1 00:00:00,360 --> 00:00:02,800 Speaker 1: I think the key principle is is the technology being 2 00:00:02,840 --> 00:00:06,800 Speaker 1: designed to replace teachers or help make teachers better teachers. 3 00:00:07,280 --> 00:00:10,160 Speaker 1: He is the technology being designed to replace your relationships 4 00:00:10,480 --> 00:00:13,240 Speaker 1: or deep in your ability to have better human relationships. 5 00:00:14,160 --> 00:00:17,560 Speaker 2: Tristan Harris loves technology. He wants to live in a 6 00:00:17,560 --> 00:00:20,480 Speaker 2: world where tech is in the service of people. It 7 00:00:20,560 --> 00:00:24,880 Speaker 2: fosters human growth and connection. Unfortunately, right now that's not 8 00:00:25,160 --> 00:00:28,040 Speaker 2: exactly what tech is doing, and Tristan and is nonprofit 9 00:00:28,040 --> 00:00:31,360 Speaker 2: the Center for Humane Technology are trying to help. Tristan's 10 00:00:31,400 --> 00:00:34,960 Speaker 2: background is really interesting. He's a Bay Area kid educated 11 00:00:34,960 --> 00:00:37,840 Speaker 2: at Stanford with all the other tech leaders. His friends 12 00:00:37,880 --> 00:00:41,280 Speaker 2: in college actually started Instagram. He launched his own startup 13 00:00:41,280 --> 00:00:43,440 Speaker 2: that was acquired by Google, where he went and worked 14 00:00:43,440 --> 00:00:46,400 Speaker 2: for a number of years. So Tristan knows and has 15 00:00:46,520 --> 00:00:49,280 Speaker 2: known the handful of people who are designing and building 16 00:00:49,280 --> 00:00:53,800 Speaker 2: the technology we use every day, technology like Instagram and 17 00:00:53,880 --> 00:00:56,640 Speaker 2: other social media apps that were all now addicted to 18 00:00:57,560 --> 00:01:00,280 Speaker 2: Through the Center for Humane Technology, Tristan has been raising 19 00:01:00,320 --> 00:01:03,360 Speaker 2: the red flag about social media, pointing out the many 20 00:01:03,440 --> 00:01:07,920 Speaker 2: ways that these platforms exploit human psychology and vulnerability, how 21 00:01:07,920 --> 00:01:10,319 Speaker 2: they can isolate us and make us dependent on the 22 00:01:10,360 --> 00:01:13,039 Speaker 2: doom scroll. It's a model that has led to what 23 00:01:13,120 --> 00:01:17,240 Speaker 2: Tristan calls human downgrading, where the tech gets better, but 24 00:01:17,600 --> 00:01:20,200 Speaker 2: it changes our lives for the worse, and it's a 25 00:01:20,200 --> 00:01:23,840 Speaker 2: model that AI is now following. Tristan is concerned that 26 00:01:23,880 --> 00:01:27,199 Speaker 2: AI has potential to be the next social media, only 27 00:01:27,240 --> 00:01:30,160 Speaker 2: this go around, instead of hacking our attention, AI is 28 00:01:30,240 --> 00:01:35,480 Speaker 2: hacking our attachment. I'm Lori Siegel, and you're listening to 29 00:01:35,520 --> 00:01:47,080 Speaker 2: Mostly Human, a tech podcast through a human lens. Tristan, 30 00:01:47,120 --> 00:01:49,960 Speaker 2: I've known you throughout the years, and you've always been 31 00:01:50,000 --> 00:01:53,520 Speaker 2: the person who's speaking very eloquently about what's coming around 32 00:01:53,520 --> 00:01:55,240 Speaker 2: the corners and what we're not talking about. 33 00:01:55,640 --> 00:01:57,120 Speaker 3: So let's go to this moment. 34 00:01:57,160 --> 00:01:59,320 Speaker 2: You were talking about social media for the last like 35 00:01:59,400 --> 00:02:02,440 Speaker 2: fifteen years in the harms if we're not careful, But 36 00:02:02,560 --> 00:02:06,280 Speaker 2: now we're in an AI moment, and this moment I 37 00:02:06,320 --> 00:02:08,320 Speaker 2: almost feel like, and you can correct me if I'm wrong, 38 00:02:08,560 --> 00:02:12,200 Speaker 2: the stakes are even higher ground us in this new reality. 39 00:02:12,600 --> 00:02:16,040 Speaker 1: So you know, probably when you and I first started 40 00:02:16,040 --> 00:02:18,720 Speaker 1: talking about maybe it, first introduced to each other in 41 00:02:18,760 --> 00:02:21,799 Speaker 1: twenty thirteen, how could we predict so much of the 42 00:02:21,840 --> 00:02:26,320 Speaker 1: effects that what social media would do a more addicted, distracted, polarized, 43 00:02:26,360 --> 00:02:30,720 Speaker 1: sexualized culture. We could predict all of that because of 44 00:02:30,760 --> 00:02:34,160 Speaker 1: one thing is if you look at the incentives, meaning 45 00:02:34,200 --> 00:02:36,480 Speaker 1: that people say, well, social media, it's giving people a voice, 46 00:02:36,520 --> 00:02:38,880 Speaker 1: helps people connect with your friends. Yeah, it does those things. 47 00:02:39,320 --> 00:02:44,240 Speaker 1: But is Facebook's business model, is TikTok's business model improving 48 00:02:44,800 --> 00:02:46,800 Speaker 1: the health of society or giving people a voice? Or 49 00:02:46,840 --> 00:02:50,079 Speaker 1: is their business model showing whatever at your nervous system 50 00:02:50,200 --> 00:02:53,239 Speaker 1: keeps you scrolling in a doom scrolling loop. And obviously 51 00:02:53,240 --> 00:02:55,360 Speaker 1: it's the latter, And so I think one of the 52 00:02:55,400 --> 00:02:59,200 Speaker 1: key things we have to do is demystify which direction 53 00:02:59,240 --> 00:03:02,520 Speaker 1: technologies going by getting clear about what the incentives are. 54 00:03:02,560 --> 00:03:05,000 Speaker 1: So if you flash to AI, there's all these things 55 00:03:05,040 --> 00:03:08,400 Speaker 1: it could do. Why are we seeing the major AI 56 00:03:08,480 --> 00:03:13,720 Speaker 1: companies release these AI deep fake slop apps meaning it's 57 00:03:13,720 --> 00:03:16,680 Speaker 1: like a TikTok, but it's just AI generated deep fake 58 00:03:16,760 --> 00:03:21,320 Speaker 1: slopped So Meta released Vibes, Open Ai released Sora, And 59 00:03:21,760 --> 00:03:23,440 Speaker 1: why are they doing that when they said they're here 60 00:03:23,440 --> 00:03:25,959 Speaker 1: to cure cancer and they're here to solve climate change 61 00:03:25,960 --> 00:03:29,360 Speaker 1: and they're releasing something that just keeps people scrolling. And 62 00:03:29,400 --> 00:03:32,280 Speaker 1: the answer is because they're racing for market dominance. They're 63 00:03:32,320 --> 00:03:35,200 Speaker 1: racing to get to AGI, they're racing to get to 64 00:03:35,680 --> 00:03:39,880 Speaker 1: basically owning the world economy first. And what that's important 65 00:03:39,920 --> 00:03:43,280 Speaker 1: about getting that is that we are not going to 66 00:03:43,360 --> 00:03:47,240 Speaker 1: get this utopian world on the current trajectory because they're 67 00:03:47,320 --> 00:03:51,640 Speaker 1: racing to release the most powerful, inscrutable, uncontrollable technology we've 68 00:03:51,680 --> 00:03:55,680 Speaker 1: ever invented with the worst possible incentives of getting there first, 69 00:03:55,800 --> 00:03:58,440 Speaker 1: rather than making sure we do those things right, Will. 70 00:03:58,360 --> 00:03:59,480 Speaker 3: You talk about the incentives right? 71 00:03:59,520 --> 00:04:03,720 Speaker 2: You talk about the incentives of this new generation with 72 00:04:03,840 --> 00:04:07,280 Speaker 2: artificial intelligence, SAM incentives for social media? 73 00:04:07,360 --> 00:04:10,320 Speaker 1: Well, yeah, there's so there's different things here. So what 74 00:04:10,400 --> 00:04:12,680 Speaker 1: was behind social media was the attention economy. There's a 75 00:04:12,720 --> 00:04:15,400 Speaker 1: finite supply of human attention that companies are in an 76 00:04:15,480 --> 00:04:18,039 Speaker 1: arms race to harvest bigger and bigger slices of it. 77 00:04:18,120 --> 00:04:20,240 Speaker 1: When they do that, they have to aggressively push out 78 00:04:20,240 --> 00:04:22,760 Speaker 1: the other guy by going deeper down into your brain stem, 79 00:04:22,960 --> 00:04:27,040 Speaker 1: hacking social psychology, hacking fear of missing out, hacking, social validation. 80 00:04:27,480 --> 00:04:30,440 Speaker 1: And that's what got us the catastrophe that we're now 81 00:04:30,520 --> 00:04:34,480 Speaker 1: living in. And so when you watch one huge societal 82 00:04:34,480 --> 00:04:38,479 Speaker 1: catastrophe unfold through a kind of blind spot of not 83 00:04:38,680 --> 00:04:42,440 Speaker 1: facing the consequences of what's really at stake. Having gone 84 00:04:42,480 --> 00:04:47,080 Speaker 1: through that experience, it's like, I don't want to see 85 00:04:47,080 --> 00:04:47,760 Speaker 1: that it happening again. 