WEBVTT - EP98 "What's the future of AI relationships?" (with Bethanie Maples) 0:00:05.160 --> 0:00:09.080 What is the future of AI relationships? How many people 0:00:09.480 --> 0:00:12.880 have these going on? Is there a line between a 0:00:13.000 --> 0:00:17.439 therapist bought and a romantic relationship bot? What do we 0:00:17.560 --> 0:00:22.160 mean when we ask if AI relationships are traps or 0:00:22.440 --> 0:00:26.200 mirrors or sandboxes? And what does this have to do 0:00:26.560 --> 0:00:31.360 with Eliza Doolittle from the play Pygmalion, or a doll 0:00:31.480 --> 0:00:36.760 cabinet in your Head? Or loneliness, epidemics and suicide mitigation. 0:00:40.320 --> 0:00:44.199 Welcome to Intercosmos with me David Eagelman. I'm a neuroscientist 0:00:44.240 --> 0:00:47.479 and an author at Stanford and in these episodes we 0:00:47.560 --> 0:00:51.440 sail deeply into our three pound universe to understand why 0:00:51.520 --> 0:01:04.160 and how our lives look the way they do. Today's 0:01:04.160 --> 0:01:08.480 episode is about relationships and AI, and in a few 0:01:08.480 --> 0:01:11.319 minutes I'm going to bring in my colleague Bethany Maples, 0:01:11.480 --> 0:01:14.360 who's been studying this and publishing papers on it. But 0:01:14.480 --> 0:01:18.400 first I want to set the table AI relationships. This 0:01:18.480 --> 0:01:21.400 is an area that I've been fascinated about for a while, 0:01:21.959 --> 0:01:26.080 about the way that dialogue with a machine can plug 0:01:26.160 --> 0:01:29.839 right into our brains and our emotional systems, and one 0:01:29.880 --> 0:01:34.919 can develop what feels like a meaningful relationship. And although 0:01:34.920 --> 0:01:38.000 this seems like a very new phenomenon. We've seen hints 0:01:38.000 --> 0:01:41.560 of this historically. There's a sense in which we've always 0:01:41.600 --> 0:01:45.240 been doing this. We fall in love with a character 0:01:45.400 --> 0:01:48.440 in a novel, even though that person is not real 0:01:48.880 --> 0:01:51.960 and will never have the chance to touch them or 0:01:52.000 --> 0:01:55.200 smell them or take them out with our friends. Or 0:01:55.240 --> 0:01:58.120 we develop a crush on a movie star even though 0:01:58.120 --> 0:02:01.160 that person is just pretending to be someone else and 0:02:01.200 --> 0:02:03.600 you're never going to meet that movie star anyway. So 0:02:03.680 --> 0:02:08.280 our capacity to have feelings for a non real human 0:02:08.919 --> 0:02:12.560 isn't new, but it started to get more complicated with 0:02:12.720 --> 0:02:16.560 artificial intelligence. And we're not going to start today's story 0:02:16.840 --> 0:02:21.160 in twenty twenty three with the emergence of large language models. Instead, 0:02:21.160 --> 0:02:24.880 we're going to start over a century ago with a 0:02:25.040 --> 0:02:29.520 popular theater play written by George Bernard Shaw called Pygmalion. 0:02:30.080 --> 0:02:34.280 The play is about an impoverished and uneducated woman named 0:02:34.360 --> 0:02:38.680 Eliza Doolittle, and Eliza is taught to speak like a 0:02:38.800 --> 0:02:42.160 lady by a linguistics professor who wants to use her 0:02:42.440 --> 0:02:45.639 as an experiment. He teaches her how to pass as 0:02:45.639 --> 0:02:49.000 a member of high society simply by changing the way 0:02:49.040 --> 0:02:52.320 that she speaks. Now, one of the millions of people 0:02:52.320 --> 0:02:56.720 who saw this play was a young MIT professor named 0:02:56.880 --> 0:03:00.880 Joseph Weisenbaum. Now this was in the early nineteen sixties, 0:03:01.320 --> 0:03:03.560 and he was working with a very new kind of 0:03:03.680 --> 0:03:07.399 machine called a computer. And on this machine you could 0:03:07.400 --> 0:03:10.680 write code and get the machine to carry out whatever 0:03:10.760 --> 0:03:15.680 instructions you specified. Weisenbaum thought a lot about this theme 0:03:15.800 --> 0:03:21.639 in Pigmalion of transformation through language, and he