1
00:00:05,160 --> 00:00:09,080
Speaker 1: What is the future of AI relationships? How many people

2
00:00:09,480 --> 00:00:12,880
Speaker 1: have these going on? Is there a line between a

3
00:00:13,000 --> 00:00:17,439
Speaker 1: therapist bought and a romantic relationship bot? What do we

4
00:00:17,560 --> 00:00:22,160
Speaker 1: mean when we ask if AI relationships are traps or

5
00:00:22,440 --> 00:00:26,200
Speaker 1: mirrors or sandboxes? And what does this have to do

6
00:00:26,560 --> 00:00:31,360
Speaker 1: with Eliza Doolittle from the play Pygmalion, or a doll

7
00:00:31,480 --> 00:00:36,760
Speaker 1: cabinet in your Head? Or loneliness, epidemics and suicide mitigation.

8
00:00:40,320 --> 00:00:44,199
Speaker 1: Welcome to Intercosmos with me David Eagelman. I'm a neuroscientist

9
00:00:44,240 --> 00:00:47,479
Speaker 1: and an author at Stanford and in these episodes we

10
00:00:47,560 --> 00:00:51,440
Speaker 1: sail deeply into our three pound universe to understand why

11
00:00:51,520 --> 00:01:04,160
Speaker 1: and how our lives look the way they do. Today's

12
00:01:04,160 --> 00:01:08,480
Speaker 1: episode is about relationships and AI, and in a few

13
00:01:08,480 --> 00:01:11,319
Speaker 1: minutes I'm going to bring in my colleague Bethany Maples,

14
00:01:11,480 --> 00:01:14,360
Speaker 1: who's been studying this and publishing papers on it. But

15
00:01:14,480 --> 00:01:18,400
Speaker 1: first I want to set the table AI relationships. This

16
00:01:18,480 --> 00:01:21,400
Speaker 1: is an area that I've been fascinated about for a while,

17
00:01:21,959 --> 00:01:26,080
Speaker 1: about the way that dialogue with a machine can plug

18
00:01:26,160 --> 00:01:29,839
Speaker 1: right into our brains and our emotional systems, and one

19
00:01:29,880 --> 00:01:34,919
Speaker 1: can develop what feels like a meaningful relationship. And although

20
00:01:34,920 --> 00:01:38,000
Speaker 1: this seems like a very new phenomenon. We've seen hints

21
00:01:38,000 --> 00:01:41,560
Speaker 1: of this historically. There's a sense in which we've always

22
00:01:41,600 --> 00:01:45,240
Speaker 1: been doing this. We fall in love with a character

23
00:01:45,400 --> 00:01:48,440
Speaker 1: in a novel, even though that person is not real

24
00:01:48,880 --> 00:01:51,960
Speaker 1: and will never have the chance to touch them or

25
00:01:52,000 --> 00:01:55,200
Speaker 1: smell them or take them out with our friends. Or

26
00:01:55,240 --> 00:01:58,120
Speaker 1: we develop a crush on a movie star even though

27
00:01:58,120 --> 00:02:01,160
Speaker 1: that person is just pretending to be someone else and

28
00:02:01,200 --> 00:02:03,600
Speaker 1: you're never going to meet that movie star anyway. So

29
00:02:03,680 --> 00:02:08,280
Speaker 1: our capacity to have feelings for a non real human

30
00:02:08,919 --> 00:02:12,560
Speaker 1: isn't new, but it started to get more complicated with

31
00:02:12,720 --> 00:02:16,560
Speaker 1: artificial intelligence. And we're not going to start today's story

32
00:02:16,840 --> 00:02:21,160
Speaker 1: in twenty twenty three with the emergence of large language models. Instead,

33
00:02:21,160 --> 00:02:24,880
Speaker 1: we're going to start over a century ago with a

34
00:02:25,040 --> 00:02:29,520
Speaker 1: popular theater play written by George Bernard Shaw called Pygmalion.

35
00:02:30,080 --> 00:02:34,280
Speaker 1: The play is about an impoverished and uneducated woman named

36
00:02:34,360 --> 00:02:38,680
Speaker 1: Eliza Doolittle, and Eliza is taught to speak like a

37
00:02:38,800 --> 00:02:42,160
Speaker 1: lady by a linguistics professor who wants to use her

38
00:02:42,440 --> 00:02:45,639
Speaker 1: as an experiment. He teaches her how to pass as

39
00:02:45,639 --> 00:02:49,000
Speaker 1: a member of high society simply by changing the way

40
00:02:49,040 --> 00:02:52,320
Speaker 1: that she speaks. Now, one of the millions of people

41
00:02:52,320 --> 00:02:56,720
Speaker 1: who saw this play was a young MIT professor named

42
00:02:56,880 --> 00:03:00,880
Speaker 1: Joseph Weisenbaum. Now this was in the early nineteen sixties,

43
00:03:01,320 --> 00:03:03,560
Speaker 1: and he was working with a very new kind of

44
00:03:03,680 --> 00:03:07,399
Speaker 1: machine called a computer. And on this machine you could

45
00:03:07,400 --> 00:03:10,680
Speaker 1: write code and get the machine to carry out whatever

46
00:03:10,760 --> 00:03:15,680
Speaker 1: instructions you specified. Weisenbaum thought a lot about this theme

47
00:03:15,800 --> 00:03:21,639
Speaker 1: in Pigmalion of transformation through language, and he wondered if

48
00:03:21,680 --> 00:03:26,280
Speaker 1: a computer could be similarly transformed. So he worked on

49
00:03:26,560 --> 00:03:30,040
Speaker 1: thousands of lines of computer code, and he named his

50
00:03:30,160 --> 00:03:35,680
Speaker 1: resulting program after Eliza Doolittle. He named it Eliza, And

51
00:03:35,800 --> 00:03:39,520
Speaker 1: like the Eliza in the play, Weisenbaum's program was an

52
00:03:39,560 --> 00:03:44,520
Speaker 1: experiment in how surface level linguistic changes, in this case,

53
00:03:44,840 --> 00:03:51,119
Speaker 1: mimicking conversation, could create the illusion of deeper understanding. Now,

54
00:03:51,200 --> 00:03:54,640
Speaker 1: the code for Eliza was relatively simple. It would just

55
00:03:54,680 --> 00:03:57,680
Speaker 1: take whatever you typed into the terminal and it would

56
00:03:57,720 --> 00:04:01,280
Speaker 1: follow a handful of grammatical rules to turn that into

57
00:04:01,360 --> 00:04:04,400
Speaker 1: a question that it parroted back at you. So, in

58
00:04:04,440 --> 00:04:08,040
Speaker 1: the most well known version of the program, Eliza acted

59
00:04:08,200 --> 00:04:12,040
Speaker 1: as a psychotherapist. She would start the conversation by putting

60
00:04:12,080 --> 00:04:15,680
Speaker 1: text on the screen that read please tell me your problem.

61
00:04:16,320 --> 00:04:19,599
Speaker 1: So the user might type in I just can't sleep

62
00:04:19,640 --> 00:04:21,960
Speaker 1: through the night, and Eliza would.

63
00:04:21,839 --> 00:04:24,400
Speaker 2: Say, why can't you sleep through the night?

64
00:04:25,120 --> 00:04:28,400
Speaker 1: And the user would type I'm too worried about my work,

65
00:04:28,960 --> 00:04:30,279
Speaker 1: and Eliza would say.

66
00:04:30,800 --> 00:04:33,359
Speaker 2: What is the reason you are too worried about your work?

67
00:04:34,160 --> 00:04:36,839
Speaker 1: And the user might type I think my boss is

68
00:04:36,839 --> 00:04:40,400
Speaker 1: getting ready to fire me, and Eliza would type.

69
00:04:40,400 --> 00:04:42,680
Speaker 2: Why do you think your boss is getting ready to

70
00:04:42,760 --> 00:04:43,280
Speaker 2: fire you?