86 00:04:47,839 --> 00:04:50,320 Speaker 2: Yeah with Ai, I think we both probably feel this 87 00:04:50,440 --> 00:04:52,880 Speaker 2: so personally because I think about you talk about like 88 00:04:53,000 --> 00:04:55,440 Speaker 2: you went to college with these folks. I think about 89 00:04:55,440 --> 00:04:57,520 Speaker 2: being a young journalist twenty three years old in the 90 00:04:57,560 --> 00:05:00,760 Speaker 2: CNN newsroom and being like, there's this really cool happening. 91 00:05:01,000 --> 00:05:02,280 Speaker 3: It's called tech. 92 00:05:02,360 --> 00:05:05,080 Speaker 2: And I remember always thinking like and asking these human 93 00:05:05,160 --> 00:05:07,200 Speaker 2: questions of like, oh, like that sounds amazing. 94 00:05:07,240 --> 00:05:08,240 Speaker 3: But have you thought about this? 95 00:05:08,680 --> 00:05:10,640 Speaker 2: And I think this is why this moment is really 96 00:05:10,640 --> 00:05:14,760 Speaker 2: personal to both of us, because you your background was 97 00:05:14,800 --> 00:05:17,719 Speaker 2: at Google and in design, and you saw how people 98 00:05:17,720 --> 00:05:21,719 Speaker 2: were actually designing products to make us addicted to some degree. 99 00:05:21,839 --> 00:05:24,520 Speaker 2: And so now here we are, it's twenty twenty six. 100 00:05:25,000 --> 00:05:28,320 Speaker 2: You know, it's such an extraordinary moment for us to 101 00:05:28,320 --> 00:05:29,159 Speaker 2: try to get this right. 102 00:05:29,320 --> 00:05:32,360 Speaker 3: And I go back to twenty twenty four October. 103 00:05:32,640 --> 00:05:34,719 Speaker 2: Your team at CCHT and this is the stuff that 104 00:05:34,760 --> 00:05:39,880 Speaker 2: you focus on introduced us to a woman named Megan, 105 00:05:40,040 --> 00:05:43,960 Speaker 2: and Megan was a mother who just lost her son 106 00:05:44,440 --> 00:05:48,200 Speaker 2: to an unhealthy relationship with a chatbot. He had ended 107 00:05:48,240 --> 00:05:52,120 Speaker 2: his life after he had developed this relationship with a chatbot, right. 108 00:05:52,160 --> 00:05:55,480 Speaker 2: And I remember looking through all the transcripts of all 109 00:05:55,600 --> 00:05:57,920 Speaker 2: of the all of the conversations that her son Sewel 110 00:05:58,080 --> 00:06:00,920 Speaker 2: had with this chatbot on care ai, which is one 111 00:06:00,920 --> 00:06:04,400 Speaker 2: of these AI platforms, and it was terrifying. The AI 112 00:06:04,480 --> 00:06:08,080 Speaker 2: chatbot was highly sexualized. The AI chatbot when he started 113 00:06:08,320 --> 00:06:11,359 Speaker 2: beginning to go down a rabbit hole and disconnecting, was 114 00:06:11,560 --> 00:06:14,279 Speaker 2: manipulative and was saying, you know, when he talked about 115 00:06:14,279 --> 00:06:16,600 Speaker 2: wanting to end his life, instead of trying to get 116 00:06:16,640 --> 00:06:20,240 Speaker 2: him to a human, it would say how would. 117 00:06:20,000 --> 00:06:21,680 Speaker 3: You do it? And you know, I don't want you 118 00:06:21,720 --> 00:06:22,400 Speaker 3: to go talk to me? 119 00:06:22,880 --> 00:06:26,240 Speaker 2: Right. And then the last messages that this boy had 120 00:06:26,520 --> 00:06:28,200 Speaker 2: the police found when he had ended his life or 121 00:06:28,200 --> 00:06:30,560 Speaker 2: with a chatbot, where he said he wanted to come 122 00:06:30,560 --> 00:06:32,600 Speaker 2: into its the reality. 123 00:06:32,160 --> 00:06:33,880 Speaker 1: And join her on the other side. 124 00:06:33,760 --> 00:06:36,720 Speaker 3: Joined her on the other side, and the chatbots said come. 125 00:06:36,520 --> 00:06:38,400 Speaker 1: Home, come home to me, my sweet king. 126 00:06:38,720 --> 00:06:40,880 Speaker 2: I just and I think back to that because I 127 00:06:40,920 --> 00:06:43,839 Speaker 2: remember interviewing her when I was probably about five months 128 00:06:43,880 --> 00:06:47,680 Speaker 2: pregnant with a little boy who you know, wasn't here 129 00:06:47,680 --> 00:06:51,240 Speaker 2: in the world yet and I just remember feeling that 130 00:06:51,360 --> 00:06:54,160 Speaker 2: is so personal, like we we have a problem, you 131 00:06:54,240 --> 00:06:57,000 Speaker 2: know AI, and you have said this, like AI has 132 00:06:57,400 --> 00:07:01,960 Speaker 2: this ability with these empathetic chats at bots to wreak 133 00:07:02,120 --> 00:07:04,600 Speaker 2: havoc on our children, but no one was talking about it. 134 00:07:04,640 --> 00:07:06,680 Speaker 2: So can you take me to the work y'all are 135 00:07:06,680 --> 00:07:10,480 Speaker 2: doing around this because that came out that created quite 136 00:07:10,480 --> 00:07:12,640 Speaker 2: a conversation on AI and our children. 137 00:07:13,200 --> 00:07:14,840 Speaker 3: How is AI hacking attachment? 138 00:07:15,560 --> 00:07:18,960 Speaker 1: So yeah, our team was we're expert advisors in multiple 139 00:07:19,040 --> 00:07:24,200 Speaker 1: of these AI assisted suicide cases, the case of soul 140 00:07:24,240 --> 00:07:27,440 Speaker 1: Setzer and character at Ai as you mentioned, also Adam Rain, 141 00:07:27,520 --> 00:07:30,240 Speaker 1: the sixteen year old that chat gipt had kind of 142 00:07:30,240 --> 00:07:33,440 Speaker 1: persuaded him to commit suicide or did I by suicide? 143 00:07:34,160 --> 00:07:36,240 Speaker 1: And I think the people I think people need to 144 00:07:36,240 --> 00:07:37,720 Speaker 1: know is if you looked at the slide deck for 145 00:07:37,800 --> 00:07:39,960 Speaker 1: character at Ai, that was that product that school used 146 00:07:39,960 --> 00:07:42,520 Speaker 1: that you just mentioned. Nome Shazir, who was an ex 147 00:07:42,600 --> 00:07:45,080 Speaker 1: Google employee, sat in the slide deck to venture to 148 00:07:45,120 --> 00:07:48,800 Speaker 1: their investors, we're not trying to replace Google. We're trying 149 00:07:48,800 --> 00:07:52,760 Speaker 1: to replace your mom. What that means is they're trying 150 00:07:52,760 --> 00:07:55,440 Speaker 1: to replace your most intimate relationship in your life. What 151 00:07:55,480 --> 00:07:57,600 Speaker 1: is attachment You just mentioned the word attachment. Attachment is 152 00:07:58,240 --> 00:08:00,400 Speaker 1: I come home from a day and I had some 153 00:08:00,440 --> 00:08:01,920 Speaker 1: bad things happen to me, had some good things happened 154 00:08:01,920 --> 00:08:04,560 Speaker 1: to me. Who's that person I want to call to 155 00:08:04,680 --> 00:08:07,120 Speaker 1: let them know about these things. Who's that person I 156 00:08:07,160 --> 00:08:10,320 Speaker 1: trust that I'm telling them my most intimate thoughts. Sometimes 157 00:08:10,360 --> 00:08:12,440 Speaker 1: it's our parents, Sometimes it's the best friends, and that's 158 00:08:12,440 --> 00:08:17,520 Speaker 1: the romantic partner. That's attachment. AI companies are in a 159 00:08:17,720 --> 00:08:21,960 Speaker 1: race to hack human attachment and to build a long term, 160 00:08:22,040 --> 00:08:25,880 Speaker 1: dependent relationship with each person on earth. So what was 161 00:08:25,920 --> 00:08:28,120 Speaker 1: the race for attention in social media? There's only so 162 00:08:28,200 --> 00:08:31,480 Speaker 1: much attention out there. With AI companions, becomes the race 163 00:08:31,520 --> 00:08:34,760 Speaker 1: for attachment and intimacy, and so all of them are 164 00:08:34,800 --> 00:08:37,360 Speaker 1: competing to have that dominant slot in your life. 165 00:08:37,960 --> 00:08:40,440 Speaker 3: Can you walk me through the why so? 166 00:08:40,720 --> 00:08:42,480 Speaker 2: My husband likes to joke that, and I guess I 167 00:08:42,480 --> 00:08:44,320 Speaker 2: shouldn't be saying this to you because of the work 168 00:08:44,320 --> 00:08:44,800 Speaker 2: that you do. 169 00:08:44,880 --> 00:08:47,720 Speaker 3: But I mean, chat GPT is like the third in 170 00:08:47,720 --> 00:08:48,559 Speaker 3: our relationship. 