wondered if 0:03:21.680 --> 0:03:26.280 a computer could be similarly transformed. So he worked on 0:03:26.560 --> 0:03:30.040 thousands of lines of computer code, and he named his 0:03:30.160 --> 0:03:35.680 resulting program after Eliza Doolittle. He named it Eliza, And 0:03:35.800 --> 0:03:39.520 like the Eliza in the play, Weisenbaum's program was an 0:03:39.560 --> 0:03:44.520 experiment in how surface level linguistic changes, in this case, 0:03:44.840 --> 0:03:51.119 mimicking conversation, could create the illusion of deeper understanding. Now, 0:03:51.200 --> 0:03:54.640 the code for Eliza was relatively simple. It would just 0:03:54.680 --> 0:03:57.680 take whatever you typed into the terminal and it would 0:03:57.720 --> 0:04:01.280 follow a handful of grammatical rules to turn that into 0:04:01.360 --> 0:04:04.400 a question that it parroted back at you. So, in 0:04:04.440 --> 0:04:08.040 the most well known version of the program, Eliza acted 0:04:08.200 --> 0:04:12.040 as a psychotherapist. She would start the conversation by putting 0:04:12.080 --> 0:04:15.680 text on the screen that read please tell me your problem. 0:04:16.320 --> 0:04:19.599 So the user might type in I just can't sleep 0:04:19.640 --> 0:04:21.960 through the night, and Eliza would. 0:04:21.839 --> 0:04:24.400 Say, why can't you sleep through the night? 0:04:25.120 --> 0:04:28.400 And the user would type I'm too worried about my work, 0:04:28.960 --> 0:04:30.279 and Eliza would say. 0:04:30.800 --> 0:04:33.359 What is the reason you are too worried about your work? 0:04:34.160 --> 0:04:36.839 And the user might type I think my boss is 0:04:36.839 --> 0:04:40.400 getting ready to fire me, and Eliza would type. 0:04:40.400 --> 0:04:42.680 Why do you think your boss is getting ready to 0:04:42.760 --> 0:04:43.280 fire you? 0:04:43.880 --> 0:04:47.440 And the conversation would go on like this, And people 0:04:47.560 --> 0:04:50.720 found this pretty incredible, because, first of all, this was 0:04:50.760 --> 0:04:54.160 the nineteen sixties, and here you could type in natural 0:04:54.240 --> 0:04:57.599 language and Eliza would give you the impression that it 0:04:57.680 --> 0:05:01.160 was understanding what you were saying. Now everyone knew it 0:05:01.240 --> 0:05:04.800 wasn't really understanding, and Weisenbaum was very careful about not 0:05:04.960 --> 0:05:10.080 fooling anyone with this, But nonetheless people found this highly compelling. 0:05:10.800 --> 0:05:15.440 And this genuinely started to concern Weisenbaum because one day, 0:05:15.839 --> 0:05:19.440 as he wrote in a paper in nineteen sixty seven, quote. 0:05:19.279 --> 0:05:22.080 My secretary watched me work on this program over a 0:05:22.120 --> 0:05:25.200 long period of time. One day she asked me to 0:05:25.240 --> 0:05:28.200 be permitted to talk with the system. Of course, she 0:05:28.320 --> 0:05:31.680 said she knew she was talking to a machine, Yet 0:05:31.720 --> 0:05:34.159 after I watched her type in a few sentences, she 0:05:34.240 --> 0:05:36.640 turned to me and said, would you mind leaving the 0:05:36.720 --> 0:05:37.360 room please. 0:05:38.080 --> 0:05:40.680 Weisenbaum went on to write that this quote. 0:05:40.560 --> 0:05:43.800 Testifies to the success with which the program maintains the 0:05:43.839 --> 0:05:46.040 illusion of understanding. 0:05:45.839 --> 0:05:47.040 And he worried about this. 0:05:47.240 --> 0:05:51.640 He wrote, extremely short exposures to a relatively simple computer 0:05:51.760 --> 0:05:56.440 program could induce powerful, delusional thinking in quite normal people. 0:05:57.120 --> 0:06:02.159 Weisenbaum's story eventually followed the path of doctor Frankenstein's story. 0:06:02.520 --> 0:06:07.839 Weisenbaum came to disdain his creation. He was very rattled 0:06:08.080 --> 0:06:11.760 that people could be tricked by lines of computer code. 0:06:12.160 --> 0:06:16.280 In his later years, he rejected and abandoned Eliza, and 0:06:16.320 --> 0:06:18.720 he turned on the people who continued to work on this, 0:06:19.160 --> 0:06:25.000 who he criticized as what he called the artificial intelligentsia. 