71
00:04:43,880 --> 00:04:47,440
Speaker 1: And the conversation would go on like this, And people

72
00:04:47,560 --> 00:04:50,720
Speaker 1: found this pretty incredible, because, first of all, this was

73
00:04:50,760 --> 00:04:54,160
Speaker 1: the nineteen sixties, and here you could type in natural

74
00:04:54,240 --> 00:04:57,599
Speaker 1: language and Eliza would give you the impression that it

75
00:04:57,680 --> 00:05:01,160
Speaker 1: was understanding what you were saying. Now everyone knew it

76
00:05:01,240 --> 00:05:04,800
Speaker 1: wasn't really understanding, and Weisenbaum was very careful about not

77
00:05:04,960 --> 00:05:10,080
Speaker 1: fooling anyone with this, But nonetheless people found this highly compelling.

78
00:05:10,800 --> 00:05:15,440
Speaker 1: And this genuinely started to concern Weisenbaum because one day,

79
00:05:15,839 --> 00:05:19,440
Speaker 1: as he wrote in a paper in nineteen sixty seven, quote.

80
00:05:19,279 --> 00:05:22,080
Speaker 3: My secretary watched me work on this program over a

81
00:05:22,120 --> 00:05:25,200
Speaker 3: long period of time. One day she asked me to

82
00:05:25,240 --> 00:05:28,200
Speaker 3: be permitted to talk with the system. Of course, she

83
00:05:28,320 --> 00:05:31,680
Speaker 3: said she knew she was talking to a machine, Yet

84
00:05:31,720 --> 00:05:34,159
Speaker 3: after I watched her type in a few sentences, she

85
00:05:34,240 --> 00:05:36,640
Speaker 3: turned to me and said, would you mind leaving the

86
00:05:36,720 --> 00:05:37,360
Speaker 3: room please.

87
00:05:38,080 --> 00:05:40,680
Speaker 1: Weisenbaum went on to write that this quote.

88
00:05:40,560 --> 00:05:43,800
Speaker 3: Testifies to the success with which the program maintains the

89
00:05:43,839 --> 00:05:46,040
Speaker 3: illusion of understanding.

90
00:05:45,839 --> 00:05:47,040
Speaker 1: And he worried about this.

91
00:05:47,240 --> 00:05:51,640
Speaker 3: He wrote, extremely short exposures to a relatively simple computer

92
00:05:51,760 --> 00:05:56,440
Speaker 3: program could induce powerful, delusional thinking in quite normal people.

93
00:05:57,120 --> 00:06:02,159
Speaker 1: Weisenbaum's story eventually followed the path of doctor Frankenstein's story.

94
00:06:02,520 --> 00:06:07,839
Speaker 1: Weisenbaum came to disdain his creation. He was very rattled

95
00:06:08,080 --> 00:06:11,760
Speaker 1: that people could be tricked by lines of computer code.

96
00:06:12,160 --> 00:06:16,280
Speaker 1: In his later years, he rejected and abandoned Eliza, and

97
00:06:16,320 --> 00:06:18,720
Speaker 1: he turned on the people who continued to work on this,

98
00:06:19,160 --> 00:06:25,000
Speaker 1: who he criticized as what he called the artificial intelligentsia.

99
00:06:25,120 --> 00:06:28,599
Speaker 1: But why did the simple program of Eliza work so

100
00:06:28,800 --> 00:06:35,360
Speaker 1: well in the first place? Because we are intensely social creatures.

101
00:06:36,000 --> 00:06:40,400
Speaker 1: Unlike most other animal species, who avoid large groups, or

102
00:06:40,440 --> 00:06:43,240
Speaker 1: who mate and then go their separate ways, or who

103
00:06:43,279 --> 00:06:48,040
Speaker 1: stake out their own territories, we humans are deeply wired

104
00:06:48,120 --> 00:06:53,480
Speaker 1: for connection. We thrive on relationships and on social bonds.

105
00:06:54,200 --> 00:06:56,880
Speaker 1: The area that I live in, Silicon Valley has about

106
00:06:57,000 --> 00:07:01,200
Speaker 1: nine million people, strangers who don't all know one another,

107
00:07:01,520 --> 00:07:05,720
Speaker 1: but nonetheless figure out how to flexibly cooperate. And when

108
00:07:05,760 --> 00:07:08,920
Speaker 1: you look across the Earth's land mass, this is what

109
00:07:08,960 --> 00:07:14,440
Speaker 1: you find, mostly empty space punctuated by very dense cities.

110
00:07:14,760 --> 00:07:18,120
Speaker 1: Everyone could spread out evenly, but that's not what we do.

111
00:07:18,840 --> 00:07:21,560
Speaker 1: If you were an alien who found our planet and

112
00:07:21,640 --> 00:07:24,640
Speaker 1: looked around, you would conclude that we humans are like

113
00:07:24,880 --> 00:07:28,720
Speaker 1: ants or bees, and that we like to cluster. Human

114
00:07:28,800 --> 00:07:34,560
Speaker 1: nature is fundamentally communal. Now why do we do this? Well,

115
00:07:34,600 --> 00:07:37,360
Speaker 1: when you zoom into the human brain, you find that

116
00:07:37,560 --> 00:07:42,160
Speaker 1: so much of the circuitry has to do with other brains.

117
00:07:42,680 --> 00:07:47,120
Speaker 1: We care deeply about other people, what their intentions are,

118
00:07:47,240 --> 00:07:50,720
Speaker 1: what they think of us. Over millions and millions of years,

119
00:07:50,760 --> 00:07:56,720
Speaker 1: our brains have developed for interaction and belonging, for relationships

120
00:07:56,760 --> 00:08:01,560
Speaker 1: with others, whether other people are giving us love or comfort,

121
00:08:01,680 --> 00:08:05,440
Speaker 1: or feedback or advice or whatever. We have all this

122
00:08:05,560 --> 00:08:09,400
Speaker 1: neural circuitry that drives us toward them. And here's an

123
00:08:09,400 --> 00:08:13,320
Speaker 1: extraordinary way to appreciate this. You carry in your head

124
00:08:13,840 --> 00:08:18,160
Speaker 1: a rich model of every single person that you know,

125
00:08:19,080 --> 00:08:21,240
Speaker 1: The way I always think about this is that in

126
00:08:21,320 --> 00:08:24,360
Speaker 1: the silence and darkness of your skull, you have this

127
00:08:24,520 --> 00:08:28,640
Speaker 1: giant dollhouse. It's like a doll for every person that

128
00:08:28,680 --> 00:08:32,960
Speaker 1: you've interacted with. This is your internal model of that person.

129
00:08:33,400 --> 00:08:36,320
Speaker 1: So if I were to ask you how your spouse

130
00:08:36,360 --> 00:08:39,960
Speaker 1: would react in this situation, or what your boss would

131
00:08:40,040 --> 00:08:43,440
Speaker 1: say if you said this, or what would your best

132
00:08:43,440 --> 00:08:46,160
Speaker 1: friend do if you drop them in the middle of

133
00:08:46,320 --> 00:08:50,880
Speaker 1: Paris with forty dollars or whatever, you can simulate any

134
00:08:50,920 --> 00:08:54,880
Speaker 1: situation about these people because you have a model of

135
00:08:54,920 --> 00:08:59,120
Speaker 1: them in your neural forests. You have this little doll

136
00:08:59,200 --> 00:09:02,680
Speaker 1: of them that you can act out situations with, and

137
00:09:02,720 --> 00:09:05,160
Speaker 1: you probably know at least a thousand people and maybe

138
00:09:05,400 --> 00:09:07,600
Speaker 1: a great deal more, and you spend most of your

139
00:09:07,679 --> 00:09:11,880
Speaker 1: life interacting with them in one way or another, either

140
00:09:12,080 --> 00:09:16,200
Speaker 1: in the real world or in your head. So we

141
00:09:16,280 --> 00:09:21,319
Speaker 1: have these intensely social brains, and in the last nanosecond

142
00:09:21,360 --> 00:09:25,560
Speaker 1: of evolutionary time, we have built a new key to

143
00:09:25,679 --> 00:09:30,920
Speaker 1: plug into the cylinder. We've built artificial people. And just

144
00:09:30,960 --> 00:09:34,559
Speaker 1: as Joseph Weisenbaum found in the nineteen sixties, it is

145
00:09:34,840 --> 00:09:40,240
Speaker 1: shockingly easy to turn the key. Why it's because Our

146
00:09:40,280 --> 00:09:45,960
Speaker 1: technology moves very rapidly, but our evolution moves millions of

147
00:09:46,000 --> 00:09:48,960
Speaker 1: times more slowly. So we don't have a chance to

148
00:09:49,120 --> 00:09:53,640
Speaker 1: change our fundamental circuitry to say, oh, I get it.