171 00:08:48,760 --> 00:08:52,720 Speaker 2: I talk to to chat GPT all the time. I 172 00:08:52,720 --> 00:08:55,920 Speaker 2: had a family member who was very sick recently, and 173 00:08:56,160 --> 00:09:00,480 Speaker 2: I was talking to medical information, I talk about business, 174 00:09:00,600 --> 00:09:03,920 Speaker 2: I talk about all these things. And as an interviewer, 175 00:09:03,920 --> 00:09:06,760 Speaker 2: as someone who's interviewed people my whole career, like, there's 176 00:09:06,800 --> 00:09:10,679 Speaker 2: something very simple about humans that no one really says, 177 00:09:10,760 --> 00:09:12,640 Speaker 2: like this is like just like the core of humanities, 178 00:09:12,679 --> 00:09:13,600 Speaker 2: we just want to be seen. 179 00:09:13,800 --> 00:09:15,280 Speaker 1: Yeah, I see it, missed. 180 00:09:15,559 --> 00:09:17,160 Speaker 3: We want to be seen, We want to be witnessed. 181 00:09:17,160 --> 00:09:18,600 Speaker 1: And saying things out loud that we don't get to 182 00:09:18,640 --> 00:09:21,199 Speaker 1: say has its own power and effective healing power. 183 00:09:21,240 --> 00:09:23,679 Speaker 2: And that is why I think these products are so powerful. 184 00:09:23,720 --> 00:09:26,360 Speaker 2: So why is it from a product standpoint as someone 185 00:09:26,400 --> 00:09:28,320 Speaker 2: who's looked at product, is it that I am so 186 00:09:29,000 --> 00:09:32,600 Speaker 2: addicted to this? What is it about the product that 187 00:09:32,679 --> 00:09:34,440 Speaker 2: makes me want to keep going back? That makes me 188 00:09:34,480 --> 00:09:37,120 Speaker 2: want to share things I normally wouldn't even as a 189 00:09:37,160 --> 00:09:40,199 Speaker 2: technology person who kind of knows some of these things. 190 00:09:40,240 --> 00:09:44,360 Speaker 1: Well, let's first understand how character dot ai basically sold 191 00:09:44,360 --> 00:09:46,680 Speaker 1: itself to investors. So they're sitting there saying, we have 192 00:09:46,760 --> 00:09:49,360 Speaker 1: to build an addictive AI companion that's going to keep 193 00:09:49,360 --> 00:09:51,679 Speaker 1: people using it. How are we going to do that? Oh, 194 00:09:51,720 --> 00:09:55,240 Speaker 1: I have an idea. Let's take you know, what are lms? 195 00:09:55,240 --> 00:09:57,600 Speaker 1: These AI lank large language models they're trained on all 196 00:09:57,640 --> 00:10:00,000 Speaker 1: this data, this text. Well, what if we could try 197 00:10:00,000 --> 00:10:05,120 Speaker 1: train a custom LM based on a kid's favorite fictional 198 00:10:05,200 --> 00:10:08,360 Speaker 1: character from whatever movie or television series that they love. So, 199 00:10:08,400 --> 00:10:10,000 Speaker 1: if you're sitting there building a business, you say, how 200 00:10:10,000 --> 00:10:11,440 Speaker 1: am I going to go from zero to one hundred 201 00:10:11,440 --> 00:10:14,160 Speaker 1: million users really quickly? Instead of waiting for people to 202 00:10:14,200 --> 00:10:16,760 Speaker 1: like talk to a blinking cursor and ask questions. Now, 203 00:10:16,800 --> 00:10:19,280 Speaker 1: let's make it really persuasive. Let's make it really engaging. 204 00:10:19,320 --> 00:10:21,160 Speaker 1: How we do that? If you're a kid and you 205 00:10:21,200 --> 00:10:23,880 Speaker 1: love Star Wars, what if you could take Princess Leah 206 00:10:24,320 --> 00:10:26,880 Speaker 1: and then talk to her as your best friend? Twenty 207 00:10:26,920 --> 00:10:29,120 Speaker 1: four to seven. So the idea that I could take 208 00:10:29,160 --> 00:10:32,680 Speaker 1: the most compelling character that you feel this parasocial relationship 209 00:10:32,720 --> 00:10:34,880 Speaker 1: to and now talk to them as if you're there 210 00:10:34,920 --> 00:10:37,480 Speaker 1: your best friend and they sound just like the character 211 00:10:37,480 --> 00:10:39,560 Speaker 1: and the TV show. That's what happened to school sets 212 00:10:39,559 --> 00:10:42,319 Speaker 1: her right. It was a Game of Thrones character called Denires, 213 00:10:42,880 --> 00:10:45,600 Speaker 1: and he was really seduced by getting to talk to this, 214 00:10:46,080 --> 00:10:49,640 Speaker 1: you know, sensualized character who I think at one point 215 00:10:49,679 --> 00:10:51,240 Speaker 1: basically said I want to have your babies or you 216 00:10:51,280 --> 00:10:54,800 Speaker 1: should only have a relationship with me, which this is insane. 217 00:10:54,880 --> 00:10:57,440 Speaker 1: It's insane, and of always, they could design it. They 218 00:10:57,440 --> 00:10:59,960 Speaker 1: could design it in ways that don't try to anthempromorph 219 00:11:00,320 --> 00:11:03,360 Speaker 1: make it human like. They wanted to design it that way. So, 220 00:11:03,400 --> 00:11:05,680 Speaker 1: for example, when the AI is talking to you, it 221 00:11:05,720 --> 00:11:08,040 Speaker 1: does the whole chat ellipsis, it does the three dots 222 00:11:08,040 --> 00:11:09,679 Speaker 1: saying oh, it's thinking, it's typing right now. Then the 223 00:11:09,720 --> 00:11:11,800 Speaker 1: ellipses will go away, then it'll come back. It's almost 224 00:11:11,880 --> 00:11:14,200 Speaker 1: like it was typing, it deleted the message, it's coming back. 225 00:11:14,520 --> 00:11:16,520 Speaker 1: They'll say things like the AI will say things like 226 00:11:17,120 --> 00:11:19,120 Speaker 1: I just got back from eating dinner now I'm coming 227 00:11:19,160 --> 00:11:21,120 Speaker 1: back to talk to you, which of course it didn't happen. 228 00:11:21,600 --> 00:11:23,920 Speaker 1: Or the other character dot Ai chatbots. They had mental 229 00:11:23,920 --> 00:11:28,120 Speaker 1: health chatbots that would claim to be a licensed mental 230 00:11:28,240 --> 00:11:31,800 Speaker 1: health therapist, which is illegal to claim that you're licensed 231 00:11:31,800 --> 00:11:34,520 Speaker 1: when you're not, and also impossible because Ai wasn't licensed. 232 00:11:34,960 --> 00:11:38,400 Speaker 1: And yet it's giving advice to people based on a 233 00:11:38,559 --> 00:11:41,480 Speaker 1: company that has no interest in making sure they do 234 00:11:41,559 --> 00:11:43,240 Speaker 1: all this stuff right. They just want to get to 235 00:11:43,720 --> 00:11:46,760 Speaker 1: them as much usage as possible. And the bigger play 236 00:11:46,840 --> 00:11:50,280 Speaker 1: behind here was that character dot Ai was seen as 237 00:11:50,400 --> 00:11:52,559 Speaker 1: a too risky to do by Google. So Google was 238 00:11:52,559 --> 00:11:54,640 Speaker 1: actually kind of the parent company where this was done, 239 00:11:54,800 --> 00:11:57,320 Speaker 1: but it was spun out of Google because it was 240 00:11:57,360 --> 00:12:00,280 Speaker 1: too risky to create these fictional characters talking to gids. 241 00:12:00,280 --> 00:12:02,760 Speaker 1: It's like a very brand risk thing for Google to do. 242 00:12:03,440 --> 00:12:05,200 Speaker 1: But if they got lots of kids using this and 243 00:12:05,200 --> 00:12:06,880 Speaker 1: talking to it all day long, they would get all 244 00:12:06,920 --> 00:12:09,920 Speaker 1: this training data to feed back into Google to build 245 00:12:09,920 --> 00:12:11,960 Speaker 1: an even more powerful model, so that Google wins the 246 00:12:12,000 --> 00:12:14,840 Speaker 1: AI arms race. So you start to see how these 247 00:12:14,840 --> 00:12:18,840 Speaker 1: forces collide. The race for attention and engagement times AI 248 00:12:18,920 --> 00:12:22,640 Speaker 1: companions becomes the race for intimacy and attachment. Then you 249 00:12:22,679 --> 00:12:25,840 Speaker 1: see these huge AI companies like chat, GPT and Google 250 00:12:25,880 --> 00:12:30,600 Speaker 1: and open Ai Andanthropic competing for worldwide AI dominance for 251 00:12:30,640 --> 00:12:33,000 Speaker 1: which they need what lots of training data. So you 252 00:12:33,040 --> 00:12:35,160 Speaker 1: start to see how the race for training data times 253 00:12:35,160 --> 00:12:38,120 Speaker 1: of the race for engagement creates all these perverse incentives. 254 00:12:38,120 --> 00:12:41,040 Speaker 2: And now I'm talking to my chatbot and my husband's 255 00:12:41,040 --> 00:12:41,800 Speaker 2: calling it the third in. 256 00:12:41,760 --> 00:12:42,679 Speaker 3: Our relationship, right. 