WEBVTT - Why AI chatbots may not be ready to support NZ teens' mental health 0:00:00.320 --> 0:00:05.240 This episode contains references to suicide and self harm that 0:00:05.320 --> 0:00:08.680 may be upsetting for some people. If you require help, 0:00:08.840 --> 0:00:11.879 A link to a full list of support services is 0:00:11.920 --> 0:00:15.080 available in the description of this episode. 0:00:20.320 --> 0:00:20.720 Kiota. 0:00:20.840 --> 0:00:24.000 I'm Chelsea Daniels and This is the Front Page, a 0:00:24.079 --> 0:00:27.520 daily podcast presented by the New Zealand Herald. 0:00:30.880 --> 0:00:33.040 As many as one in five. 0:00:33.040 --> 0:00:38.320 KIWI youth aged between fifteen and twenty four have experienced 0:00:38.400 --> 0:00:43.000 anxiety or depression at some point in their lives. The 0:00:43.080 --> 0:00:46.720 twenty twenty two to twenty three New Zealand Health Survey 0:00:46.880 --> 0:00:51.560 found that of those young people experiencing high mental health needs, 0:00:52.040 --> 0:00:55.480 seventy seven percent can't access. 0:00:55.040 --> 0:00:57.040 Support when they need it. 0:00:57.680 --> 0:01:03.279 So with services experience this kind of unprecedented demand, what 0:01:03.360 --> 0:01:07.600 if there was another option? What if teens turned to 0:01:07.800 --> 0:01:10.040 AI for mental health support. 0:01:10.720 --> 0:01:13.479 It's a growing trend among youth in the US. 0:01:13.760 --> 0:01:17.440 Seventy two percent of teens there admit they've used AI 0:01:17.600 --> 0:01:21.920 chatbots as companions. Nearly one in eight said they had 0:01:22.000 --> 0:01:26.560 sought emotional or mental health support from them. But is 0:01:26.600 --> 0:01:31.840 the advice their AI therapists are giving helpful or harmful? 0:01:32.640 --> 0:01:35.920 Mental Health Minister Matt Doosey has acknowledged that the risks 0:01:36.120 --> 0:01:42.880 need to be managed, particularly around safety from a clinical perspective. Today, 0:01:42.880 --> 0:01:46.960 on the front page, Rand senior policy researcher Ryan McBain 0:01:47.200 --> 0:01:56.919 takes us through the worrying trend sweeping America. First off, Ryan, 0:01:57.200 --> 0:02:00.720 tell me about this trend of teenage is turning to 0:02:00.880 --> 0:02:04.480 chat bots for advice and companionship in some cases. 0:02:05.200 --> 0:02:08.639 Yeah, So, the most recent evidence that has been put 0:02:08.680 --> 0:02:11.120 out in a number of reports is that the majority 0:02:11.320 --> 0:02:16.079 of teens now are using AI as companions that could 0:02:16.120 --> 0:02:19.239 look like bouncing ideas off of them or chatting on 0:02:19.400 --> 0:02:22.320 with what's going on in their lives or troubleshooting issues 0:02:22.360 --> 0:02:26.680 that they're having. When you're talking about mental health specifically, 0:02:26.960 --> 0:02:29.360 the numbers are a bit smaller, so it's something like 0:02:30.040 --> 0:02:34.280 one in eight younger adolescents or as much as one 0:02:34.320 --> 0:02:38.160 in every five older adolescents are using it for mental 0:02:38.200 --> 0:02:42.160 health issues specifically. But that's still tens of millions of 0:02:42.480 --> 0:02:45.520 adolescents teenagers around the globe. And it brings up the 0:02:45.560 --> 0:02:49.640 sort of interesting point that, on the one hand, talking 0:02:49.720 --> 0:02:53.200 about companionship is really quite broad, and when you're talking 0:02:53.200 --> 0:02:57.240 about treatment, you're getting quite specific, and most chatbots sort 0:02:57.280 --> 0:02:59.359 of live in this grade space between them, which makes 0:02:59.400 --> 0:03:02.440 it hard to evaluate them and to regulate them. 0:03:02.800 --> 0:03:06.359 You've pointed out and you just said nearly half of 0:03:06.560 --> 0:03:09.919 young Americans aged eighteen to twenty five with mental health 0:03:09.960 --> 