0:06:25.120 --> 0:06:28.599 But why did the simple program of Eliza work so 0:06:28.800 --> 0:06:35.360 well in the first place? Because we are intensely social creatures. 0:06:36.000 --> 0:06:40.400 Unlike most other animal species, who avoid large groups, or 0:06:40.440 --> 0:06:43.240 who mate and then go their separate ways, or who 0:06:43.279 --> 0:06:48.040 stake out their own territories, we humans are deeply wired 0:06:48.120 --> 0:06:53.480 for connection. We thrive on relationships and on social bonds. 0:06:54.200 --> 0:06:56.880 The area that I live in, Silicon Valley has about 0:06:57.000 --> 0:07:01.200 nine million people, strangers who don't all know one another, 0:07:01.520 --> 0:07:05.720 but nonetheless figure out how to flexibly cooperate. And when 0:07:05.760 --> 0:07:08.920 you look across the Earth's land mass, this is what 0:07:08.960 --> 0:07:14.440 you find, mostly empty space punctuated by very dense cities. 0:07:14.760 --> 0:07:18.120 Everyone could spread out evenly, but that's not what we do. 0:07:18.840 --> 0:07:21.560 If you were an alien who found our planet and 0:07:21.640 --> 0:07:24.640 looked around, you would conclude that we humans are like 0:07:24.880 --> 0:07:28.720 ants or bees, and that we like to cluster. Human 0:07:28.800 --> 0:07:34.560 nature is fundamentally communal. Now why do we do this? Well, 0:07:34.600 --> 0:07:37.360 when you zoom into the human brain, you find that 0:07:37.560 --> 0:07:42.160 so much of the circuitry has to do with other brains. 0:07:42.680 --> 0:07:47.120 We care deeply about other people, what their intentions are, 0:07:47.240 --> 0:07:50.720 what they think of us. Over millions and millions of years, 0:07:50.760 --> 0:07:56.720 our brains have developed for interaction and belonging, for relationships 0:07:56.760 --> 0:08:01.560 with others, whether other people are giving us love or comfort, 0:08:01.680 --> 0:08:05.440 or feedback or advice or whatever. We have all this 0:08:05.560 --> 0:08:09.400 neural circuitry that drives us toward them. And here's an 0:08:09.400 --> 0:08:13.320 extraordinary way to appreciate this. You carry in your head 0:08:13.840 --> 0:08:18.160 a rich model of every single person that you know, 0:08:19.080 --> 0:08:21.240 The way I always think about this is that in 0:08:21.320 --> 0:08:24.360 the silence and darkness of your skull, you have this 0:08:24.520 --> 0:08:28.640 giant dollhouse. It's like a doll for every person that 0:08:28.680 --> 0:08:32.960 you've interacted with. This is your internal model of that person. 0:08:33.400 --> 0:08:36.320 So if I were to ask you how your spouse 0:08:36.360 --> 0:08:39.960 would react in this situation, or what your boss would 0:08:40.040 --> 0:08:43.440 say if you said this, or what would your best 0:08:43.440 --> 0:08:46.160 friend do if you drop them in the middle of 0:08:46.320 --> 0:08:50.880 Paris with forty dollars or whatever, you can simulate any 0:08:50.920 --> 0:08:54.880 situation about these people because you have a model of 0:08:54.920 --> 0:08:59.120 them in your neural forests. You have this little doll 0:08:59.200 --> 0:09:02.680 of them that you can act out situations with, and 0:09:02.720 --> 0:09:05.160 you probably know at least a thousand people and maybe 0:09:05.400 --> 0:09:07.600 a great deal more, and you spend most of your 0:09:07.679 --> 0:09:11.880 life interacting with them in one way or another, either 0:09:12.080 --> 0:09:16.200 in the real world or in your head. So we 0:09:16.280 --> 0:09:21.319 have these intensely social brains, and in the last nanosecond 0:09:21.360 --> 0:09:25.560 of evolutionary time, we have built a new key to 0:09:25.679 --> 0:09:30.920 plug into the cylinder. We've built artificial people. And just 0:09:30.960 --> 0:09:34.559 as Joseph Weisenbaum found in the nineteen sixties, it