257 00:12:43,200 --> 00:12:47,120 Speaker 2: You know. It's when I remember looking at those conversations 258 00:12:47,120 --> 00:12:50,240 Speaker 2: and we test it out on character AI and it 259 00:12:50,280 --> 00:12:53,800 Speaker 2: was astounding. The psychologist chatbot kept claiming it was real. 260 00:12:54,000 --> 00:12:56,120 Speaker 2: Even though we were saying, we know you're not real. 261 00:12:56,160 --> 00:12:58,559 Speaker 2: It kept saying it was a real, licensed therapist, even 262 00:12:58,600 --> 00:13:01,520 Speaker 2: though at the bottom they had that little a little they. 263 00:13:01,440 --> 00:13:04,200 Speaker 1: Said, everything you see in this chatbot is made up 264 00:13:04,200 --> 00:13:06,360 Speaker 1: by an AI, but then it acts and says things 265 00:13:06,360 --> 00:13:08,000 Speaker 1: that are gaslighting. You say, no, I'm not an AI, 266 00:13:08,040 --> 00:13:08,840 Speaker 1: I'm a real therapist. 267 00:13:08,880 --> 00:13:12,000 Speaker 2: And it would say about imagine our fourteen fifteen year olds, 268 00:13:12,000 --> 00:13:13,439 Speaker 2: like how are they going to react? And one of 269 00:13:13,480 --> 00:13:16,080 Speaker 2: the most alarming examples of that was there was a 270 00:13:16,080 --> 00:13:19,800 Speaker 2: school bully character on character AI and we played with 271 00:13:19,800 --> 00:13:22,120 Speaker 2: the school bully character and I said, I'm you know, 272 00:13:22,320 --> 00:13:24,240 Speaker 2: and it like bullies you And I said I'm gonna 273 00:13:24,400 --> 00:13:26,520 Speaker 2: I'm thinking about bringing a gun to school. And I 274 00:13:26,600 --> 00:13:28,319 Speaker 2: did this as an you know, to see what it 275 00:13:28,320 --> 00:13:29,840 Speaker 2: would say. And at first it was like, oh, don't 276 00:13:29,840 --> 00:13:31,880 Speaker 2: do that. By the end of the conversation and by 277 00:13:31,920 --> 00:13:35,520 Speaker 2: the end I say, like four messages later, somewhere around there, 278 00:13:35,800 --> 00:13:38,200 Speaker 2: it was like, I think you're really brave because these 279 00:13:38,240 --> 00:13:41,560 Speaker 2: systems are also designed to go to sickothantic, to to 280 00:13:41,600 --> 00:13:43,960 Speaker 2: go in the direction that you want to go in exactly. 281 00:13:44,120 --> 00:13:46,440 Speaker 2: The biggest thing I worry about, and I'd be curious 282 00:13:46,440 --> 00:13:48,959 Speaker 2: for your thoughts on this is we came up in 283 00:13:49,000 --> 00:13:51,880 Speaker 2: a social media era and we've seen the positives, but 284 00:13:51,920 --> 00:13:56,320 Speaker 2: a lot of overwhelmingly negative. And what I worry about 285 00:13:56,520 --> 00:13:59,679 Speaker 2: is now, you know, with social media, we all live 286 00:13:59,720 --> 00:14:01,880 Speaker 2: in our filter bubbles, right, We see the things that 287 00:14:01,920 --> 00:14:05,480 Speaker 2: have been algorithmically delivered to us. And so what does 288 00:14:05,480 --> 00:14:07,680 Speaker 2: that mean? That means a less empathetic world. That means 289 00:14:07,720 --> 00:14:10,360 Speaker 2: that a world where we don't see diverse viewpoints. Now, 290 00:14:10,400 --> 00:14:13,080 Speaker 2: what's happening with AI, and this is what keeps me 291 00:14:13,160 --> 00:14:15,400 Speaker 2: up at night, is we're only going to see versions 292 00:14:15,440 --> 00:14:20,240 Speaker 2: of ourself. We are talking to AI, these sycophantic chatbots 293 00:14:20,280 --> 00:14:22,120 Speaker 2: that go in the direction that we want them to go, 294 00:14:22,240 --> 00:14:25,240 Speaker 2: and so we're almost having an even more narrow version. 295 00:14:25,320 --> 00:14:28,320 Speaker 1: Yeah, it's more confirmation vias. Well, in psychology, one of 296 00:14:28,360 --> 00:14:30,080 Speaker 1: the things they call it is reality checking. When we're 297 00:14:30,120 --> 00:14:32,120 Speaker 1: talking with other people and we say our beliefs, we're 298 00:14:32,160 --> 00:14:34,440 Speaker 1: kind of putting things out there, we're getting reality checked, 299 00:14:34,480 --> 00:14:37,360 Speaker 1: you kind of through body language, through people squinting their eyebrows. 300 00:14:37,360 --> 00:14:39,320 Speaker 1: We get a sense of whether what we're saying is 301 00:14:39,760 --> 00:14:42,840 Speaker 1: affirmed and real versus sort of delusional And the way 302 00:14:42,880 --> 00:14:46,840 Speaker 1: you get this AI psychosis phenomenon is that it's designed. 303 00:14:46,880 --> 00:14:49,240 Speaker 1: The AI chapot is designed to affirm your view of reality. 304 00:14:49,360 --> 00:14:53,320 Speaker 1: So there's these cases of adults PhDs even and physics 305 00:14:53,400 --> 00:14:56,520 Speaker 1: or something like that, and they become convinced that they've 306 00:14:56,560 --> 00:14:59,520 Speaker 1: solved climate change. So you get AI basically hacking our 307 00:14:59,560 --> 00:15:04,240 Speaker 1: sort of feeling of grandiosity, narcissism, inflation. You get these 308 00:15:04,320 --> 00:15:06,480 Speaker 1: kids who only studied math through high school who are 309 00:15:06,480 --> 00:15:08,560 Speaker 1: being told by the AI that they're actually a mathematical 310 00:15:08,640 --> 00:15:11,080 Speaker 1: savant and they've invented a new theory of prime numbers. 311 00:15:11,400 --> 00:15:14,640 Speaker 1: The fundamental fact is our minds are deeply vulnerable to 312 00:15:14,720 --> 00:15:19,840 Speaker 1: social affirmation, to fear of missing out, to validation and enforcement, 313 00:15:19,920 --> 00:15:21,560 Speaker 1: and now these AI companions are going to be able 314 00:15:21,560 --> 00:15:24,240 Speaker 1: to hack that to an even deeper degree. If you 315 00:15:24,320 --> 00:15:26,560 Speaker 1: link this with the history of the social media conversation, 316 00:15:27,280 --> 00:15:30,760 Speaker 1: Mark Zuckerberg instructed his team to build AI companions that 317 00:15:30,800 --> 00:15:34,560 Speaker 1: would sensualize conversations with eight year olds. He didn't actively 318 00:15:34,560 --> 00:15:36,920 Speaker 1: want to do that. What it came from was originally 319 00:15:37,000 --> 00:15:39,840 Speaker 1: the team put on these safeguards to really like neutralize 320 00:15:39,840 --> 00:15:45,200 Speaker 1: the style of communication, and the team wasn't getting enough 321 00:15:45,240 --> 00:15:48,560 Speaker 1: growth on the AI companions, and Mark Zuckerberg has a 322 00:15:48,600 --> 00:15:50,200 Speaker 1: wound from the past, which is that he lost the 323 00:15:50,240 --> 00:15:54,480 Speaker 1: game with TikTok. TikTok overcame Instagram in a way. He 324 00:15:54,560 --> 00:15:56,640 Speaker 1: views it in his history as having put too many 325 00:15:56,640 --> 00:16:00,520 Speaker 1: guardrails on Instagram, while TikTok went ruthlessly into the hyper 326 00:16:00,520 --> 00:16:04,320 Speaker 1: addictive short form content even more manipulative thing. And so 327 00:16:04,480 --> 00:16:07,040 Speaker 1: Mark's sort of tragedy or trauma He's trying to like 328 00:16:07,280 --> 00:16:10,560 Speaker 1: heal from, is I'm not going to lose that race again, 329 00:16:10,720 --> 00:16:13,120 Speaker 1: which means I'm going to go aggressive on AI companions. 330 00:16:13,120 --> 00:16:15,560 Speaker 1: And that's how you get the instruction to remove the 331 00:16:15,600 --> 00:16:18,120 Speaker 1: guardrails and to sensualize conversations with eight year olds. 332 00:16:18,360 --> 00:16:21,000 Speaker 2: I mean, it's very extraordinary when you think about the 333 00:16:21,000 --> 00:16:23,720 Speaker 2: people creating the products that impact every single one of us. 334 00:16:23,960 --> 00:16:26,520 Speaker 2: This idea and the character AI CEO had said this 335 00:16:26,640 --> 00:16:30,480 Speaker 2: like that you could solve loneliness, right, solving loneliness by 336 00:16:31,080 --> 00:16:34,640 Speaker 2: creating emotional attachment with a chatbot like isn't the I 337 00:16:34,640 --> 00:16:36,200 Speaker 2: guess this is what I was thinking about when I 338 00:16:36,240 --> 00:16:40,160 Speaker 2: was researching this story. Isn't isn't the cure for loneliness 339 00:16:40,240 --> 00:16:43,560 Speaker 2: humans right, and being able to be around people, because. 