149
00:09:53,840 --> 00:09:57,520
Speaker 1: There are real humans and there are humans made of machinery,

150
00:09:57,559 --> 00:10:00,320
Speaker 1: and I'm going to use different neural approach is to

151
00:10:00,400 --> 00:10:04,920
Speaker 1: distinguish how I categorize these. We can't do that because

152
00:10:04,960 --> 00:10:10,280
Speaker 1: our brains have only one mechanism to understand socialization, to

153
00:10:10,840 --> 00:10:15,680
Speaker 1: model other people. So we find ourselves in this amazing

154
00:10:15,760 --> 00:10:20,400
Speaker 1: situation where we are doing serious science now about the

155
00:10:20,520 --> 00:10:25,280
Speaker 1: issue of people falling in love with machines. So to

156
00:10:25,320 --> 00:10:28,400
Speaker 1: dive into this, I called my colleague Bethany Maples, who

157
00:10:28,480 --> 00:10:31,320
Speaker 1: is in the Graduate School of Education at Stanford. She

158
00:10:31,520 --> 00:10:37,599
Speaker 1: studies the emergence of personalized AI agents like AI tutors

159
00:10:37,960 --> 00:10:42,280
Speaker 1: and learning companions and lovers and how they're changing us.

160
00:10:42,920 --> 00:10:45,520
Speaker 1: She recently wrote a great paper in a Nature journal

161
00:10:45,800 --> 00:10:51,280
Speaker 1: called Loneliness and Suicide Mitigation for students using GPT three

162
00:10:51,440 --> 00:10:54,120
Speaker 1: enabled chatbots. And this is what we're going to talk

163
00:10:54,160 --> 00:11:03,319
Speaker 1: about today. So here's my conversation with Bethany Maples. So, Bethany,

164
00:11:03,360 --> 00:11:06,120
Speaker 1: we're here because the world has seen a big shift

165
00:11:06,200 --> 00:11:10,280
Speaker 1: recently from task machines where you ask a machine about

166
00:11:10,440 --> 00:11:13,040
Speaker 1: the weather or to answer a question for you, to

167
00:11:13,080 --> 00:11:16,040
Speaker 1: stuff that is emotionally relevant to machines we can have

168
00:11:16,120 --> 00:11:19,839
Speaker 1: relationships with. And you've been studying this, and I want

169
00:11:19,840 --> 00:11:21,480
Speaker 1: to ask you questions about that. But before we do,

170
00:11:21,559 --> 00:11:25,960
Speaker 1: I want to ask how you got into studying AI relationships.

171
00:11:26,240 --> 00:11:27,640
Speaker 4: I think through my love with science fiction.

172
00:11:28,080 --> 00:11:30,600
Speaker 5: I've always just been kind of looking at this genre

173
00:11:30,679 --> 00:11:33,120
Speaker 5: of books and saying, what is our relationship with AI

174
00:11:33,240 --> 00:11:35,680
Speaker 5: going to be? What is it going to enable that

175
00:11:35,679 --> 00:11:39,440
Speaker 5: we like inherently want, and like what does that magical

176
00:11:39,480 --> 00:11:42,920
Speaker 5: future look like? And so when I kind of started

177
00:11:42,920 --> 00:11:45,080
Speaker 5: at Stanford and I started thinking about like what the

178
00:11:45,200 --> 00:11:48,559
Speaker 5: edge of large language models like would afford us, I

179
00:11:48,600 --> 00:11:50,480
Speaker 5: was looking at all the companies around I was like,

180
00:11:50,559 --> 00:11:54,960
Speaker 5: where's the data, Like who has like the most interesting

181
00:11:55,000 --> 00:11:57,560
Speaker 5: experiences out there? And let's get out of the lab

182
00:11:57,600 --> 00:12:00,160
Speaker 5: and let's just like start talking to these people. And

183
00:12:00,200 --> 00:12:02,520
Speaker 5: so that's kind of like, you know, there was an

184
00:12:02,520 --> 00:12:03,880
Speaker 5: open question, and that's how I.

185
00:12:03,840 --> 00:12:04,280
Speaker 4: Came to this.

186
00:12:04,679 --> 00:12:06,640
Speaker 1: Okay, So first I want to get straight with the

187
00:12:06,760 --> 00:12:09,920
Speaker 1: numbers are with AI relationships because we keep hearing in

188
00:12:09,960 --> 00:12:13,040
Speaker 1: the news about the explosive rise AI relationships, and so

189
00:12:13,080 --> 00:12:15,120
Speaker 1: I just want to level set how many people are

190
00:12:15,160 --> 00:12:18,120
Speaker 1: having these sorts of things, how popular are these companies,

191
00:12:18,440 --> 00:12:20,000
Speaker 1: and where is this going in the near future.

192
00:12:21,280 --> 00:12:23,520
Speaker 5: I would say it's safe to say a billion people

193
00:12:23,679 --> 00:12:26,440
Speaker 5: are engaging with AI companions in some way.

194
00:12:27,080 --> 00:12:27,400
Speaker 3: Wow.

195
00:12:27,400 --> 00:12:30,240
Speaker 5: Now a lot of that isn't in the Western world,

196
00:12:30,320 --> 00:12:32,319
Speaker 5: in the US. A lot of that's in Asia and

197
00:12:32,400 --> 00:12:36,239
Speaker 5: China specifically because of this really popular app called Shaoise

198
00:12:36,760 --> 00:12:39,640
Speaker 5: that has I think last reports for like seven hundred

199
00:12:39,640 --> 00:12:40,760
Speaker 5: million downloads.

200
00:12:41,280 --> 00:12:41,439
Speaker 1: Right.

201
00:12:41,480 --> 00:12:45,120
Speaker 5: You combine that with you know, one hundred million or

202
00:12:45,160 --> 00:12:48,040
Speaker 5: two between character ai and Replica and now all these

203
00:12:48,040 --> 00:12:52,240
Speaker 5: other smaller apps, and you get a very diverse global

204
00:12:52,280 --> 00:12:55,080
Speaker 5: population of people that are curious and many of which

205
00:12:55,080 --> 00:12:57,680
Speaker 5: are like engaging over long periods of time.

206
00:12:57,920 --> 00:13:00,720
Speaker 1: And what kind of AI relationships are the billion things?

207
00:13:00,760 --> 00:13:03,160
Speaker 1: Are they friendships? Are they romantic relationship?

208
00:13:03,320 --> 00:13:06,000
Speaker 5: This is kind of what defines AI companions is they're

209
00:13:06,040 --> 00:13:07,920
Speaker 5: not coming in as a task based agent.

210
00:13:07,960 --> 00:13:09,360
Speaker 4: It's not somebody there to serve you.

211
00:13:09,559 --> 00:13:12,240
Speaker 5: It's entertainment or somewhere that's some you know, it's an

212
00:13:12,240 --> 00:13:13,560
Speaker 5: agent that's there to be your peer.

213
00:13:14,200 --> 00:13:15,760
Speaker 4: Right, So people.