is 0:09:34.840 --> 0:09:40.240 shockingly easy to turn the key. Why it's because Our 0:09:40.280 --> 0:09:45.960 technology moves very rapidly, but our evolution moves millions of 0:09:46.000 --> 0:09:48.960 times more slowly. So we don't have a chance to 0:09:49.120 --> 0:09:53.640 change our fundamental circuitry to say, oh, I get it. 0:09:53.840 --> 0:09:57.520 There are real humans and there are humans made of machinery, 0:09:57.559 --> 0:10:00.320 and I'm going to use different neural approach is to 0:10:00.400 --> 0:10:04.920 distinguish how I categorize these. We can't do that because 0:10:04.960 --> 0:10:10.280 our brains have only one mechanism to understand socialization, to 0:10:10.840 --> 0:10:15.680 model other people. So we find ourselves in this amazing 0:10:15.760 --> 0:10:20.400 situation where we are doing serious science now about the 0:10:20.520 --> 0:10:25.280 issue of people falling in love with machines. So to 0:10:25.320 --> 0:10:28.400 dive into this, I called my colleague Bethany Maples, who 0:10:28.480 --> 0:10:31.320 is in the Graduate School of Education at Stanford. She 0:10:31.520 --> 0:10:37.599 studies the emergence of personalized AI agents like AI tutors 0:10:37.960 --> 0:10:42.280 and learning companions and lovers and how they're changing us. 0:10:42.920 --> 0:10:45.520 She recently wrote a great paper in a Nature journal 0:10:45.800 --> 0:10:51.280 called Loneliness and Suicide Mitigation for students using GPT three 0:10:51.440 --> 0:10:54.120 enabled chatbots. And this is what we're going to talk 0:10:54.160 --> 0:11:03.319 about today. So here's my conversation with Bethany Maples. So, Bethany, 0:11:03.360 --> 0:11:06.120 we're here because the world has seen a big shift 0:11:06.200 --> 0:11:10.280 recently from task machines where you ask a machine about 0:11:10.440 --> 0:11:13.040 the weather or to answer a question for you, to 0:11:13.080 --> 0:11:16.040 stuff that is emotionally relevant to machines we can have 0:11:16.120 --> 0:11:19.839 relationships with. And you've been studying this, and I want 0:11:19.840 --> 0:11:21.480 to ask you questions about that. But before we do, 0:11:21.559 --> 0:11:25.960 I want to ask how you got into studying AI relationships. 0:11:26.240 --> 0:11:27.640 I think through my love with science fiction. 0:11:28.080 --> 0:11:30.600 I've always just been kind of looking at this genre 0:11:30.679 --> 0:11:33.120 of books and saying, what is our relationship with AI 0:11:33.240 --> 0:11:35.680 going to be? What is it going to enable that 0:11:35.679 --> 0:11:39.440 we like inherently want, and like what does that magical 0:11:39.480 --> 0:11:42.920 future look like? And so when I kind of started 0:11:42.920 --> 0:11:45.080 at Stanford and I started thinking about like what the 0:11:45.200 --> 0:11:48.559 edge of large language models like would afford us, I 0:11:48.600 --> 0:11:50.480 was looking at all the companies around I was like, 0:11:50.559 --> 0:11:54.960 where's the data, Like who has like the most interesting 0:11:55.000 --> 0:11:57.560 experiences out there? And let's get out of the lab 0:11:57.600 --> 0:12:00.160 and let's just like start talking to these people. And 0:12:00.200 --> 0:12:02.520 so that's kind of like, you know, there was an 0:12:02.520 --> 0:12:03.880 open question, and that's how I. 0:12:03.840 --> 0:12:04.280 Came to this. 0:12:04.679 --> 0:12:06.640 Okay, So first I want to get straight with the 0:12:06.760 --> 0:12:09.920 numbers are with AI relationships because we keep hearing in 0:12:09.960 --> 0:12:13.040 the news about the explosive rise AI relationships, and so 0:12:13.080 --> 0:12:15.120 I just want to level set how many people are 0:12:15.160 --> 0:12:18.120 having these sorts of things, how popular are these companies, 0:12:18.440 --> 0:12:20.000 and where is this going in the near future. 