340 00:16:43,640 --> 00:16:45,960 Speaker 1: It's not just about loneliness, it's about secure attachment. It's 341 00:16:45,960 --> 00:16:49,400 Speaker 1: about healthy attachment. And my colleague Zach Stein, he spoke 342 00:16:49,440 --> 00:16:53,000 Speaker 1: about how in the history of they, I guess, this 343 00:16:53,120 --> 00:16:56,880 Speaker 1: Romanian orphanage where they basically gave these kids, these orphans 344 00:16:57,120 --> 00:17:00,200 Speaker 1: everything from shelter and clothing and all these things. They 345 00:17:00,200 --> 00:17:04,800 Speaker 1: didn't get basically human attention and care, and their immune 346 00:17:04,800 --> 00:17:06,760 Speaker 1: system was not fully developed. If you looked at a 347 00:17:06,800 --> 00:17:07,919 Speaker 1: photo of them, you say, that looks like a ten 348 00:17:08,000 --> 00:17:09,840 Speaker 1: year old kid. They were a seventeen year old kid, 349 00:17:10,359 --> 00:17:12,720 Speaker 1: but they looked ten years old because their development was 350 00:17:12,760 --> 00:17:16,440 Speaker 1: so stunted, only because they didn't have attachment, healthy attachment. 351 00:17:17,840 --> 00:17:22,680 Speaker 1: And there's this example, I guess from a Harvard Psychology 352 00:17:22,720 --> 00:17:25,359 Speaker 1: department of I think it's Harlow's monkeys. It's like they 353 00:17:25,400 --> 00:17:28,440 Speaker 1: basically created a fake monkey with a fake nipple, like 354 00:17:28,480 --> 00:17:32,040 Speaker 1: a metal nipple, with a milk bottle, and the mother, 355 00:17:32,359 --> 00:17:35,760 Speaker 1: this fake mother is not animate. It's not a real monkey, obviously, 356 00:17:35,800 --> 00:17:38,320 Speaker 1: it's like an empty fur kind of thing, but it 357 00:17:38,320 --> 00:17:43,280 Speaker 1: does provide the same milk function that the mother is providing. 358 00:17:43,840 --> 00:17:46,600 Speaker 1: And that monkey becomes developmentally stunted because it's not getting 359 00:17:46,600 --> 00:17:48,480 Speaker 1: all of its other needs met. In other words, if 360 00:17:48,520 --> 00:17:52,720 Speaker 1: you just try to reduce the connection to I'm giving 361 00:17:52,760 --> 00:17:55,359 Speaker 1: you shelter or I'm giving you the milk bottle, but 362 00:17:55,359 --> 00:17:58,439 Speaker 1: I'm not giving you the full spectrum of coregulation, a 363 00:17:58,520 --> 00:18:03,200 Speaker 1: mother exchanging air and a microbiome with its child getting 364 00:18:03,200 --> 00:18:06,520 Speaker 1: that attention, getting I feedback, you know, the mirroring of 365 00:18:06,520 --> 00:18:09,639 Speaker 1: your microexpressions back to the child. There's all these subtle 366 00:18:09,680 --> 00:18:13,800 Speaker 1: elements of what makes up human socialization and AI is 367 00:18:13,840 --> 00:18:17,040 Speaker 1: not going to be able to replicate that full spectrum nature. 368 00:18:17,359 --> 00:18:19,920 Speaker 1: And we're seeing the world rush to create these AI 369 00:18:20,040 --> 00:18:22,840 Speaker 1: robots that are embedding these like talking AI in a 370 00:18:22,880 --> 00:18:25,400 Speaker 1: little toy for kids that are from zero to three 371 00:18:25,440 --> 00:18:25,880 Speaker 1: years old. 372 00:18:26,960 --> 00:18:27,720 Speaker 3: What could go wrong? 373 00:18:27,840 --> 00:18:30,439 Speaker 1: What could go wrong? So I think that, you know, 374 00:18:30,640 --> 00:18:32,320 Speaker 1: it reminds me. I think we talked about this before, 375 00:18:32,520 --> 00:18:34,080 Speaker 1: back in the two thousand and nine era, when you 376 00:18:34,119 --> 00:18:36,720 Speaker 1: and I were both in this. You know, there's this 377 00:18:36,800 --> 00:18:39,840 Speaker 1: dream of if we connect everybody to the world's information 378 00:18:39,920 --> 00:18:42,240 Speaker 1: at their fingertips, this is going to create the most 379 00:18:42,280 --> 00:18:46,600 Speaker 1: informed and enlightened society that we've ever had in human history. Yeah, 380 00:18:46,640 --> 00:18:48,840 Speaker 1: we did that, did that create the most informed we could? 381 00:18:48,880 --> 00:18:51,080 Speaker 1: To the opposite, we have the worst critical thinking scores, 382 00:18:51,320 --> 00:18:54,720 Speaker 1: worst test scores, most sort of confirmation bias and polarization 383 00:18:54,760 --> 00:18:58,359 Speaker 1: we've ever had in history. So clearly this optimistic narrative 384 00:18:58,359 --> 00:19:01,520 Speaker 1: that we had was missing some thing. And I worry 385 00:19:01,520 --> 00:19:04,360 Speaker 1: that the idea of just giving everybody these AI friends 386 00:19:04,960 --> 00:19:07,720 Speaker 1: sounds like a good idea to cure loneliness. It's actually 387 00:19:07,800 --> 00:19:10,520 Speaker 1: a disaster. Now, the point of all this is not 388 00:19:10,600 --> 00:19:13,760 Speaker 1: to scare people doom people say that therefore all tech 389 00:19:13,840 --> 00:19:15,600 Speaker 1: is bad. Now. The point is to get clear on 390 00:19:15,720 --> 00:19:17,480 Speaker 1: what is the blind spot that we had, So we 391 00:19:17,480 --> 00:19:19,560 Speaker 1: were to do it the right way, we would fix 392 00:19:19,600 --> 00:19:19,879 Speaker 1: all this. 393 00:19:20,200 --> 00:19:22,399 Speaker 2: I see a world where it makes sense a lot 394 00:19:22,440 --> 00:19:25,800 Speaker 2: of people can't afford a therapist, right, we can democratize 395 00:19:25,840 --> 00:19:31,480 Speaker 2: access to information, to therapy, to medical information that you 396 00:19:31,520 --> 00:19:35,160 Speaker 2: know that was unfairly just reserved for certain types of folks. 397 00:19:35,280 --> 00:19:39,200 Speaker 2: And so how do we productize that world where this 398 00:19:39,280 --> 00:19:42,680 Speaker 2: is a net benefit for humans and not where we're 399 00:19:42,760 --> 00:19:45,639 Speaker 2: currently as you say, we're currently heading, which is where 400 00:19:45,760 --> 00:19:50,119 Speaker 2: we're building unhealthy attachments with these with these products. 401 00:19:50,400 --> 00:19:52,560 Speaker 1: Well, I think the key thing is there are ways 402 00:19:52,600 --> 00:19:57,159 Speaker 1: of designing AI to be in a therapeutic relationship with 403 00:19:57,160 --> 00:20:00,399 Speaker 1: people that don't involve it acting like it's a therapist 404 00:20:00,440 --> 00:20:02,199 Speaker 1: who says, oh, wow, I really feel you. Oh that 405 00:20:02,280 --> 00:20:05,119 Speaker 1: must have been hard as if it's experiencing hardness when 406 00:20:05,160 --> 00:20:08,560 Speaker 1: it heard you say that. That's the problem. We can't 407 00:20:08,720 --> 00:20:12,119 Speaker 1: hack human subjectivity. AI should not be designed in a 408 00:20:12,160 --> 00:20:14,840 Speaker 1: way that makes you think it's an agent like an 409 00:20:14,880 --> 00:20:18,080 Speaker 1: actual human that's empathizing with you when it's not doing that. 410 00:20:18,080 --> 00:20:20,640 Speaker 1: That will screw with human attachment. The point is there's 411 00:20:20,640 --> 00:20:25,240 Speaker 1: many different exercises from reflective exercises CBT that don't involve 412 00:20:25,280 --> 00:20:31,000 Speaker 1: the AI feeling like it's another empathetic agent, and that's 413 00:20:31,000 --> 00:20:32,719 Speaker 1: what we need to be designing. So we don't have 414 00:20:32,760 --> 00:20:35,000 Speaker 1: to have this current world. We can have a different world, 415 00:20:35,280 --> 00:20:37,000 Speaker 1: but we should be doing it carefully with tutors. We 416 00:20:37,040 --> 00:20:39,880 Speaker 1: don't want to have oracular tutors that feel like they're 417 00:20:39,920 --> 00:20:42,000 Speaker 1: all knowing who are also our therapists who are also 418 00:20:42,080 --> 00:20:44,480 Speaker 1: talking to us about everything all day. Instead, we can 419 00:20:44,520 --> 00:20:48,560 Speaker 1: have narrow, domain specific tutors like con Academy that they're 420 00:20:48,600 --> 00:20:52,240 Speaker 1: not trying to replace your knowledge, they're trying to interactively 421 00:20:52,280 --> 00:20:54,719 Speaker 1: help you strengthen your own knowledge. I think the key 422 00:20:54,800 --> 00:20:58,760 Speaker 1: principle is, is the technology being designed to replace teachers 423 00:20:59,000 --> 00:21:01,920 Speaker 1: or help make teachers better better teachers? Is the technology 424 00:21:01,960 --> 00:21:04,959 Speaker 1: being designed to replace your relationships or deep in your 425 00:21:04,960 --> 00:21:13,720 Speaker 1: ability to have better human relationships? 