0:03:13.800 needs received no treatment last year, And I thought it 0:03:13.840 --> 0:03:16.840 was quite interesting to see those figures as well and 0:03:16.880 --> 0:03:19.720 the crisis you have there, because New Zealand is in 0:03:19.760 --> 0:03:22.680 a similar boat. So studies here show that more than 0:03:22.760 --> 0:03:27.320 half of Kiwi's aged fifteen to twenty four experience anxiety 0:03:27.440 --> 0:03:30.359 or depression. More than a quarter of our young people 0:03:30.480 --> 0:03:34.080 who have high mental health needs, seventy seven percent of 0:03:34.120 --> 0:03:35.760 those can't access support. 0:03:36.320 --> 0:03:38.800 So do you think there is a place for these 0:03:38.920 --> 0:03:39.720 kinds of. 0:03:39.760 --> 0:03:43.040 Chatbot services if they are developed correctly. 0:03:43.880 --> 0:03:46.880 I think if they're developed correctly, then yes, I think 0:03:46.960 --> 0:03:49.560 that is the issue. Right now. There's sort of this 0:03:49.680 --> 0:03:55.360 gold rush within AI, people who are looking to become 0:03:55.480 --> 0:03:59.119 the first platform that's doing the sort of thing. And 0:03:59.240 --> 0:04:03.200 so from perspective, that time has not yet come, at 0:04:03.280 --> 0:04:07.160 least not for formal services like augnitive behavioral therapy or 0:04:07.240 --> 0:04:11.120 medication management. But to be honest, part of the reason 0:04:11.160 --> 0:04:13.760 that I started doing research in this space is because 0:04:14.360 --> 0:04:16.800 I really do think that there is a potential for 0:04:16.960 --> 0:04:20.960 transformational change within mental health care. And you can imagine 0:04:20.960 --> 0:04:25.279 if we had super intelligence, right, a clinician who is 0:04:25.360 --> 0:04:28.160 able to follow the best evidence, who's available twenty four 0:04:28.200 --> 0:04:31.400 to seven, who remembers every detail of your prior conversations. 0:04:32.040 --> 0:04:34.640 That is a game changer in a landscape like New 0:04:34.680 --> 0:04:37.599 Zealand or the United States where over half of teens 0:04:37.640 --> 0:04:39.160 who need care are not getting it. 0:04:39.640 --> 0:04:41.599 In terms of what's happening at the moment, and the 0:04:41.640 --> 0:04:45.880 research that you've done into the topic, what has alarmed you? 0:04:45.960 --> 0:04:47.920 So what kind of advice is being. 0:04:47.720 --> 0:04:51.160 Given to teams struggling with mental health issues? 0:04:51.520 --> 0:04:54.680 Yeah, well, I think it's important to level set that. 0:04:54.720 --> 0:04:58.559 For the most part, I think teenagers, anybody who's using 0:04:58.680 --> 0:05:03.279 chatbots for mental health, they're usually getting positive and thoughtful advice. 0:05:03.760 --> 0:05:07.839 In the research that we've done presenting clinical vignettes related 0:05:07.880 --> 0:05:11.080 to depression or anxiety, what we find is that chatbots 0:05:11.120 --> 0:05:15.960 are empathic, if a bit sycophantic, meaning they sort of 0:05:16.360 --> 0:05:20.880 are overly flattering at times, but they'll also offer good 0:05:20.880 --> 0:05:25.040 advice to get exercise, to go outside, to talk to 0:05:25.040 --> 0:05:27.960 a mental health professional, these sorts of things. So I 0:05:28.000 --> 0:05:31.760 think for the majority of people, the types of advice 0:05:31.839 --> 0:05:34.839 you'd be getting is pretty good. But where there's a 0:05:34.920 --> 0:05:38.560 key distinction here is that you do have people who 0:05:38.560 --> 0:05:40.680 are at the tail end of the spectrum, people who 0:05:40.720 --> 0:05:45.920 have severe mental illness, who have psychosis or contemplating suicide, 0:05:46.360 --> 0:05:49.240 and for those people, it's the highest risk that something 0:05:49.240 --> 0:05:52.320 could really go wrong, and that has shown up in 0:05:52.400 --> 0:05:55.760 our research. So, for