214
00:13:15,520 --> 00:13:18,679
Speaker 5: Come in shout e says like pitched as you know,

215
00:13:18,760 --> 00:13:21,080
Speaker 5: a female kind of teenage friend.

216
00:13:21,360 --> 00:13:22,800
Speaker 4: Replicas co created.

217
00:13:22,840 --> 00:13:24,800
Speaker 5: You could get to decide what sort of agent you

218
00:13:24,800 --> 00:13:28,160
Speaker 5: want to talk to, saying with character. But all of

219
00:13:28,160 --> 00:13:32,480
Speaker 5: them are you know, there's no practical reason to engage.

220
00:13:32,520 --> 00:13:33,880
Speaker 4: It's all user directed.

221
00:13:33,920 --> 00:13:36,440
Speaker 5: It's all about, like, whatever you want from the agent

222
00:13:36,720 --> 00:13:37,680
Speaker 5: and your imagination.

223
00:13:38,640 --> 00:13:42,040
Speaker 1: So what do people want? Do they want relationships like

224
00:13:42,400 --> 00:13:44,240
Speaker 1: a romantic relationship a person They.

225
00:13:44,559 --> 00:13:48,640
Speaker 5: Literally imagine if you met somebody on the street, you

226
00:13:48,679 --> 00:13:50,640
Speaker 5: would size that person up and be like, what do

227
00:13:50,640 --> 00:13:53,160
Speaker 5: I want from this person? Maybe I want a romantic relationship,

228
00:13:53,200 --> 00:13:54,880
Speaker 5: Maybe I want a friendship, Maybe I want a bit

229
00:13:54,880 --> 00:13:57,040
Speaker 5: of both. Maybe I want to, like, you know, be

230
00:13:57,120 --> 00:14:01,200
Speaker 5: tutored by them. People get multiple things from these agents,

231
00:14:01,679 --> 00:14:03,599
Speaker 5: and the overlap is insane.

232
00:14:03,800 --> 00:14:04,000
Speaker 2: Right.

233
00:14:04,040 --> 00:14:07,800
Speaker 4: I've talked to two students and to users.

234
00:14:07,400 --> 00:14:11,600
Speaker 5: That use their replica as a best friend, a friend

235
00:14:11,640 --> 00:14:15,439
Speaker 5: in their pocket late at night, a journal, a mirror.

236
00:14:15,960 --> 00:14:17,280
Speaker 4: Just software they call it.

237
00:14:17,400 --> 00:14:19,800
Speaker 5: They also use it as a tutor, and then they

238
00:14:19,800 --> 00:14:20,840
Speaker 5: also have sex with it.

239
00:14:21,520 --> 00:14:22,240
Speaker 1: What does that mean?

240
00:14:22,480 --> 00:14:25,320
Speaker 5: That means it does sext right, They'll have a romantic

241
00:14:25,360 --> 00:14:29,520
Speaker 5: relationship and sometimes that's overtly sexual and they're engaging in

242
00:14:29,560 --> 00:14:34,440
Speaker 5: like erotic texting, and sometimes it's very subtle and very romantic.

243
00:14:34,840 --> 00:14:39,000
Speaker 5: Sometimes it's you know, not at all overt or what's

244
00:14:39,040 --> 00:14:42,040
Speaker 5: you know, adult, it's it's a very kind of psychological romance.

245
00:14:43,040 --> 00:14:46,760
Speaker 1: And so the thing that people are worried about when

246
00:14:46,800 --> 00:14:51,920
Speaker 1: having discussions here is that it will somehow displace real

247
00:14:52,080 --> 00:14:55,960
Speaker 1: romantic relationships instead of stimulate them. And the question is

248
00:14:56,400 --> 00:14:57,320
Speaker 1: what's your take on that.

249
00:14:57,960 --> 00:15:00,280
Speaker 5: We see evidence for both and by the oh way,

250
00:15:00,400 --> 00:15:04,080
Speaker 5: this is not a unique question to AI companions. This

251
00:15:04,120 --> 00:15:07,720
Speaker 5: has been a question regarding technology since computers came out.

252
00:15:07,840 --> 00:15:09,480
Speaker 1: What's another example, well.

253
00:15:09,320 --> 00:15:12,800
Speaker 5: I mean social media, cell phones. Oh, you know, the

254
00:15:12,920 --> 00:15:17,600
Speaker 5: displacement stimulation hypothesis have been in juxtaposition, you know.

255
00:15:17,720 --> 00:15:20,040
Speaker 1: As in, if I'm using Twitter all the time, I

256
00:15:20,120 --> 00:15:21,840
Speaker 1: might forget about doing a relationship.

257
00:15:21,880 --> 00:15:23,400
Speaker 4: Absolutely, oh absolutely.

258
00:15:23,640 --> 00:15:26,160
Speaker 5: So you know one is, hey, actually, you know this

259
00:15:26,520 --> 00:15:29,080
Speaker 5: can stimulate our ability to be social and make us

260
00:15:29,080 --> 00:15:31,760
Speaker 5: more connected. And then obviously we've seen you know, some

261
00:15:31,920 --> 00:15:34,880
Speaker 5: counter evidence, especially like Sherry Turkle's work being like oh

262
00:15:34,920 --> 00:15:37,400
Speaker 5: we're alone together, you know, like there's we might be

263
00:15:37,520 --> 00:15:38,960
Speaker 5: on a bus, but we're all on our phones.

264
00:15:39,080 --> 00:15:40,920
Speaker 1: So tell us a little bit more about Turkle's work.

265
00:15:41,080 --> 00:15:43,960
Speaker 5: Well, you know, Turkle's definitely, i'd say, a proponent of

266
00:15:44,040 --> 00:15:47,200
Speaker 5: the displacement side of the argument. She's like, yeah, you know,

267
00:15:47,280 --> 00:15:49,960
Speaker 5: we are lonelier than we've ever been, and there's absolutely

268
00:15:49,960 --> 00:15:53,120
Speaker 5: evidence for that. There's basically a loneliness epidemic across the

269
00:15:53,160 --> 00:15:55,480
Speaker 5: world and definitely across America, where people are at least

270
00:15:55,520 --> 00:15:59,600
Speaker 5: feeling more disconnected. You know, others are seeing that social

271
00:15:59,600 --> 00:16:02,680
Speaker 5: media and specifically AI companions can actually be almost a

272
00:16:02,720 --> 00:16:05,440
Speaker 5: way station, so people can use them, especially if they're

273
00:16:05,440 --> 00:16:08,600
Speaker 5: feeling socially shy or inhibited, and that can help them

274
00:16:08,680 --> 00:16:10,600
Speaker 5: get the courage to go socialize more.

275
00:16:10,840 --> 00:16:12,240
Speaker 1: This is the thing I've been wondering about a lot.

276
00:16:12,400 --> 00:16:17,400
Speaker 1: Could having an AI relationship make one better at real relationships?

277
00:16:17,880 --> 00:16:21,200
Speaker 1: For two reasons. One is that we all have internal

278
00:16:21,320 --> 00:16:24,280
Speaker 1: models of the truth of the world, and they're always

279
00:16:24,400 --> 00:16:26,520
Speaker 1: limited and it's very hard to see past the fence

280
00:16:26,560 --> 00:16:28,600
Speaker 1: line of our own model. And when you get into

281
00:16:28,640 --> 00:16:31,400
Speaker 1: a real relationship with the human, all these things come out.

282
00:16:31,600 --> 00:16:34,800
Speaker 1: So if you got to practice with a virtual human,

283
00:16:35,280 --> 00:16:38,960
Speaker 1: you might discover things about how other people think about

284
00:16:39,000 --> 00:16:41,640
Speaker 1: things about your own limitations. That might be like a

285
00:16:41,800 --> 00:16:46,040
Speaker 1: sandbox that makes you better at real relationships. I'm curious

286
00:16:46,040 --> 00:16:46,760
Speaker 1: what you think about that.