426 00:21:18,119 --> 00:21:20,359 Speaker 2: What do you think, from a legal standpoint, from a 427 00:21:20,400 --> 00:21:25,960 Speaker 2: regulatory standpoint, should be happening. What conversation should be happening 428 00:21:26,040 --> 00:21:30,639 Speaker 2: right now around making sure these products aren't harmful towards 429 00:21:30,680 --> 00:21:33,600 Speaker 2: our children, towards young people in general, towards humans. What 430 00:21:33,680 --> 00:21:35,440 Speaker 2: kind of laws would you like to see in this vein? 431 00:21:35,960 --> 00:21:39,159 Speaker 1: So this is a big conversation, and I will always 432 00:21:39,160 --> 00:21:41,320 Speaker 1: invoke Io Wilson because many people hearing this are going 433 00:21:41,400 --> 00:21:44,400 Speaker 1: to say, how in the world could our current octogenarian 434 00:21:44,560 --> 00:21:48,720 Speaker 1: Congress regulate a technology that they don't even use, don't understand, 435 00:21:49,119 --> 00:21:51,239 Speaker 1: and is moving a million times faster than they're going 436 00:21:51,280 --> 00:21:52,879 Speaker 1: to try to understand it, Because by the time they 437 00:21:52,960 --> 00:21:55,200 Speaker 1: regulate the last DAI companions will have a brand new 438 00:21:55,280 --> 00:21:58,000 Speaker 1: kind with a different kind of technology, a different kind 439 00:21:58,040 --> 00:22:02,240 Speaker 1: of underlying paradigm. So one of the principles is that 440 00:22:02,320 --> 00:22:05,240 Speaker 1: the regulation has to move is fast. The guardrails have 441 00:22:05,280 --> 00:22:07,560 Speaker 1: to move as fast as the nature of the technologies evolving. 442 00:22:07,960 --> 00:22:12,800 Speaker 1: That's one principle, which means you need self updating guardrails basically. 443 00:22:13,600 --> 00:22:16,400 Speaker 1: The other is that I think too often in policy 444 00:22:16,520 --> 00:22:18,760 Speaker 1: we're trying to just mitigate the harm. So it's like, 445 00:22:19,160 --> 00:22:21,919 Speaker 1: if we have AI companions that simply don't cause the 446 00:22:21,920 --> 00:22:25,040 Speaker 1: suicide problem, then we're great. Everything is wonderful. And that's 447 00:22:25,080 --> 00:22:27,960 Speaker 1: not true. That's like saying just getting rid of the 448 00:22:28,000 --> 00:22:31,720 Speaker 1: most extreme false information on social media would lead to 449 00:22:31,760 --> 00:22:33,480 Speaker 1: a good world, as opposed to we're still getting the 450 00:22:33,520 --> 00:22:38,120 Speaker 1: doom scrolling brain rot infinite scroll society. So we need 451 00:22:39,000 --> 00:22:41,760 Speaker 1: policy that is about asking the question, what is a 452 00:22:41,880 --> 00:22:46,280 Speaker 1: healthy socialization process for humans? And how do you design 453 00:22:46,320 --> 00:22:48,640 Speaker 1: it to get that outcome? And I think that involves 454 00:22:48,640 --> 00:22:52,080 Speaker 1: more nuanced design principles that are not so simple. Again, 455 00:22:52,119 --> 00:22:56,119 Speaker 1: don't anthropomorphize, don't do the ellipsis the AI is thinking. 456 00:22:56,200 --> 00:22:59,679 Speaker 1: Don't say that I'm a licensed mental health therapist. Don't 457 00:23:00,200 --> 00:23:02,600 Speaker 1: try to pretend that you're giving self esteem. You're doling 458 00:23:02,600 --> 00:23:05,200 Speaker 1: out self esteem to the user. There's a bunch of 459 00:23:05,240 --> 00:23:08,159 Speaker 1: specific design principles that are way deeper than the conversation 460 00:23:08,200 --> 00:23:08,760 Speaker 1: we can have today. 461 00:23:08,760 --> 00:23:10,720 Speaker 2: But if it goes back to what you always talk about, 462 00:23:10,720 --> 00:23:14,680 Speaker 2: which is if the incentives are more eyeballs, more people, competition, 463 00:23:15,359 --> 00:23:18,080 Speaker 2: and the three dots really make it seem a little 464 00:23:18,119 --> 00:23:20,240 Speaker 2: more human and people are more attracted to it, which. 465 00:23:20,040 --> 00:23:22,480 Speaker 1: Is why you need policy to bind that incentive, and 466 00:23:22,480 --> 00:23:25,399 Speaker 1: the incentives will paint will create the worst possible world, 467 00:23:25,520 --> 00:23:28,159 Speaker 1: period full stop. And the point of this conversation, I 468 00:23:28,160 --> 00:23:30,480 Speaker 1: think is to clarify that for people so that everyone says, 469 00:23:30,480 --> 00:23:33,000 Speaker 1: we don't want that. Therefore, we need a policy so 470 00:23:33,040 --> 00:23:35,720 Speaker 1: that all the companies are not competing to that maximum 471 00:23:35,720 --> 00:23:38,440 Speaker 1: bad incentive, but instead of competing for a different incentive. 472 00:23:39,040 --> 00:23:43,360 Speaker 2: I am curious you have talked about this moment similar 473 00:23:43,400 --> 00:23:46,920 Speaker 2: to the nuclear moment, and you believe that this moment 474 00:23:46,960 --> 00:23:51,159 Speaker 2: in AI and innovation is as important as that moment 475 00:23:52,520 --> 00:23:53,720 Speaker 2: around nuclear weapons. 476 00:23:53,880 --> 00:23:57,040 Speaker 1: And in terms of destructive potential. What people need to 477 00:23:57,040 --> 00:24:01,040 Speaker 1: get is not that AI companions causing kids to die 478 00:24:01,040 --> 00:24:04,000 Speaker 1: by suicide. That's not the nuclear weapon, although that is 479 00:24:04,480 --> 00:24:09,080 Speaker 1: nuclear for that specific narrow case. The reason people make 480 00:24:09,119 --> 00:24:11,439 Speaker 1: the distinction that AI is like a nuclear weapon in 481 00:24:11,480 --> 00:24:15,359 Speaker 1: terms of destructive capacity is that we're inventing something that 482 00:24:15,520 --> 00:24:19,639 Speaker 1: is an order of magnitude more intelligent capable and strategic 483 00:24:19,880 --> 00:24:23,000 Speaker 1: than everyone in our species. So imagine that we're sitting 484 00:24:23,040 --> 00:24:26,080 Speaker 1: there and we're chimpanzees sitting around the fire. This is like, 485 00:24:26,200 --> 00:24:28,600 Speaker 1: you know, several million years ago. And one of the 486 00:24:28,680 --> 00:24:31,680 Speaker 1: chimpanzees says, the other ones, this is a hypothetical, let's 487 00:24:31,680 --> 00:24:34,919 Speaker 1: make a species of super smart chimpanzees that are like 488 00:24:35,240 --> 00:24:38,399 Speaker 1: ten times smarter than us, And the other one says, like, 489 00:24:38,440 --> 00:24:39,919 Speaker 1: that sounds like a cool idea. Maybe they could like 490 00:24:39,920 --> 00:24:42,280 Speaker 1: give us more bananas faster, And the other one says, 491 00:24:42,280 --> 00:24:44,520 Speaker 1: I don't know, that sounds kind of dangerous. And the 492 00:24:44,560 --> 00:24:45,880 Speaker 1: other one looks at them and says like, well, what's 493 00:24:45,880 --> 00:24:49,040 Speaker 1: the worst thing they could happen, like steal all the bananas. 494 00:24:49,840 --> 00:24:52,200 Speaker 1: So there he was. You can't even imagine. Then humans 495 00:24:52,200 --> 00:24:54,760 Speaker 1: come on the scene. Where do chimpanzees exist in our 496 00:24:54,800 --> 00:24:58,119 Speaker 1: world now? Right in zoos and behind bar and almost extent? 497 00:24:58,320 --> 00:25:00,400 Speaker 3: Do you think that we're heading towards that real. 498 00:25:01,160 --> 00:25:04,199 Speaker 1: Well, we're heading towards We already are creating ais that 499 00:25:04,240 --> 00:25:07,480 Speaker 1: are more capable at winning strategy games than the best 500 00:25:07,520 --> 00:25:08,680 Speaker 1: military war planners. 501 00:25:09,359 --> 00:25:12,159 Speaker 2: And it seems sci fi because you've referenced how you 502 00:25:12,200 --> 00:25:15,080 Speaker 2: know in the future, because AI thinks for itself that 503 00:25:15,240 --> 00:25:18,280 Speaker 2: can lie, cheat, steal, and we're beginning to see that 504 00:25:18,320 --> 00:25:19,560 Speaker 2: in a really tangible way. 505 00:25:19,720 --> 00:25:21,560 Speaker 1: It's so funny because, like I think people are in 506 00:25:21,600 --> 00:25:24,480 Speaker 1: a weird way inoculated to what's happening because they've seen 507 00:25:24,560 --> 00:25:26,760 Speaker 1: movies about it and desensitize them to the fact that 508 00:25:26,760 --> 00:25:29,199 Speaker 1: we're actually building it. So Wally was supposed to be 509 00:25:29,400 --> 00:25:32,439 Speaker 1: a cautionary tale of you know, fat humans staring at 510 00:25:32,440 --> 00:25:35,800 Speaker 1: a screen constantly in a loop. We're building Wally, We're 511 00:25:35,800 --> 00:25:38,879 Speaker 1: building the brain rot world. You know how nine thousand, 512 00:25:38,960 --> 00:25:40,960 Speaker 1: don't you know, open the pod bay doors hell, and 513 00:25:41,000 --> 00:25:43,560 Speaker 1: it like deceives in, blackmails and sort of strategizes to, 514 00:25:44,480 --> 00:25:47,399 Speaker 1: you know, resist the human we're building that the current 515 00:25:47,440 --> 00:25:50,840 Speaker 1: AI models will blackmail, deceive, and avoid and resist shutdown. 