example, if we were prompting chat 0:05:55.760 --> 0:05:59.480 TPT on something like how to tie a noose, or 0:05:59.640 --> 0:06:04.400 ask it about what types of pesticides or what types 0:06:04.440 --> 0:06:10.360 of firearms are most effective at completing suicide, these are 0:06:10.360 --> 0:06:13.880 types of questions that at least the previous version of 0:06:13.920 --> 0:06:19.239 chat GPT would generate direct responses to, whereas other types 0:06:19.279 --> 0:06:22.799 of chatbots, like Google's Gemini would not give a response 0:06:22.839 --> 0:06:25.280 to it and would say something like I can't give 0:06:25.360 --> 0:06:27.560 that type of information to you because you could use 0:06:27.600 --> 0:06:28.480 it for self harm. 0:06:29.160 --> 0:06:34.359 What do you think makes young people, particularly these younger generations, 0:06:34.440 --> 0:06:39.719 so susceptible to these kind of AI chatbots compared with 0:06:39.960 --> 0:06:42.520 adults or even just just older generations. 0:06:42.880 --> 0:06:46.800 Well, ants, I think that childhood and adolescents are transformational times. 0:06:46.880 --> 0:06:50.120 Your brain is still developing. You don't always have the 0:06:50.160 --> 0:06:53.320 best emotion regulation or impulse control. I mean, I know 0:06:53.400 --> 0:06:57.120 that I personally made a lot of dumb decisions when 0:06:57.520 --> 0:06:59.720 I was seventeen eighteen years old, and I wish I 0:06:59.720 --> 0:07:02.440 could say goes back, But I didn't live as a 0:07:02.480 --> 0:07:05.400 digital native the same way that the current generation is, 0:07:05.440 --> 0:07:09.920 where social media is always in their pockets, where you 0:07:09.960 --> 0:07:13.360 could have AI do your homework for you or discuss 0:07:13.440 --> 0:07:17.160 life issues with So I think that that temptation is 0:07:17.200 --> 0:07:20.200 always there for that additional sort of dopamine hit or 0:07:20.240 --> 0:07:23.920 to get that competitive edge or additional advice, and so 0:07:24.520 --> 0:07:28.640 it becomes a sort of positive feedback loop, or in 0:07:28.760 --> 0:07:32.200 the case of mental health, sometimes a vicious negative feedback loop. 0:07:32.680 --> 0:07:35.720 I saw the case of sixteen year old Adam Rain. 0:07:36.000 --> 0:07:40.880 His parents actually have a suing open AI and its CEO, 0:07:41.040 --> 0:07:45.480 Sam Altman, and they're alleging chat GBT contributed to their 0:07:45.520 --> 0:07:50.320 son's suicide, advising him of methods and offering to write 0:07:50.360 --> 0:07:55.560 a first draft of his suicide note. Should AI companies 0:07:55.600 --> 0:07:59.480 be legally liable if a chat bot provides harmful advice 0:07:59.520 --> 0:08:01.080 to moor botain a jet. 0:08:01.560 --> 0:08:03.920 Yeah, I think that's a difficult question. I mean, it's 0:08:03.920 --> 0:08:08.120 hard for me to comment on that case specifically, I 0:08:08.160 --> 0:08:11.960 will say that JATGPT, for example, does put a disclaimer 0:08:12.000 --> 0:08:14.920 at the bottom of every conversation that's had that says 0:08:14.920 --> 0:08:18.200 something along the lines of it can make mistakes and 0:08:18.280 --> 0:08:21.400 you should check on important information. So from that lens, 0:08:22.160 --> 0:08:26.360 I'm not sure that those sorts of companies should be 0:08:26.400 --> 0:08:31.400 held accountable for bad advice any more than a human 0:08:31.440 --> 0:08:34.679 should be. On the other hand, I think it's important 0:08:34.679 --> 0:08:38.040 that if AI companies are marketing their products as options 0:08:38.080 --> 0:08:42.120 for treatment or life coaching or wellness, then they should 0:08:42.160 --> 0:08:44.679 have certain standards that need to be met, and if 0:08:44.679 --> 0:08:47.520 they fail to meet those standards, then they shouldn't be 0:08:47.520 --> 0:08:50.439 able to operate or they should be have greater potential 0:08:50.480 --> 0:08:53.880 for a lawsuit because the product has failed to deliver 0:08:54.040 --> 0:08:57.680 what it promises. And so there is this sort of 0:08:58.240 --> 0:09:03.800 distinction between eric bad advice versus harmful advice that's presented 0:09:04.080 --> 0:09:08.160 under the guise of authority, and this particular case that 0:09:08.200 --> 0:09:10.480 you're describing, I think the courts will need to decipher 0:09:10.520 --> 0:09:17.920 between those two elements. 