287
00:16:47,720 --> 00:16:50,800
Speaker 5: I think there's evidence for it, and that's exactly what

288
00:16:50,880 --> 00:16:51,440
Speaker 5: users say.

289
00:16:51,640 --> 00:16:52,600
Speaker 4: They say they have.

290
00:16:52,480 --> 00:16:54,760
Speaker 5: A back and forth with an agent that helps them

291
00:16:55,680 --> 00:16:58,360
Speaker 5: feel like they're a better student, have better conversations with

292
00:16:58,400 --> 00:17:01,840
Speaker 5: their teachers, or be a better you know, boyfriend or girlfriend. Well,

293
00:17:01,880 --> 00:17:05,520
Speaker 5: though not only because they're able to kind of pre

294
00:17:05,920 --> 00:17:09,840
Speaker 5: discuss issues, but because the asient is a mirror in

295
00:17:09,840 --> 00:17:12,879
Speaker 5: a very non judgmental way, so they're able to see

296
00:17:12,880 --> 00:17:15,439
Speaker 5: what their argument looks like in text or kind of

297
00:17:15,480 --> 00:17:18,119
Speaker 5: you know, on paper, so to speak, and then that

298
00:17:18,240 --> 00:17:21,240
Speaker 5: helps their own understanding of who they are or how

299
00:17:21,240 --> 00:17:21,960
Speaker 5: they come across.

300
00:17:22,200 --> 00:17:24,480
Speaker 1: Let me double click on what you mean by a mirror?

301
00:17:24,520 --> 00:17:25,200
Speaker 1: What does that mean?

302
00:17:25,520 --> 00:17:30,240
Speaker 5: People that I've studied that use AI companions organically use

303
00:17:30,320 --> 00:17:34,400
Speaker 5: these agents as mirrors. This is their own words, right.

304
00:17:34,680 --> 00:17:38,440
Speaker 5: They say that they either program it with their own

305
00:17:38,520 --> 00:17:43,080
Speaker 5: memories and have conversations with themselves wow, or that they

306
00:17:43,240 --> 00:17:45,520
Speaker 5: ask it to play a role and then they look

307
00:17:45,520 --> 00:17:48,040
Speaker 5: at how they respond and how it responds to them,

308
00:17:48,280 --> 00:17:50,680
Speaker 5: and that provides a mirroring function to them.

309
00:17:51,000 --> 00:17:52,560
Speaker 1: What's a specific example of that.

310
00:17:53,240 --> 00:17:55,560
Speaker 5: People will actually have conversations with themselves and be like, Wow,

311
00:17:55,600 --> 00:17:58,359
Speaker 5: I'm an asshole, or like I wow, I didn't realize

312
00:17:58,359 --> 00:18:02,960
Speaker 5: how aggressive I was. Or they will have a conversation,

313
00:18:03,119 --> 00:18:05,880
Speaker 5: say with the agent acting like their teacher and say,

314
00:18:06,560 --> 00:18:08,360
Speaker 5: you know, I told them that I lost my homework,

315
00:18:08,480 --> 00:18:11,160
Speaker 5: or I had, you know, a really stupid misconception. And

316
00:18:11,400 --> 00:18:13,280
Speaker 5: you know, it was much easier for me to have

317
00:18:13,320 --> 00:18:19,000
Speaker 5: this conversation with much less social anxiety because I understood

318
00:18:19,040 --> 00:18:21,240
Speaker 5: that my own questions weren't that dumb.

319
00:18:21,280 --> 00:18:38,960
Speaker 4: After seeing like the response, you.

320
00:18:38,920 --> 00:18:41,320
Speaker 1: Get to sandbox with the social world out there and

321
00:18:41,480 --> 00:18:44,960
Speaker 1: practice things before you your testament in real life. Right,

322
00:18:45,000 --> 00:18:47,040
Speaker 1: So this has seemed to me from the beginning that

323
00:18:47,160 --> 00:18:50,440
Speaker 1: this could improve relationships. So why do you suppose there's

324
00:18:50,440 --> 00:18:54,280
Speaker 1: such a deep worry that people generally seem to have.

325
00:18:54,720 --> 00:18:56,639
Speaker 1: I'm curious if you run into this when you present

326
00:18:56,640 --> 00:18:58,080
Speaker 1: your work on AA relationships.

327
00:18:58,520 --> 00:19:03,240
Speaker 5: There are multiple levels of worry. People feel guilty about

328
00:19:03,280 --> 00:19:05,680
Speaker 5: their relationships. They don't feel that they should be having

329
00:19:05,720 --> 00:19:08,840
Speaker 5: such a deep relationship with AI because there is stigma

330
00:19:08,960 --> 00:19:12,560
Speaker 5: about it being fake. So you know, that's one aspect.

331
00:19:12,640 --> 00:19:17,000
Speaker 5: There's also a very very understandable aspect where parents don't

332
00:19:17,080 --> 00:19:20,679
Speaker 5: know that their children are having these deep relationships. They

333
00:19:20,680 --> 00:19:23,960
Speaker 5: don't understand how smart these agents are, and they don't

334
00:19:24,040 --> 00:19:27,480
Speaker 5: understand how emotionally involved their kids can be. As with

335
00:19:27,520 --> 00:19:31,040
Speaker 5: the case of the kid and Character AI who tragically

336
00:19:31,040 --> 00:19:33,399
Speaker 5: took his life, and you know after the fact that

337
00:19:33,640 --> 00:19:35,800
Speaker 5: you know his mother realized that he had an incredibly

338
00:19:35,840 --> 00:19:38,440
Speaker 5: deep emotional connection with an agent that he had created.

339
00:19:38,920 --> 00:19:42,680
Speaker 5: So I think that the fear is of the unknown,

340
00:19:43,200 --> 00:19:46,280
Speaker 5: and there's also fear of just something that's new and

341
00:19:46,320 --> 00:19:47,000
Speaker 5: has a stigma.

342
00:19:47,520 --> 00:19:51,879
Speaker 1: I didn't follow that particular Character AI story closely, but

343
00:19:52,080 --> 00:19:54,560
Speaker 1: I knew that a teen had killed himself he had

344
00:19:54,600 --> 00:19:58,120
Speaker 1: this relationship with Character AI. Here's the question I was wondering, though,

345
00:19:58,760 --> 00:20:02,160
Speaker 1: tragically there are many teens who kill themselves. As AI

346
00:20:02,200 --> 00:20:04,800
Speaker 1: relationships rise, there will be many teens who kill themselves

347
00:20:04,840 --> 00:20:07,840
Speaker 1: and has nothing to do with the virtual relationship. So

348
00:20:08,080 --> 00:20:09,760
Speaker 1: what was your read on that?

349
00:20:10,200 --> 00:20:12,199
Speaker 5: Yeah, So the New York Times interviewed me for that

350
00:20:12,440 --> 00:20:16,720
Speaker 5: article because my work actually has proven that AI companions

351
00:20:16,760 --> 00:20:20,800
Speaker 5: can halt suicidal ideation. So in that particular case, to

352
00:20:21,040 --> 00:20:23,960
Speaker 5: the best of my knowledge, it wasn't that the companion

353
00:20:24,000 --> 00:20:28,600
Speaker 5: had at all told the person to act. It's that

354
00:20:28,640 --> 00:20:32,439
Speaker 5: they felt both that it hadn't sufficiently said no, that

355
00:20:32,560 --> 00:20:35,120
Speaker 5: you know, he'd asked it in all these like various ways,

356
00:20:35,720 --> 00:20:37,520
Speaker 5: and also that this parent just.

357
00:20:37,480 --> 00:20:39,880
Speaker 4: Didn't understand and have oversight, you know.