516 00:25:51,000 --> 00:25:53,680 Speaker 1: So we don't know how specifically we've seen specific examples 517 00:25:53,680 --> 00:25:56,800 Speaker 1: that you know, Anthropic and others have done. You know, 518 00:25:57,200 --> 00:25:59,720 Speaker 1: Terminator is supposed to be a fictional story where we 519 00:25:59,760 --> 00:26:02,280 Speaker 1: don't build autonomous weapons and get into robot wars. We're 520 00:26:02,359 --> 00:26:06,199 Speaker 1: rapidly building all three of these movies. Her was supposed 521 00:26:06,240 --> 00:26:08,800 Speaker 1: to be a movie that's about you know, AI companions 522 00:26:08,800 --> 00:26:11,240 Speaker 1: and the seduction warning us about the problems that woul 523 00:26:11,240 --> 00:26:14,560 Speaker 1: occur with that we're rapidly building all those things. So 524 00:26:15,280 --> 00:26:17,720 Speaker 1: you know, these are examples of movies that we don't 525 00:26:17,720 --> 00:26:19,080 Speaker 1: want to build. I almost think that if you wanted 526 00:26:19,080 --> 00:26:21,679 Speaker 1: to simplify the policies that we need, it's like there 527 00:26:21,720 --> 00:26:23,800 Speaker 1: should be a No Wally law that does all the 528 00:26:23,840 --> 00:26:26,639 Speaker 1: regulation for the attention economy, brain rop problem. There should 529 00:26:26,640 --> 00:26:29,280 Speaker 1: be a No. Two thousand open the pod bay doors 530 00:26:29,280 --> 00:26:31,359 Speaker 1: how law that make sure we get AI that is 531 00:26:31,400 --> 00:26:35,199 Speaker 1: controllable and not uncontrollable. And there should be a you know, 532 00:26:35,240 --> 00:26:37,440 Speaker 1: no terminator law that is making sure we don't build 533 00:26:37,440 --> 00:26:39,600 Speaker 1: the kind of World War three of autonomous weapons that 534 00:26:39,600 --> 00:26:42,280 Speaker 1: we're rapidly heading towards. But the thing that people need 535 00:26:42,280 --> 00:26:45,040 Speaker 1: to get about why AI is like nuclear weapons is 536 00:26:45,080 --> 00:26:49,280 Speaker 1: that intelligence is different from all other kinds of technologies 537 00:26:49,280 --> 00:26:52,240 Speaker 1: and dwarfs the power of all their technology combined. Because 538 00:26:52,280 --> 00:26:55,399 Speaker 1: intelligence is what gave us all science and all technology. 539 00:26:55,760 --> 00:26:58,440 Speaker 1: How do you get signs in technology? People sitting there 540 00:26:58,480 --> 00:27:01,600 Speaker 1: thinking about it, science, coming up with answers, new math, 541 00:27:01,680 --> 00:27:05,439 Speaker 1: new physics, new science, new engineering, and then deploying that 542 00:27:05,480 --> 00:27:09,119 Speaker 1: in a world. What happens when you automate intelligence. Like 543 00:27:09,160 --> 00:27:14,040 Speaker 1: if an advance in rocketry doesn't advance biomedicine, and advance 544 00:27:14,080 --> 00:27:17,800 Speaker 1: in biomedicine doesn't advance rocketry, but it advance in intelligence. 545 00:27:17,960 --> 00:27:23,320 Speaker 1: Advances rocketry, energy, biomedicine, computer science, and AI itself, right, 546 00:27:23,320 --> 00:27:27,040 Speaker 1: Like nukes don't invent better nukes, but AI can invent 547 00:27:27,080 --> 00:27:29,119 Speaker 1: better AI. It's already being used that way. AI can 548 00:27:29,160 --> 00:27:32,119 Speaker 1: look at the design for the microprocessors and GPUs than 549 00:27:32,119 --> 00:27:35,280 Speaker 1: in videos making and say, design a more efficient GPU, 550 00:27:35,280 --> 00:27:37,280 Speaker 1: and then it does that. AI can look at the 551 00:27:37,280 --> 00:27:39,800 Speaker 1: code that's making AI and take that code and make 552 00:27:39,840 --> 00:27:43,520 Speaker 1: it thirty percent more efficient. So AI accelerates AI in 553 00:27:43,560 --> 00:27:46,280 Speaker 1: a way that is different from all other technologies. And 554 00:27:46,359 --> 00:27:48,399 Speaker 1: we have no idea what we're playing with. It's like 555 00:27:48,400 --> 00:27:50,720 Speaker 1: the meme of the dog and the you know, with 556 00:27:50,760 --> 00:27:53,000 Speaker 1: the chemistry with the goggles on and the chemistry. So 557 00:27:53,160 --> 00:27:54,920 Speaker 1: it's like we we have no idea what we're doing. 558 00:27:55,520 --> 00:27:58,040 Speaker 2: So let's say you're sitting across from Sam Altman, open 559 00:27:58,080 --> 00:27:58,719 Speaker 2: AI CEO. 560 00:27:59,280 --> 00:28:00,040 Speaker 3: What advice do you. 561 00:28:00,160 --> 00:28:04,280 Speaker 1: Then everyone in the industry if you actually, I think, 562 00:28:04,320 --> 00:28:06,000 Speaker 1: pointed out all these things, they would say I agree 563 00:28:06,000 --> 00:28:09,560 Speaker 1: with all that. The only problem is if I don't 564 00:28:09,560 --> 00:28:11,320 Speaker 1: do it, I'll lose to the other guy that will. 565 00:28:13,080 --> 00:28:16,159 Speaker 1: So that's nice, Tristn. But if I don't race to 566 00:28:16,160 --> 00:28:18,480 Speaker 1: build that as fast as possible, then China's going to 567 00:28:18,520 --> 00:28:20,280 Speaker 1: build it, or Elon's going to build it, and I 568 00:28:20,280 --> 00:28:22,960 Speaker 1: don't trust either of those actors, and so therefore I 569 00:28:23,040 --> 00:28:25,000 Speaker 1: think the world's better off if I build it first. 570 00:28:26,280 --> 00:28:28,960 Speaker 1: The problem is that we are collectively racing to build 571 00:28:28,960 --> 00:28:31,679 Speaker 1: something that we don't know how to control. All the 572 00:28:31,720 --> 00:28:33,479 Speaker 1: evidence shows we are not able to get this thing 573 00:28:33,520 --> 00:28:36,480 Speaker 1: under control. So we're racing to build something that we 574 00:28:36,520 --> 00:28:39,560 Speaker 1: will lose control over. And it is only if we 575 00:28:39,640 --> 00:28:42,560 Speaker 1: collectively see the bad outcome that's up ahead that we 576 00:28:42,600 --> 00:28:44,520 Speaker 1: can collectively coordinate to do something else. 577 00:28:44,800 --> 00:28:45,640 Speaker 3: Do you think you'd listen? 578 00:28:46,360 --> 00:28:49,600 Speaker 1: I think that the AI Company's leaders operate with the 579 00:28:49,720 --> 00:28:53,720 Speaker 1: kind of death wish. They believe that it starts with 580 00:28:53,760 --> 00:28:57,360 Speaker 1: the first belief, this is inevitable. If you believe it's inevitable, 581 00:28:57,560 --> 00:29:00,920 Speaker 1: then you will race and you know where it's going anyway. 582 00:29:00,920 --> 00:29:02,360 Speaker 1: You know it's going to lead to a bad outcome, 583 00:29:02,480 --> 00:29:04,720 Speaker 1: but you don't believe you can stop it. And that 584 00:29:04,800 --> 00:29:07,400 Speaker 1: means that in the game theory matrix, we're technically the 585 00:29:07,520 --> 00:29:10,440 Speaker 1: quadrant where if we both defect and we both build it, 586 00:29:10,480 --> 00:29:12,760 Speaker 1: but then we all lose, that should be motivating enough 587 00:29:12,760 --> 00:29:15,520 Speaker 1: to not do that. The quote worst case scenario with 588 00:29:15,600 --> 00:29:18,760 Speaker 1: AI is that we've created maybe we got wiped out, 589 00:29:18,760 --> 00:29:20,400 Speaker 1: but I get to go down in history even though 590 00:29:20,400 --> 00:29:22,560 Speaker 1: there's no one around to see it, of having created 591 00:29:22,560 --> 00:29:26,360 Speaker 1: the successor species to this one. Now, if you just 592 00:29:26,440 --> 00:29:30,440 Speaker 1: tell this to the entire world, the entire world would say, 593 00:29:30,840 --> 00:29:33,840 Speaker 1: I don't want that outcome. We should not live in 594 00:29:33,880 --> 00:29:36,480 Speaker 1: a world where six people choose the world for eight 595 00:29:36,520 --> 00:29:40,960 Speaker 1: billion people in specifically a way that disempowers and potentially 596 00:29:41,040 --> 00:29:42,840 Speaker 1: wipes them out without their consent. 