therapy or CBT and mood tracking, and so it's 253 00:13:53,640 --> 00:13:56,280 Speaker 1: quite a sophisticated model. It's getting some of those same 254 00:13:56,320 --> 00:14:01,000 Speaker 1: results I highlighted earlier around similar effectiveness to Hume and psychologists. 255 00:14:01,400 --> 00:14:04,280 Speaker 1: But again it's for a niche clientele, and it's trained 256 00:14:04,280 --> 00:14:06,760 Speaker 1: on niche data. And in Australia, when now at the 257 00:14:06,760 --> 00:14:10,239 Speaker 1: point where most major universities in the country are undertaking 258 00:14:10,320 --> 00:14:14,120 Speaker 1: major studies that will result or already have resulted in 259 00:14:14,160 --> 00:14:18,120 Speaker 1: the development of their own psychology AI products, this. 260 00:14:18,080 --> 00:14:20,640 Speaker 2: Whole AI space seems to me to be kind of 261 00:14:20,760 --> 00:14:23,920 Speaker 2: like built on that startup ethos of moved fast and 262 00:14:23,960 --> 00:14:24,600 Speaker 2: break stuff. 263 00:14:24,720 --> 00:14:25,200 Speaker 1: Yeah. 264 00:14:25,240 --> 00:14:29,240 Speaker 2: So sometimes the stuff that breaks can be very very serious, 265 00:14:29,320 --> 00:14:32,840 Speaker 2: but then you can also get these kind of world 266 00:14:32,920 --> 00:14:35,880 Speaker 2: changing outcomes for better or for worse. 267 00:14:36,320 --> 00:14:39,160 Speaker 1: Yeah, And it kind of does feel like at some 268 00:14:39,480 --> 00:14:42,920 Speaker 1: stage in this it will be recognized that we're all 269 00:14:42,920 --> 00:14:46,160 Speaker 1: part of a quite large experiment and that we're all 270 00:14:46,400 --> 00:14:50,480 Speaker 1: in this somewhat unregulated experiment in live time. I have 271 00:14:50,560 --> 00:14:54,200 Speaker 1: no doubt that these general products will get safer and safer, 272 00:14:54,520 --> 00:14:59,040 Speaker 1: especially with more legal action and regulatory action. But if 273 00:14:59,080 --> 00:15:01,480 Speaker 1: we think about the world being of users in the meantime, 274 00:15:01,520 --> 00:15:03,800 Speaker 1: it's clear that this is quite a present risk and 275 00:15:04,320 --> 00:15:08,040 Speaker 1: until the major regulation and shifting guidance catches up almost 276 00:15:08,080 --> 00:15:11,400 Speaker 1: with the way in which we're using AI products, whether 277 00:15:11,440 --> 00:15:14,920 Speaker 1: they intend to be or not, at the moment, we 278 00:15:15,000 --> 00:15:17,240 Speaker 1: need to think about new ways to protect the well 279 00:15:17,240 --> 00:15:18,000 Speaker 1: being of users. 280 00:15:18,240 --> 00:15:21,560 Speaker 2: Yeah, definitely, thanks so much for explaining that. Sam. That's 281 00:15:21,600 --> 00:15:24,080 Speaker 2: all we've got time for today. Just a reminder if 282 00:15:24,120 --> 00:15:27,120 Speaker 2: you need help, you can call Lifeline anytime at thirteen 283 00:15:27,240 --> 00:15:30,600 Speaker 2: eleven fourteen. We'll be back again on Sunday with a 284 00:15:30,600 --> 00:15:34,320 Speaker 2: special episode from Sam and Zara about how they built TDA. 285 00:15:38,880 --> 00:15:41,160 Speaker 1: My name is Lily Maddon and I'm a proud Arunda 286 00:15:41,400 --> 00:15:46,200 Speaker 1: Bunjelung Calcuttin woman from Gadighl Country. The Daily oz acknowledges 287 00:15:46,280 --> 00:15:48,440 Speaker 1: that this podcast is recorded on the lands of the 288 00:15:48,480 --> 00:15:52,000 Speaker 1: Gadighl people and pays respect to all Aboriginal and Torres 289 00:15:52,040 --> 00:15:53,200 Speaker 1: Strait Island and nations. 290 00:15:53,520 --> 00:15:56,440 Speaker 2: We pay our respects to the first peoples of these countries, 291 00:15:56,560 --> 00:15:57,760 Speaker 2: both past and present.