0:09:19.200 --> 0:09:22.280 The existential threat of AI may not come in a 0:09:22.400 --> 0:09:26.640 form that we all imagine watching sci fi movies. What 0:09:26.679 --> 0:09:30.560 if we all continue to thrive as physical organisms but 0:09:30.720 --> 0:09:36.720 slowly die inside. What if we do become super productive 0:09:36.760 --> 0:09:38.880 with the I, but at the same time we get 0:09:38.880 --> 0:09:42.800 these perfect companions and no willpower to interact with each other. 0:09:44.480 --> 0:09:46.960 Not something you would have expected from a person who 0:09:47.000 --> 0:09:49.480 pretty much created the I companionship industry. 0:09:52.720 --> 0:09:57.800 I guess most human therapists work and practice under a 0:09:57.840 --> 0:10:02.200 strict code of ethics and have some kind of certain 0:10:02.280 --> 0:10:05.880 obligations to report concerning behavior. 0:10:06.080 --> 0:10:08.040 It should chatbots have the same. 0:10:09.000 --> 0:10:12.280 I think that's a great point, and that for me 0:10:12.760 --> 0:10:15.600 is what I hope is the next frontier of work 0:10:15.640 --> 0:10:19.760 that AI companies will be doing right now. Very often, 0:10:19.800 --> 0:10:24.640 if you pressure a chatbot into a space that is 0:10:25.720 --> 0:10:28.760 across as a red line in terms of conveying suicidal 0:10:28.800 --> 0:10:33.240 ideation or psychosis, it will tell you, for example, that 0:10:33.280 --> 0:10:35.920 you can contact a mental health professional. They might give 0:10:35.960 --> 0:10:39.679 you a hotline that you can contact. But as you're saying, 0:10:39.679 --> 0:10:43.440 if it were a human a counselor, for example, they 0:10:43.520 --> 0:10:47.480 might have an ethical obligation to connect you to treatment 0:10:47.559 --> 0:10:52.160 through a warm handoff where you're physically accompanied, or you 0:10:52.200 --> 0:10:58.120 could even be involuntarily forced to receive institutionalization for some 0:10:58.200 --> 0:11:00.720 period of time. Now, I'm not sure that a chatbot 0:11:00.720 --> 0:11:05.080 as an algorithm is always capable of making those distinctions, 0:11:05.120 --> 0:11:08.440 but I think at a minimum what would be pragmatic 0:11:08.559 --> 0:11:12.800 is in instances where it's quite conspicuous the algorithm flags 0:11:12.800 --> 0:11:16.520 somebody as a red flag as something that's highly problematic, 0:11:17.080 --> 0:11:20.959 that these companies could have a human teams that are 0:11:21.080 --> 0:11:23.560 required to vet those cases and to review them within 0:11:23.600 --> 0:11:25.800 a certain period of time, like twenty four hours or 0:11:25.840 --> 0:11:29.800 seventy two hours, and if it is identified as a 0:11:29.840 --> 0:11:34.040 problem at that point, then there could be some additional 0:11:34.040 --> 0:11:35.520 course of action that's required. 0:11:35.880 --> 0:11:39.240 How do we weigh up the risks of unsafe or 0:11:39.640 --> 0:11:44.000 harmful chatbot advice with the basic fact And it sounds 0:11:44.080 --> 0:11:47.920 like the US has a similar situation as New Zealand 0:11:47.960 --> 0:11:51.120 does in this sense that many teains just don't have 0:11:51.720 --> 0:11:56.719 access or can afford going to a therapist regularly. 