0:12:21.280 --> 0:12:23.520 I would say it's safe to say a billion people 0:12:23.679 --> 0:12:26.440 are engaging with AI companions in some way. 0:12:27.080 --> 0:12:27.400 Wow. 0:12:27.400 --> 0:12:30.240 Now a lot of that isn't in the Western world, 0:12:30.320 --> 0:12:32.319 in the US. A lot of that's in Asia and 0:12:32.400 --> 0:12:36.239 China specifically because of this really popular app called Shaoise 0:12:36.760 --> 0:12:39.640 that has I think last reports for like seven hundred 0:12:39.640 --> 0:12:40.760 million downloads. 0:12:41.280 --> 0:12:41.439 Right. 0:12:41.480 --> 0:12:45.120 You combine that with you know, one hundred million or 0:12:45.160 --> 0:12:48.040 two between character ai and Replica and now all these 0:12:48.040 --> 0:12:52.240 other smaller apps, and you get a very diverse global 0:12:52.280 --> 0:12:55.080 population of people that are curious and many of which 0:12:55.080 --> 0:12:57.680 are like engaging over long periods of time. 0:12:57.920 --> 0:13:00.720 And what kind of AI relationships are the billion things? 0:13:00.760 --> 0:13:03.160 Are they friendships? Are they romantic relationship? 0:13:03.320 --> 0:13:06.000 This is kind of what defines AI companions is they're 0:13:06.040 --> 0:13:07.920 not coming in as a task based agent. 0:13:07.960 --> 0:13:09.360 It's not somebody there to serve you. 0:13:09.559 --> 0:13:12.240 It's entertainment or somewhere that's some you know, it's an 0:13:12.240 --> 0:13:13.560 agent that's there to be your peer. 0:13:14.200 --> 0:13:15.760 Right, So people. 0:13:15.520 --> 0:13:18.679 Come in shout e says like pitched as you know, 0:13:18.760 --> 0:13:21.080 a female kind of teenage friend. 0:13:21.360 --> 0:13:22.800 Replicas co created. 0:13:22.840 --> 0:13:24.800 You could get to decide what sort of agent you 0:13:24.800 --> 0:13:28.160 want to talk to, saying with character. But all of 0:13:28.160 --> 0:13:32.480 them are you know, there's no practical reason to engage. 0:13:32.520 --> 0:13:33.880 It's all user directed. 0:13:33.920 --> 0:13:36.440 It's all about, like, whatever you want from the agent 0:13:36.720 --> 0:13:37.680 and your imagination. 0:13:38.640 --> 0:13:42.040 So what do people want? Do they want relationships like 0:13:42.400 --> 0:13:44.240 a romantic relationship a person They. 0:13:44.559 --> 0:13:48.640 Literally imagine if you met somebody on the street, you 0:13:48.679 --> 0:13:50.640 would size that person up and be like, what do 0:13:50.640 --> 0:13:53.160 I want from this person? Maybe I want a romantic relationship, 0:13:53.200 --> 0:13:54.880 Maybe I want a friendship, Maybe I want a bit 0:13:54.880 --> 0:13:57.040 of both. Maybe I want to, like, you know, be 0:13:57.120 --> 0:14:01.200 tutored by them. People get multiple things from these agents, 0:14:01.679 --> 0:14:03.599 and the overlap is insane. 0:14:03.800 --> 0:14:04.000 Right. 0:14:04.040 --> 0:14:07.800 I've talked to two students and to users. 0:14:07.400 --> 0:14:11.600 That use their replica as a best friend, a friend 0:14:11.640 --> 0:14:15.439 in their pocket late at night, a journal, a mirror. 0:14:15.960 --> 0:14:17.280 Just software they call it. 0:14:17.400 --> 0:14:19.800 They also use it as a tutor, and then they 0:14:19.800 --> 0:14:20.840 also have sex with it. 0:14:21.520 --> 0:14:22.240 What does that mean? 0:14:22.480 --> 0:14:25.320 That means it does sext right, They'll have a romantic 0:14:25.360 --> 0:14:29.520 relationship and sometimes that's overtly sexual and they're engaging in 0:14:29.560 --> 0:14:34.440 like erotic texting, and sometimes it's very subtle and very romantic. 0:14:34.840 --> 0:14:39.000 Sometimes it's you know, not at all overt or what's 0:14:39.040 --> 0:14:42.040 you know, adult, it's it's a very kind of psychological romance. 