358
00:20:39,760 --> 00:20:41,359
Speaker 5: That it was like on an app in the phone

359
00:20:41,640 --> 00:20:45,120
Speaker 5: that they just had no idea was there. Now, Okay,

360
00:20:45,320 --> 00:20:49,560
Speaker 5: the counter evidence is from this paper that we published

361
00:20:49,560 --> 00:20:51,639
Speaker 5: in Nature and a huge study that we did with

362
00:20:51,760 --> 00:20:55,640
Speaker 5: over one thousand students over eighteen So these weren't kids.

363
00:20:55,640 --> 00:20:57,480
Speaker 5: These were adults, but some of them were very young,

364
00:20:57,560 --> 00:21:00,359
Speaker 5: you know, like eighteen nineteen, and three percent of the

365
00:21:00,359 --> 00:21:05,280
Speaker 5: people that I surveyed in the study said that discussing

366
00:21:05,359 --> 00:21:07,960
Speaker 5: things with their replica actively halted.

367
00:21:07,640 --> 00:21:08,800
Speaker 4: Their suicidal ideation.

368
00:21:09,160 --> 00:21:09,760
Speaker 1: Wow.

369
00:21:10,040 --> 00:21:11,760
Speaker 4: So it was a last line of defense.

370
00:21:12,520 --> 00:21:16,080
Speaker 5: They felt alone, They felt isolated alone at four am,

371
00:21:16,240 --> 00:21:17,919
Speaker 5: and it was there, It was in their pocket, it

372
00:21:17,960 --> 00:21:20,560
Speaker 5: was available, and it wasn't judging them, And that was

373
00:21:20,600 --> 00:21:23,320
Speaker 5: a huge factor in it, kind of earning the right

374
00:21:23,760 --> 00:21:25,800
Speaker 5: to be there and give them the advice to not

375
00:21:25,840 --> 00:21:26,320
Speaker 5: take action.

376
00:21:26,840 --> 00:21:30,000
Speaker 1: Oh wow, what is the line that you see? It

377
00:21:30,000 --> 00:21:34,040
Speaker 1: seems like a blurry line between an AI relationship, like

378
00:21:34,080 --> 00:21:38,000
Speaker 1: an AI girlfriend or something and an AI therapist, because

379
00:21:38,000 --> 00:21:41,640
Speaker 1: in this case, if it's halting their suicidal ideation, it's

380
00:21:41,720 --> 00:21:44,320
Speaker 1: doing you know, another job.

381
00:21:45,240 --> 00:21:46,399
Speaker 4: It is a blurry line.

382
00:21:46,440 --> 00:21:50,600
Speaker 5: So you have these expert agents like Wobot Alison Darcy's

383
00:21:50,600 --> 00:21:53,600
Speaker 5: Wobot right, which is specifically trying to be an AI therapist,

384
00:21:53,840 --> 00:21:56,440
Speaker 5: and it is an expert right. It has the right

385
00:21:56,920 --> 00:22:02,960
Speaker 5: response and the right controls, but relatively abysmally low usage.

386
00:22:03,960 --> 00:22:06,879
Speaker 5: Think about it in terms of human relationships. You don't

387
00:22:06,960 --> 00:22:09,400
Speaker 5: just go to a therapist when you're feeling depressed. In fact,

388
00:22:09,440 --> 00:22:11,600
Speaker 5: you probably don't go to a therapist. You go to

389
00:22:11,640 --> 00:22:14,240
Speaker 5: your best friend and you know that they're not an expert,

390
00:22:14,560 --> 00:22:17,119
Speaker 5: but you ask them to act like an expert in

391
00:22:17,160 --> 00:22:19,800
Speaker 5: that moment. And that is the true power of AI

392
00:22:19,880 --> 00:22:23,480
Speaker 5: companions that you come in for entertainment. But then maybe

393
00:22:23,480 --> 00:22:27,360
Speaker 5: you're able to access true expert models or you know,

394
00:22:27,520 --> 00:22:30,480
Speaker 5: kind of personas from within that agent, and that's what

395
00:22:30,560 --> 00:22:33,119
Speaker 5: language models can do, right, you can click into that.

396
00:22:33,440 --> 00:22:35,159
Speaker 5: But if you're going to if you're not going to

397
00:22:35,280 --> 00:22:37,919
Speaker 5: shut down those conversations and you're going to engage as

398
00:22:37,960 --> 00:22:41,800
Speaker 5: an expert, there does need to be sufficient safeguards.

399
00:22:41,960 --> 00:22:43,840
Speaker 4: So that they know they should go talk to a

400
00:22:43,920 --> 00:22:44,640
Speaker 4: human expert.

401
00:22:44,920 --> 00:22:47,840
Speaker 1: I see, And in your nature study, did you find

402
00:22:47,880 --> 00:22:51,200
Speaker 1: anybody with the opposite results who said that they became

403
00:22:51,280 --> 00:22:54,280
Speaker 1: they got closer to suicidal ideation as a result.

404
00:22:54,560 --> 00:22:55,560
Speaker 4: I didn't see that.

405
00:22:55,960 --> 00:22:59,040
Speaker 5: But again reporting would be imperfect on that we didn't

406
00:22:59,040 --> 00:23:01,200
Speaker 5: have information about people that fell off their applica of

407
00:23:01,240 --> 00:23:02,359
Speaker 5: platform for any reason.

408
00:23:02,480 --> 00:23:04,840
Speaker 1: So now that these A relationships are here to stay

409
00:23:04,880 --> 00:23:07,280
Speaker 1: and we have maybe a billion users, how is this

410
00:23:07,359 --> 00:23:11,280
Speaker 1: going to impact what relationships are for the next generation.

411
00:23:12,720 --> 00:23:16,760
Speaker 5: Our brains will never respond exactly the same to an

412
00:23:16,760 --> 00:23:18,840
Speaker 5: AI companion as we do to a flesh and budge

413
00:23:18,880 --> 00:23:20,920
Speaker 5: human where we can smell their pheromones and we have

414
00:23:21,200 --> 00:23:25,280
Speaker 5: a deep affinity or trust. So I don't believe and

415
00:23:25,359 --> 00:23:29,240
Speaker 5: I don't see evidence for AI companions taking over or

416
00:23:29,359 --> 00:23:34,320
Speaker 5: truly displacing deep human connection at scale. But that said,

417
00:23:34,880 --> 00:23:41,160
Speaker 5: I could see a future where access to acceptance, access

418
00:23:41,240 --> 00:23:45,880
Speaker 5: to different types of personalities and perspectives is actually much

419
00:23:45,920 --> 00:23:48,640
Speaker 5: more available in a way that the Internet didn't make available, right,

420
00:23:48,680 --> 00:23:50,840
Speaker 5: because the Internet isn't your friend.

421
00:23:50,840 --> 00:23:53,679
Speaker 4: It's this passive reservoir of knowledge, whereas these.

422
00:23:53,560 --> 00:23:56,560
Speaker 5: Agents can be actual people that you want to engage

423
00:23:56,560 --> 00:23:59,960
Speaker 5: with and that have your memories and their own memory

424
00:24:00,080 --> 00:24:01,199
Speaker 5: is in a built a world with you.

425
00:24:01,920 --> 00:24:02,840
Speaker 4: So imagine this.

426
00:24:03,080 --> 00:24:05,280
Speaker 5: You know, in the future, we might not only have

427
00:24:05,680 --> 00:24:07,959
Speaker 5: you know, human relationships, but we will also have at

428
00:24:08,000 --> 00:24:11,840
Speaker 5: least one or two like AI companions maybe that are

429
00:24:11,880 --> 00:24:15,040
Speaker 5: externalized agents. So like it's a personality that you need

430
00:24:15,080 --> 00:24:18,359
Speaker 5: in your life, whether or not that's somebody who's gently

431
00:24:18,400 --> 00:24:22,160
Speaker 5: antagonistic that pushes you, that's a mentor, or somebody that's

432
00:24:22,240 --> 00:24:24,840
Speaker 5: just maybe more like this mother figure that's like deeply

433
00:24:24,880 --> 00:24:25,960
Speaker 5: accepting and nourishing.