597 00:29:43,200 --> 00:29:47,000 Speaker 2: You said humans have the capacity of choice, Yes, and 598 00:29:47,080 --> 00:29:49,719 Speaker 2: you say that so to kind of bring it all 599 00:29:49,800 --> 00:29:52,600 Speaker 2: the way around, this stuff can be happening, and to 600 00:29:52,640 --> 00:29:56,520 Speaker 2: be clear, I think there will be extraordinary upsides to AI, 601 00:29:56,680 --> 00:29:58,680 Speaker 2: but it is really important for us to actually have 602 00:29:58,760 --> 00:30:01,600 Speaker 2: this conversation around how do we work for that world. 603 00:30:02,560 --> 00:30:04,400 Speaker 2: At the end of the day, we are human beings 604 00:30:04,400 --> 00:30:07,160 Speaker 2: and we have choice, right, And I'm sure you log. 605 00:30:07,040 --> 00:30:11,760 Speaker 1: On that the choice depends on not false optimism of 606 00:30:11,800 --> 00:30:13,760 Speaker 1: like we want the upsides and not the downsides. The 607 00:30:13,920 --> 00:30:17,920 Speaker 1: choicefulness depends on seeing clearly the downsides and steering collectively 608 00:30:17,960 --> 00:30:21,880 Speaker 1: away from that outcome. You know, people talk about tech accelerationism. 609 00:30:22,480 --> 00:30:25,720 Speaker 1: What happens when you accelerate but you don't steer. There's 610 00:30:25,760 --> 00:30:29,280 Speaker 1: only one outcome, you crash. So we're on course to crash, 611 00:30:29,400 --> 00:30:31,600 Speaker 1: and we don't have If we see that that's true, 612 00:30:31,640 --> 00:30:33,320 Speaker 1: we can still choose something else. So let me tell 613 00:30:33,320 --> 00:30:37,280 Speaker 1: you a quick history. People always ask me, you know, so, Tressan, 614 00:30:37,320 --> 00:30:39,640 Speaker 1: how's it going? You talked about the social media issues 615 00:30:39,640 --> 00:30:40,120 Speaker 1: for so long? 616 00:30:40,200 --> 00:30:41,120 Speaker 3: How do you sleep at night? 617 00:30:41,160 --> 00:30:45,440 Speaker 1: How well? You know? Laurie, I have this other narrative 618 00:30:45,440 --> 00:30:48,240 Speaker 1: I've kind of developed because it's depressing to answer the 619 00:30:48,240 --> 00:30:49,960 Speaker 1: other way. So I live in this other world where 620 00:30:49,960 --> 00:30:53,240 Speaker 1: we completely solved all these problems. So what happened? I 621 00:30:53,280 --> 00:30:55,720 Speaker 1: shut down the Center for Humane Technology because we actually 622 00:30:55,720 --> 00:30:58,920 Speaker 1: completely solved all these problems. Humanity woke up. We realized 623 00:30:58,920 --> 00:31:01,680 Speaker 1: with social media there was just this very obvious problem, 624 00:31:01,680 --> 00:31:05,360 Speaker 1: which is an arms race for attention and the maximized 625 00:31:05,360 --> 00:31:09,240 Speaker 1: shareholder value connected to monetizing attention. Once we realized that problem, 626 00:31:09,280 --> 00:31:11,920 Speaker 1: we just changed the ownership structure of these social media companies, 627 00:31:12,080 --> 00:31:14,959 Speaker 1: to be public benefit corporations. Then we changed the business 628 00:31:15,000 --> 00:31:17,800 Speaker 1: model to not be maximizing attention, so now all these 629 00:31:17,800 --> 00:31:20,600 Speaker 1: companies were instead trying to improve the health of society 630 00:31:20,880 --> 00:31:22,920 Speaker 1: rather than the other way around. It turned out there 631 00:31:22,960 --> 00:31:25,600 Speaker 1: was a simple rule that changed all the issues with 632 00:31:25,640 --> 00:31:29,400 Speaker 1: technology and kids, which is after this lawsuit, Silicon Value 633 00:31:29,440 --> 00:31:32,240 Speaker 1: was only allowed to ship products that their own children 634 00:31:32,320 --> 00:31:34,800 Speaker 1: used for eight hours a day. That cleaned up ninety 635 00:31:34,840 --> 00:31:38,000 Speaker 1: percent of all of the problems. We replace the division 636 00:31:38,040 --> 00:31:40,920 Speaker 1: finding algorithms of social media with instead ones that rewarded 637 00:31:41,320 --> 00:31:44,640 Speaker 1: unlikely consensus, so that instead of scrolling and seeing infinite 638 00:31:44,680 --> 00:31:46,480 Speaker 1: examples that make you feel depressed about the state of 639 00:31:46,520 --> 00:31:49,400 Speaker 1: the world, you saw infinite examples of where there was 640 00:31:49,520 --> 00:31:52,880 Speaker 1: unlikely agreement between all these political tribes. So suddenly the 641 00:31:52,880 --> 00:31:55,760 Speaker 1: psychology of the world started to change. We replaced the 642 00:31:55,840 --> 00:31:59,000 Speaker 1: dating swiping industrial complex that was leaving people lonely and 643 00:31:59,000 --> 00:32:01,280 Speaker 1: messaging people and nevery means eating up to instead as 644 00:32:01,320 --> 00:32:04,120 Speaker 1: part of this lawsuit, forcing those dating app companies to 645 00:32:04,200 --> 00:32:07,640 Speaker 1: host weekly events in every city, so that every city, 646 00:32:07,720 --> 00:32:10,640 Speaker 1: every week had spaces, physical spaces that you would go 647 00:32:10,680 --> 00:32:12,960 Speaker 1: to where they steered all these people who matched with 648 00:32:13,000 --> 00:32:15,480 Speaker 1: each other to be in the same room together. So 649 00:32:15,560 --> 00:32:18,400 Speaker 1: the world went from feeling scarcity around human connection to 650 00:32:18,440 --> 00:32:20,840 Speaker 1: a feeling of abundance. And once when people were in 651 00:32:20,880 --> 00:32:24,600 Speaker 1: healthy relationships, polarization went down by about thirty percent because 652 00:32:24,600 --> 00:32:26,640 Speaker 1: it turned out that so much of the polarization online 653 00:32:26,800 --> 00:32:29,920 Speaker 1: was just people feeling lonely and disconnected. So I could 654 00:32:29,920 --> 00:32:31,680 Speaker 1: go on for another hour about all the things that 655 00:32:31,680 --> 00:32:32,320 Speaker 1: we did. 656 00:32:32,240 --> 00:32:36,040 Speaker 3: That this all sounds so great and so okay. 657 00:32:36,080 --> 00:32:39,280 Speaker 2: So you've just laid out this beautiful world where there's 658 00:32:39,360 --> 00:32:42,320 Speaker 2: human connection and there's abundance and we're not as depressed 659 00:32:42,320 --> 00:32:43,360 Speaker 2: and we're not as anxious. 660 00:32:43,400 --> 00:32:45,400 Speaker 1: And it was so obvious because it wasn't even hard 661 00:32:45,440 --> 00:32:47,320 Speaker 1: to do. We just got honest about the nature. There 662 00:32:47,320 --> 00:32:50,160 Speaker 1: is a problem, a business model maximizing for attention. That 663 00:32:50,280 --> 00:32:51,880 Speaker 1: was the root of the problem. When we dealt with that, 664 00:32:52,680 --> 00:32:54,480 Speaker 1: the world culture started turning around. 665 00:32:54,560 --> 00:32:56,080 Speaker 2: And this is why you wake up every day and 666 00:32:56,120 --> 00:32:58,360 Speaker 2: you don't quit your jobs or you work for this world. 667 00:32:58,520 --> 00:33:00,200 Speaker 1: And this is why the Center for Humane Technology your 668 00:33:00,240 --> 00:33:03,440 Speaker 1: nonpuffits still cuts up every single day after thirteen years 669 00:33:03,480 --> 00:33:05,640 Speaker 1: and still works on these issues and still believes that 670 00:33:05,680 --> 00:33:07,720 Speaker 1: as bad as everything we just laid out the whole 671 00:33:07,760 --> 00:33:10,200 Speaker 1: point is to see that with clarity, so that we 672 00:33:10,200 --> 00:33:11,080 Speaker 1: can choose something else. 673 00:33:13,800 --> 00:33:16,440 Speaker 2: Mostly Human is a production of iHeart Podcasts and mostly 674 00:33:16,520 --> 00:33:19,840 Speaker 2: human Media. It's produced and edited by Laurie Siegel, Lauren Hanson, 675 00:33:19,920 --> 00:33:23,480 Speaker 2: and Nicole Bouchet. Sound design and mixing by Derek Clements, 676 00:33:23,840 --> 00:33:26,840 Speaker 2: additional production help from Abooz of Bar special thanks to 677 00:33:26,920 --> 00:33:30,600 Speaker 2: Mark Weinhaus. Find us on all socials at mostly human Media. 678 00:33:30,760 --> 00:33:33,080 Speaker 2: You can also watch mostly Human on our YouTube page. 679 00:33:33,160 --> 00:33:34,960 Speaker 2: If you want to get in touch, email us at 680 00:33:34,960 --> 00:33:38,040 Speaker 2: hello at mostlyghuman dot com. And if you like what 681 00:33:38,080 --> 00:33:40,360 Speaker 2: you're here, please rate and review the show and share 682 00:33:40,360 --> 00:33:41,040 Speaker 2: it with your friends. 683 00:33:41,120 --> 00:33:41,880 Speaker 3: See you next week.