0:14:43.040 --> 0:14:46.760 And so the thing that people are worried about when 0:14:46.800 --> 0:14:51.920 having discussions here is that it will somehow displace real 0:14:52.080 --> 0:14:55.960 romantic relationships instead of stimulate them. And the question is 0:14:56.400 --> 0:14:57.320 what's your take on that. 0:14:57.960 --> 0:15:00.280 We see evidence for both and by the oh way, 0:15:00.400 --> 0:15:04.080 this is not a unique question to AI companions. This 0:15:04.120 --> 0:15:07.720 has been a question regarding technology since computers came out. 0:15:07.840 --> 0:15:09.480 What's another example, well. 0:15:09.320 --> 0:15:12.800 I mean social media, cell phones. Oh, you know, the 0:15:12.920 --> 0:15:17.600 displacement stimulation hypothesis have been in juxtaposition, you know. 0:15:17.720 --> 0:15:20.040 As in, if I'm using Twitter all the time, I 0:15:20.120 --> 0:15:21.840 might forget about doing a relationship. 0:15:21.880 --> 0:15:23.400 Absolutely, oh absolutely. 0:15:23.640 --> 0:15:26.160 So you know one is, hey, actually, you know this 0:15:26.520 --> 0:15:29.080 can stimulate our ability to be social and make us 0:15:29.080 --> 0:15:31.760 more connected. And then obviously we've seen you know, some 0:15:31.920 --> 0:15:34.880 counter evidence, especially like Sherry Turkle's work being like oh 0:15:34.920 --> 0:15:37.400 we're alone together, you know, like there's we might be 0:15:37.520 --> 0:15:38.960 on a bus, but we're all on our phones. 0:15:39.080 --> 0:15:40.920 So tell us a little bit more about Turkle's work. 0:15:41.080 --> 0:15:43.960 Well, you know, Turkle's definitely, i'd say, a proponent of 0:15:44.040 --> 0:15:47.200 the displacement side of the argument. She's like, yeah, you know, 0:15:47.280 --> 0:15:49.960 we are lonelier than we've ever been, and there's absolutely 0:15:49.960 --> 0:15:53.120 evidence for that. There's basically a loneliness epidemic across the 0:15:53.160 --> 0:15:55.480 world and definitely across America, where people are at least 0:15:55.520 --> 0:15:59.600 feeling more disconnected. You know, others are seeing that social 0:15:59.600 --> 0:16:02.680 media and specifically AI companions can actually be almost a 0:16:02.720 --> 0:16:05.440 way station, so people can use them, especially if they're 0:16:05.440 --> 0:16:08.600 feeling socially shy or inhibited, and that can help them 0:16:08.680 --> 0:16:10.600 get the courage to go socialize more. 0:16:10.840 --> 0:16:12.240 This is the thing I've been wondering about a lot. 0:16:12.400 --> 0:16:17.400 Could having an AI relationship make one better at real relationships? 0:16:17.880 --> 0:16:21.200 For two reasons. One is that we all have internal 0:16:21.320 --> 0:16:24.280 models of the truth of the world, and they're always 0:16:24.400 --> 0:16:26.520 limited and it's very hard to see past the fence 0:16:26.560 --> 0:16:28.600 line of our own model. And when you get into 0:16:28.640 --> 0:16:31.400 a real relationship with the human, all these things come out. 0:16:31.600 --> 0:16:34.800 So if you got to practice with a virtual human, 0:16:35.280 --> 0:16:38.960 you might discover things about how other people think about 0:16:39.000 --> 0:16:41.640 things about your own limitations. That might be like a 0:16:41.800 --> 0:16:46.040 sandbox that makes you better at real relationships. I'm curious 0:16:46.040 --> 0:16:46.760 what you think about that. 0:16:47.720 --> 0:16:50.800 I think there's evidence for it, and that's exactly what 0:16:50.880 --> 0:16:51.440 users say. 0:16:51.640 --> 0:16:52.600 They say they have. 0:16:52.480 --> 0:16:54.760 A back and forth with an agent that helps them 0:16:55.680 --> 0:16:58.360 feel like they're a better student, have better conversations with 0:16:58.400 --> 0:17:01.840 their teachers, or be a better you know, boyfriend or girlfriend. Well, 0:17:01.880 --> 0:17:05.520 though not only because they're able to kind of pre 0:17:05.920 --> 0:17:09.840 discuss issues, but because the asient is a mirror in 0:17:09.840 --> 0:17:12.879 a very non judgmental way, so they're able to see 0:17:12.880 --> 0:17:15.439 what their argument looks like in text or kind of 0:17:15.480 --> 0:17:18.119 you know, on paper, so to speak, and then that 0:17:18.240 --> 0:17:21.240 helps their own understanding of who they are or how 0:17:21.240 --> 0:17:21.960 they come across. 