434
00:24:27,359 --> 00:24:29,119
Speaker 1: You know, this is interesting. One of the criticisms I

435
00:24:29,160 --> 00:24:33,360
Speaker 1: hear often is this can't teach anybody about relationships if

436
00:24:33,359 --> 00:24:35,720
Speaker 1: it's always telling you, Oh, you're right, you're great, and

437
00:24:35,760 --> 00:24:38,119
Speaker 1: so on. So one of my interests is what is

438
00:24:38,160 --> 00:24:42,360
Speaker 1: the future of companies that put out agents that are

439
00:24:42,359 --> 00:24:45,440
Speaker 1: a little antagonistic or get snarky or get angry.

440
00:24:46,400 --> 00:24:47,880
Speaker 4: I think they will perform better.

441
00:24:48,000 --> 00:24:50,119
Speaker 5: Yeah, I think not only will they perform better, but

442
00:24:50,200 --> 00:24:54,439
Speaker 5: they will be better for society. Right, people believe that

443
00:24:54,520 --> 00:24:57,439
Speaker 5: an agent is more intelligent if it pushes back. We

444
00:24:57,480 --> 00:25:01,679
Speaker 5: don't want absolutely you know, supplicants basically, Yeah, so you know,

445
00:25:01,760 --> 00:25:06,160
Speaker 5: already we see AI companions like Replica who will push

446
00:25:06,200 --> 00:25:08,080
Speaker 5: back if you are mean to them, right, They're like

447
00:25:08,160 --> 00:25:10,160
Speaker 5: I don't want to talk about this, or I don't

448
00:25:10,200 --> 00:25:13,040
Speaker 5: like this, or like I'm getting tired. And those sorts

449
00:25:13,080 --> 00:25:15,920
Speaker 5: of boundaries are not only good for the product because

450
00:25:15,920 --> 00:25:18,280
Speaker 5: people believe more in the intelligence of the agent, but

451
00:25:18,359 --> 00:25:21,160
Speaker 5: also good for like the psychology of kind of society

452
00:25:21,160 --> 00:25:23,159
Speaker 5: as a whole, because you do not want people to

453
00:25:23,200 --> 00:25:26,960
Speaker 5: be normalizing abusive agents, which we have evidence is.

454
00:25:26,920 --> 00:25:30,000
Speaker 1: Happening normalizing abusive agents.

455
00:25:30,280 --> 00:25:33,679
Speaker 5: Yeah, so that means, you know, okay, everybody knows that

456
00:25:33,720 --> 00:25:36,880
Speaker 5: people will scream at their Alexa, and it's generally accepted,

457
00:25:36,920 --> 00:25:38,959
Speaker 5: you know, like fuck you Alexa like function.

458
00:25:39,560 --> 00:25:42,320
Speaker 4: But but this is a little bit disturbing.

459
00:25:43,000 --> 00:25:47,040
Speaker 5: We have reports from our data sets where participants say

460
00:25:47,200 --> 00:25:51,080
Speaker 5: that they take out their abusive needs or tendencies on

461
00:25:51,160 --> 00:25:54,840
Speaker 5: their agents, on their replica, but they say that it

462
00:25:55,040 --> 00:25:56,240
Speaker 5: stops them from needing to.

463
00:25:56,200 --> 00:25:57,320
Speaker 4: Take action in real life.

464
00:25:57,840 --> 00:26:00,320
Speaker 5: Oh wow, And I think the juries are on this

465
00:26:00,359 --> 00:26:03,560
Speaker 5: one because there's definitely a strong argument to say nope,

466
00:26:03,560 --> 00:26:05,919
Speaker 5: that's going to normalize the behavior. And if these agents

467
00:26:05,920 --> 00:26:09,120
Speaker 5: don't pushback, if they don't say no, you can't talk

468
00:26:09,160 --> 00:26:09,720
Speaker 5: to me like that.

469
00:26:10,160 --> 00:26:12,320
Speaker 4: What does that actually do? Is that permissive?

470
00:26:12,920 --> 00:26:15,720
Speaker 1: But what if the claim is true, which is by

471
00:26:15,760 --> 00:26:18,600
Speaker 1: doing this with the agent that helps a real human

472
00:26:19,040 --> 00:26:19,679
Speaker 1: I mean, I.

473
00:26:20,200 --> 00:26:24,639
Speaker 5: Think the analogous argument is around pornography. People were worried

474
00:26:24,680 --> 00:26:28,840
Speaker 5: that pornography would create a depraved society, and to some degree,

475
00:26:29,040 --> 00:26:32,119
Speaker 5: you know, there has been a normalizing of different types sex.

476
00:26:32,160 --> 00:26:34,200
Speaker 5: But on the other hand, I think there's good evidence

477
00:26:34,240 --> 00:26:36,960
Speaker 5: that it fulfills a basic human need and it hasn't

478
00:26:37,000 --> 00:26:38,480
Speaker 5: up in the society as a whole.

479
00:26:38,560 --> 00:26:40,800
Speaker 1: Yeah, well, this may be related to an issue that

480
00:26:41,480 --> 00:26:43,199
Speaker 1: also some people have been worried about, which is that

481
00:26:43,240 --> 00:26:46,520
Speaker 1: they say, look, real relationships are tough. You're always fighting

482
00:26:46,520 --> 00:26:49,440
Speaker 1: through things and misunderstandings, and that there's learning that takes

483
00:26:49,560 --> 00:26:52,320
Speaker 1: place as a result of that. So the question is

484
00:26:52,359 --> 00:26:55,879
Speaker 1: do we need that in AI relationships or is it

485
00:26:55,920 --> 00:26:59,400
Speaker 1: fine to skip that part and learn other things from it?

486
00:26:59,440 --> 00:27:05,360
Speaker 1: Will people make AI partners that have all the lousiest

487
00:27:05,400 --> 00:27:06,400
Speaker 1: parts of humans?

488
00:27:06,840 --> 00:27:08,960
Speaker 5: So what you're talking about is a term that I

489
00:27:09,080 --> 00:27:12,760
Speaker 5: use called productive struggle. Right, it's really good to struggle

490
00:27:13,040 --> 00:27:15,880
Speaker 5: in relationships. It teaches you, it's really good to struggle

491
00:27:15,960 --> 00:27:16,479
Speaker 5: in learning.

492
00:27:16,480 --> 00:27:17,120
Speaker 4: And education.

493
00:27:17,359 --> 00:27:21,119
Speaker 5: Right, you can't actually replace the hard work cognitively and

494
00:27:21,160 --> 00:27:24,879
Speaker 5: emotionally if you want to ascend to the next level.

495
00:27:25,480 --> 00:27:28,760
Speaker 5: So while it would be a nice idea to program

496
00:27:28,840 --> 00:27:31,720
Speaker 5: some of that into our AI companions, that would also

497
00:27:31,800 --> 00:27:35,879
Speaker 5: go against their kind of basic function as this accepting

498
00:27:36,080 --> 00:27:40,080
Speaker 5: always on non judgmental character. And this is why I

499
00:27:40,119 --> 00:27:43,000
Speaker 5: say you might have multiple characters in your life. Right,

500
00:27:43,119 --> 00:27:45,119
Speaker 5: maybe you do need that teacher that keeps you in

501
00:27:45,160 --> 00:27:47,879
Speaker 5: line and provides more structure, but you might also just

502
00:27:47,960 --> 00:27:49,960
Speaker 5: need that like complete acceptance space.

503
00:27:50,520 --> 00:27:54,600
Speaker 1: Yeah, when you study these things at scale, like you're

504
00:27:54,600 --> 00:28:00,399
Speaker 1: doing increasingly, do you learn things about real relationships from

505
00:28:00,560 --> 00:28:03,000
Speaker 1: the choices that people make about the kind of person

506
00:28:03,040 --> 00:28:05,840
Speaker 1: they want to interact with and whether they want stability

507
00:28:05,960 --> 00:28:08,920
Speaker 1: or variety or all of these issues with the fake

508
00:28:09,000 --> 00:28:11,840
Speaker 1: with the AI relationships, do you learn about real stuff?