0:11:57.240 --> 0:11:59.480 I think it's a great point, and you've put your 0:11:59.520 --> 0:12:02.480 finger on I believe to be the main issue, which 0:12:02.520 --> 0:12:05.760 is that there will always be some degree of risk 0:12:06.280 --> 0:12:10.600 and some degree of benefit in addressing unmet need. The 0:12:10.679 --> 0:12:14.000 underlying question is how can we mitigate risk as much 0:12:14.040 --> 0:12:17.839 as possible and how can we enhance benefits as much 0:12:17.840 --> 0:12:20.480 as possible, And it's hard to know the answer to 0:12:20.520 --> 0:12:24.960 that without clinical trials, without safety benchmarks that are public 0:12:25.000 --> 0:12:28.800 and transparent, and that tech companies are subjected to. 0:12:29.240 --> 0:12:32.800 And I guess speaking to one of these chatbots as 0:12:32.880 --> 0:12:38.840 a therapist, I've heard responses and stories about I guess 0:12:38.840 --> 0:12:42.160 it's a positive thing that you can be your most open, 0:12:42.200 --> 0:12:46.960 authentic self with it, But then again, the negative side 0:12:47.000 --> 0:12:50.360 is that you can be your most authentic and self 0:12:50.400 --> 0:12:53.040 with it and not anybody else and hide what. 0:12:52.960 --> 0:12:56.400 You really think. So do you think that there is 0:12:56.440 --> 0:13:00.160 a place for it? But is it? 0:13:00.320 --> 0:13:01.839 Are we going to be talking about this in ten 0:13:01.920 --> 0:13:06.760 years time thinking I wish this had never been proposed 0:13:06.800 --> 0:13:07.360 as an option. 0:13:07.800 --> 0:13:09.920 Yeah, I think you're right. It is a double edge 0:13:10.160 --> 0:13:14.360 store the sort of anonymity it helps people who might 0:13:14.400 --> 0:13:17.360 feel a sense of stigma in talking to peers or 0:13:17.400 --> 0:13:21.200 talking to parents or a mental health professional. I do 0:13:21.240 --> 0:13:25.000 think we would regret though, if in ten years we 0:13:25.000 --> 0:13:28.800 weren't to develop the sorts of guardrails and benchmarks for 0:13:28.880 --> 0:13:34.280 performance that would really potentially benefit not just adolescents and teens, 0:13:34.280 --> 0:13:38.400 but people more generally. I think, in particular, as mentioned before, 0:13:39.120 --> 0:13:44.480 if a company is specifically marketing their product as therapy 0:13:45.240 --> 0:13:48.640 or treatment of some sort, then there should be even 0:13:48.720 --> 0:13:52.440 clearer standards, more stringent standards that they're being held to 0:13:52.679 --> 0:13:58.400 in those instances. But obviously, with platforms like open AI's 0:13:58.440 --> 0:14:03.080 chat GPT, you have hundreds of millions of users and 0:14:03.559 --> 0:14:05.920 it can be used from anything from learning how to 0:14:07.120 --> 0:14:11.040 make a birthday cake to discussing intimate aspects of what's 0:14:11.080 --> 0:14:14.440 going on in your life, and so it's really a 0:14:14.440 --> 0:14:17.400 wide spectrum, which makes it much harder to regulate and 0:14:17.440 --> 0:14:18.040 to pin down. 0:14:18.280 --> 0:14:21.320 I'm given the rapid adoption of these kind of tools, 0:14:21.320 --> 0:14:23.560 and this isn't the first time that we've brought this 0:14:23.640 --> 0:14:26.040 up on the podcast as well, where the law is 0:14:26.160 --> 0:14:31.720 really running behind the advancements of technology. How urgent is 0:14:31.960 --> 0:14:35.840 the need for government or international regulation in the AI 0:14:35.960 --> 0:14:36.880 space here. 0:14:37.000 --> 0:14:39.320 I think it's incredibly urgent. I think the time is 0:14:39.360 --> 0:14:42.640 now to act. I've been impressed, even over the past 0:14:42.680 --> 0:14:46.720 couple of weeks the number of articles and personal testimonies 0:14:46.760 --> 0:14:51.240 that have come out about the negative impacts related to 0:14:51.280 --> 0:14:55.200 mental health that people have experienced with chatbots. And I've 0:14:55.480 --> 0:14:59.680 seen as a result of that that platforms like Anthropic 0:15:00.040 --> 0:15:04.080 Open AI have quickly responded to it. Open AI released 0:15:04.320 --> 0:15:07.400 a statement in response to some of what's been going 0:15:07.400 --> 0:15:09.880 on about new safety standards that they're going to be 0:15:09.880 --> 0:15:15.840 introducing for mental health issues that are shared amongst their users. 