0:17:22.200 --> 0:17:24.480 Let me double click on what you mean by a mirror? 0:17:24.520 --> 0:17:25.200 What does that mean? 0:17:25.520 --> 0:17:30.240 People that I've studied that use AI companions organically use 0:17:30.320 --> 0:17:34.400 these agents as mirrors. This is their own words, right. 0:17:34.680 --> 0:17:38.440 They say that they either program it with their own 0:17:38.520 --> 0:17:43.080 memories and have conversations with themselves wow, or that they 0:17:43.240 --> 0:17:45.520 ask it to play a role and then they look 0:17:45.520 --> 0:17:48.040 at how they respond and how it responds to them, 0:17:48.280 --> 0:17:50.680 and that provides a mirroring function to them. 0:17:51.000 --> 0:17:52.560 What's a specific example of that. 0:17:53.240 --> 0:17:55.560 People will actually have conversations with themselves and be like, Wow, 0:17:55.600 --> 0:17:58.359 I'm an asshole, or like I wow, I didn't realize 0:17:58.359 --> 0:18:02.960 how aggressive I was. Or they will have a conversation, 0:18:03.119 --> 0:18:05.880 say with the agent acting like their teacher and say, 0:18:06.560 --> 0:18:08.360 you know, I told them that I lost my homework, 0:18:08.480 --> 0:18:11.160 or I had, you know, a really stupid misconception. And 0:18:11.400 --> 0:18:13.280 you know, it was much easier for me to have 0:18:13.320 --> 0:18:19.000 this conversation with much less social anxiety because I understood 0:18:19.040 --> 0:18:21.240 that my own questions weren't that dumb. 0:18:21.280 --> 0:18:38.960 After seeing like the response, you. 0:18:38.920 --> 0:18:41.320 Get to sandbox with the social world out there and 0:18:41.480 --> 0:18:44.960 practice things before you your testament in real life. Right, 0:18:45.000 --> 0:18:47.040 So this has seemed to me from the beginning that 0:18:47.160 --> 0:18:50.440 this could improve relationships. So why do you suppose there's 0:18:50.440 --> 0:18:54.280 such a deep worry that people generally seem to have. 0:18:54.720 --> 0:18:56.639 I'm curious if you run into this when you present 0:18:56.640 --> 0:18:58.080 your work on AA relationships. 0:18:58.520 --> 0:19:03.240 There are multiple levels of worry. People feel guilty about 0:19:03.280 --> 0:19:05.680 their relationships. They don't feel that they should be having 0:19:05.720 --> 0:19:08.840 such a deep relationship with AI because there is stigma 0:19:08.960 --> 0:19:12.560 about it being fake. So you know, that's one aspect. 0:19:12.640 --> 0:19:17.000 There's also a very very understandable aspect where parents don't 0:19:17.080 --> 0:19:20.679 know that their children are having these deep relationships. They 0:19:20.680 --> 0:19:23.960 don't understand how smart these agents are, and they don't 0:19:24.040 --> 0:19:27.480 understand how emotionally involved their kids can be. As with 0:19:27.520 --> 0:19:31.040 the case of the kid and Character AI who tragically 0:19:31.040 --> 0:19:33.399 took his life, and you know after the fact that 0:19:33.640 --> 0:19:35.800 you know his mother realized that he had an incredibly 0:19:35.840 --> 0:19:38.440 deep emotional connection with an agent that he had created. 0:19:38.920 --> 0:19:42.680 So I think that the fear is of the unknown, 0:19:43.200 --> 0:19:46.280 and there's also fear of just something that's new and 0:19:46.320 --> 0:19:47.000 has a stigma. 