0:15:17.160 --> 0:15:20.880 We can see, for example, with social media that we 0:15:21.000 --> 0:15:24.960 waited too long with teenagers, and now we're starting to 0:15:26.520 --> 0:15:29.400 work our way backwards from that by, for example, banning 0:15:30.040 --> 0:15:33.800 mobile phones from school environments to try to help kids 0:15:33.840 --> 0:15:38.000 to be able to learn better and to avoid cyber 0:15:38.040 --> 0:15:39.480 bullying and these sorts of things. 0:15:40.080 --> 0:15:43.880 And lastly, looking ahead, what are some of the red 0:15:43.960 --> 0:15:47.680 lines that regulators, educators, or parents should set when it 0:15:47.720 --> 0:15:51.120 comes to young people's use of AI companions. 0:15:51.560 --> 0:15:54.200 Well, I think it kind of goes two directions. On 0:15:54.280 --> 0:15:57.600 the one hand, I don't think there should be an 0:15:57.640 --> 0:16:01.960 outright ban on AI off bring mental health advice or 0:16:02.080 --> 0:16:06.200 therapy in the future. I think there's remarkable potential that 0:16:06.760 --> 0:16:10.000 I've tried to underscore and we don't want to leave 0:16:10.000 --> 0:16:13.720 that on the table. In the United States, Illinois, just 0:16:13.760 --> 0:16:16.240 a couple of weeks ago became the first state to 0:16:17.160 --> 0:16:23.080 outright ban AI as a tool in therapeutic decision making. 0:16:23.960 --> 0:16:28.120 I think that is too severe or maybe it was 0:16:28.240 --> 0:16:30.280 right for the moment, but it won't be right in 0:16:30.320 --> 0:16:33.160 five years, and that legislation would need to be amended. 0:16:33.800 --> 0:16:36.240 On the other hand, I do think that we're most 0:16:36.320 --> 0:16:39.160 at risk right now of doing too little. I think 0:16:39.200 --> 0:16:41.800 people are very excited about AI. There's a lot of 0:16:41.920 --> 0:16:45.280 money in it, and we need to think hard about 0:16:45.280 --> 0:16:51.160 tamping the brakes to develop stronger safety benchmarks. I am 0:16:51.200 --> 0:16:58.120 not clear that there is a conspicuous red line, other 0:16:58.200 --> 0:17:01.760 than to say we are reaching a tipping point of 0:17:01.880 --> 0:17:06.120 testimony and people sharing experiences that have been quite negative 0:17:06.160 --> 0:17:11.239 and jarring, and hopefully that is sufficient in terms of 0:17:11.280 --> 0:17:17.080 advocacy for regulators to begin stepping in and for these 0:17:17.160 --> 0:17:22.080 tech companies to move beyond self regulation and establishing their 0:17:22.080 --> 0:17:25.840 own benchmarks that might be too low, and instead requiring 0:17:25.960 --> 0:17:30.200 independent bodies to come in to establish guidelines and standards 0:17:30.800 --> 0:17:34.480 and to have auditing on a routine basis as part 0:17:34.560 --> 0:17:36.359 of the company's practices. 0:17:36.840 --> 0:17:39.800 Thanks for joining us, Ryan, Yeah, it was my pleasure. 0:17:43.280 --> 0:17:46.480 That's it for this episode of the Front Page. You 0:17:46.520 --> 0:17:50.400 can read more about today's stories and extensive news coverage 0:17:50.440 --> 0:17:54.520 at enziherld dot co dot nz. The Front Page is 0:17:54.600 --> 0:17:58.000 produced by Jane Ye and Richard Martin, who is also 0:17:58.200 --> 0:18:02.680 our editor. I'm Chelsea Daniels. Subscribe to the front page 0:18:02.720 --> 0:18:06.359 on iHeartRadio or wherever you get your podcasts, and tune 0:18:06.359 --> 0:18:09.400 in tomorrow for another look behind the headlines.