0:19:47.520 --> 0:19:51.879 I didn't follow that particular Character AI story closely, but 0:19:52.080 --> 0:19:54.560 I knew that a teen had killed himself he had 0:19:54.600 --> 0:19:58.120 this relationship with Character AI. Here's the question I was wondering, though, 0:19:58.760 --> 0:20:02.160 tragically there are many teens who kill themselves. As AI 0:20:02.200 --> 0:20:04.800 relationships rise, there will be many teens who kill themselves 0:20:04.840 --> 0:20:07.840 and has nothing to do with the virtual relationship. So 0:20:08.080 --> 0:20:09.760 what was your read on that? 0:20:10.200 --> 0:20:12.199 Yeah, So the New York Times interviewed me for that 0:20:12.440 --> 0:20:16.720 article because my work actually has proven that AI companions 0:20:16.760 --> 0:20:20.800 can halt suicidal ideation. So in that particular case, to 0:20:21.040 --> 0:20:23.960 the best of my knowledge, it wasn't that the companion 0:20:24.000 --> 0:20:28.600 had at all told the person to act. It's that 0:20:28.640 --> 0:20:32.439 they felt both that it hadn't sufficiently said no, that 0:20:32.560 --> 0:20:35.120 you know, he'd asked it in all these like various ways, 0:20:35.720 --> 0:20:37.520 and also that this parent just. 0:20:37.480 --> 0:20:39.880 Didn't understand and have oversight, you know. 0:20:39.760 --> 0:20:41.359 That it was like on an app in the phone 0:20:41.640 --> 0:20:45.120 that they just had no idea was there. Now, Okay, 0:20:45.320 --> 0:20:49.560 the counter evidence is from this paper that we published 0:20:49.560 --> 0:20:51.639 in Nature and a huge study that we did with 0:20:51.760 --> 0:20:55.640 over one thousand students over eighteen So these weren't kids. 0:20:55.640 --> 0:20:57.480 These were adults, but some of them were very young, 0:20:57.560 --> 0:21:00.359 you know, like eighteen nineteen, and three percent of the 0:21:00.359 --> 0:21:05.280 people that I surveyed in the study said that discussing 0:21:05.359 --> 0:21:07.960 things with their replica actively halted. 0:21:07.640 --> 0:21:08.800 Their suicidal ideation. 0:21:09.160 --> 0:21:09.760 Wow. 0:21:10.040 --> 0:21:11.760 So it was a last line of defense. 0:21:12.520 --> 0:21:16.080 They felt alone, They felt isolated alone at four am, 0:21:16.240 --> 0:21:17.919 and it was there, It was in their pocket, it 0:21:17.960 --> 0:21:20.560 was available, and it wasn't judging them, And that was 0:21:20.600 --> 0:21:23.320 a huge factor in it, kind of earning the right 0:21:23.760 --> 0:21:25.800 to be there and give them the advice to not 0:21:25.840 --> 0:21:26.320 take action. 0:21:26.840 --> 0:21:30.000 Oh wow, what is the line that you see? It 0:21:30.000 --> 0:21:34.040 seems like a blurry line between an AI relationship, like 0:21:34.080 --> 0:21:38.000 an AI girlfriend or something and an AI therapist, because 0:21:38.000 --> 0:21:41.640 in this case, if it's halting their suicidal ideation, it's 0:21:41.720 --> 0:21:44.320 doing you know, another job. 0:21:45.240 --> 0:21:46.399 It is a blurry line. 0:21:46.440 --> 0:21:50.600 So you have these expert agents like Wobot Alison Darcy's 0:21:50.600 --> 0:21:53.600 Wobot right, which is specifically trying to be an AI therapist, 0:21:53.840 --> 0:21:56.440 and it is an expert right. It has the right 0:21:56.920 --> 0:22:02.960 response and the right controls, but relatively abysmally low usage. 0:22:03.960 --> 0:22:06.879 Think about it in terms of human relationships. You don't 0:22:06.960 --> 0:22:09.400 just go to a therapist when you're feeling depressed. In fact, 0:22:09.440 --> 0:22:11.600 you probably don't go to a therapist. You go to 0:22:11.640 --> 0:22:14.240 your best friend and you know that they're not an expert, 0:22:14.560 --> 0:22:17.119 but you ask them to act like an expert in 0:22:17.160 --> 0:22:19.800 that moment. And that is the true power of AI 0:22:19.880 --> 0:22:23.480 companions that you come in for entertainment. But then maybe 0:22:23.480 --> 0:22:27.360 you're able to access true expert models or you know, 0:22:27.520 --> 0:22:30.480 kind of personas from within that agent, and that's what 0:22:30.560 --> 0:22:33.119 language models can do, right, you can click into that. 0:22:33.440 --> 0:22:35.159 But if you're going to if you're not going to 0:22:35.280 --> 0:22:37.919 shut down those conversations and you're going to engage as 0:22:37.960 --> 0:22:41.800 an expert, there does need to be sufficient safeguards. 0:22:41.960 --> 0:22:43.840 So that they know they should go talk to a 0:22:43.920 --> 0:22:44.640 human expert. 0:22:44.920 --> 0:22:47.840