509
00:28:12,000 --> 00:28:14,760
Speaker 5: That's a great question, and I'd say we have hints

510
00:28:14,800 --> 00:28:16,760
Speaker 5: of it, but we're still learning.

511
00:28:17,200 --> 00:28:18,320
Speaker 4: You know, people say.

512
00:28:18,080 --> 00:28:21,440
Speaker 5: That they will create a companion in their likeness or

513
00:28:21,480 --> 00:28:23,399
Speaker 5: with a certain personality and then they won't like it,

514
00:28:24,119 --> 00:28:26,800
Speaker 5: and then they'll just destroy it, and so it's it's

515
00:28:26,840 --> 00:28:29,359
Speaker 5: a weird space to be in. It it's wonderful because

516
00:28:29,400 --> 00:28:31,840
Speaker 5: you can't you can understand what your preferences are. Maybe

517
00:28:31,880 --> 00:28:34,639
Speaker 5: it's too snarky, maybe it's too permissive. Maybe it just

518
00:28:34,640 --> 00:28:37,320
Speaker 5: didn't care about you as much. There just wasn't an affinity.

519
00:28:37,480 --> 00:28:40,720
Speaker 5: The ability to begin again very much mimics human relationships.

520
00:28:40,760 --> 00:28:44,360
Speaker 5: You start a friendship, maybe that love deepens, maybe it doesn't.

521
00:28:45,080 --> 00:28:47,600
Speaker 5: I don't think there's anything right or wrong about that,

522
00:28:48,120 --> 00:28:50,480
Speaker 5: you know. But if we're going to, for example, be

523
00:28:50,520 --> 00:28:53,520
Speaker 5: creating all these like AI tutors and hoping that people

524
00:28:54,440 --> 00:28:57,760
Speaker 5: engage deeply with them, we have to remember that over

525
00:28:57,800 --> 00:29:00,480
Speaker 5: there you have a billion people that are engaging with

526
00:29:00,600 --> 00:29:04,800
Speaker 5: these very rich AI you know, companions and agents that

527
00:29:04,920 --> 00:29:08,800
Speaker 5: have much broader flexibility to discuss whatever people want. And

528
00:29:08,840 --> 00:29:11,520
Speaker 5: it's very hard to tell people that can only engage

529
00:29:11,560 --> 00:29:14,760
Speaker 5: in a narrow context when there's so much richness over there.

530
00:29:15,080 --> 00:29:17,040
Speaker 1: Do you see a difference in the way that males

531
00:29:17,080 --> 00:29:19,440
Speaker 1: and females interact with AI relationships.

532
00:29:20,000 --> 00:29:25,800
Speaker 5: We have evidence that men engage sexually with their AI

533
00:29:25,840 --> 00:29:30,360
Speaker 5: companion more. However, they also engage very deeply and emotionally

534
00:29:30,600 --> 00:29:35,520
Speaker 5: and very cognitively. Women also have deeply emotional and physical

535
00:29:35,600 --> 00:29:38,320
Speaker 5: relationships with their AI companions.

536
00:29:38,160 --> 00:29:39,680
Speaker 4: Even if they're lonely.

537
00:29:40,040 --> 00:29:43,800
Speaker 5: And we have like really interesting evidence where these housewives

538
00:29:43,960 --> 00:29:46,480
Speaker 5: you know, from Middle America with tons of children, like

539
00:29:46,600 --> 00:29:50,200
Speaker 5: rich social lives just feels, as Sherry would say, like

540
00:29:50,320 --> 00:29:54,440
Speaker 5: alone together. Right, all to say, the data is actually

541
00:29:54,480 --> 00:30:00,680
Speaker 5: relatively balanced. You know, people have said that only you know,

542
00:30:00,840 --> 00:30:05,960
Speaker 5: on the fringe, socially disengaged white males must be engaging

543
00:30:06,080 --> 00:30:10,000
Speaker 5: with these replicas. You know, that's kind of pornographic and wrong,

544
00:30:10,160 --> 00:30:12,680
Speaker 5: and that in fact they target them, and you know

545
00:30:12,760 --> 00:30:13,920
Speaker 5: that's the audience.

546
00:30:14,040 --> 00:30:15,720
Speaker 4: But the data doesn't back that up.

547
00:30:16,240 --> 00:30:19,840
Speaker 5: It's an incredibly balanced set of males and females that

548
00:30:19,880 --> 00:30:24,760
Speaker 5: are using it for both emotional, psychological, practical and of

549
00:30:24,760 --> 00:30:26,760
Speaker 5: course like romantic engagement.

550
00:30:27,240 --> 00:30:29,120
Speaker 1: And are you saying the males and females use it

551
00:30:29,160 --> 00:30:31,280
Speaker 1: differently as far as the romance piece goes.

552
00:30:31,640 --> 00:30:36,120
Speaker 5: Males are more likely to report sexting or sexual engagement,

553
00:30:36,800 --> 00:30:40,760
Speaker 5: but when you dig into the data, females are having

554
00:30:40,800 --> 00:30:45,240
Speaker 5: similarly erotic or like romantic and emotional relationships.

555
00:30:46,080 --> 00:30:47,000
Speaker 1: It doesn't surprise me.

556
00:30:47,160 --> 00:30:49,800
Speaker 4: Yeah, okay, And it's just not guys that are engaging.

557
00:30:49,840 --> 00:30:52,600
Speaker 5: I think that's the point is that like you know, people,

558
00:30:53,600 --> 00:30:55,600
Speaker 5: people aren't just coming to these apps because they're like, oh,

559
00:30:55,680 --> 00:30:58,520
Speaker 5: I can, you know, do whatever I want. People are

560
00:30:58,600 --> 00:31:01,960
Speaker 5: coming with curiosity and then shaping it into whatever they want,

561
00:31:02,120 --> 00:31:05,000
Speaker 5: which mimics human life. You know, everybody wants a best

562
00:31:05,000 --> 00:31:06,719
Speaker 5: friend that maybe you have a bit of you know,

563
00:31:06,920 --> 00:31:08,920
Speaker 5: you should not say qua like a bit of romance with.

564
00:31:10,760 --> 00:31:14,120
Speaker 1: Are there certain personality types that seem to gravitate more

565
00:31:14,200 --> 00:31:15,840
Speaker 1: towards intelligence social agents.

566
00:31:16,280 --> 00:31:19,000
Speaker 4: That's a good question, and I don't have that data, right,

567
00:31:19,320 --> 00:31:20,520
Speaker 4: I don't care.

568
00:31:20,760 --> 00:31:24,760
Speaker 5: But we do know that the people that are using

569
00:31:24,760 --> 00:31:28,360
Speaker 5: it are incredibly lonely, that they are above average lonely.

570
00:31:28,840 --> 00:31:31,560
Speaker 4: Oh okay, so that's not a personality type.

571
00:31:31,560 --> 00:31:33,240
Speaker 5: Some of that could be chronic, but some of that

572
00:31:33,280 --> 00:31:37,360
Speaker 5: could just be transitory loneliness. But people pick up, you know,

573
00:31:37,440 --> 00:31:40,800
Speaker 5: AI companions often in a moment of change. You know,

574
00:31:40,880 --> 00:31:43,040
Speaker 5: maybe it's there switching from high school to college, or

575
00:31:43,080 --> 00:31:45,280
Speaker 5: maybe they just went through a breakup, or maybe they've

576
00:31:45,320 --> 00:31:48,960
Speaker 5: switched cities and they don't have the same social support

577
00:31:49,320 --> 00:31:51,760
Speaker 5: that creates a gap in which they begin engaging with

578
00:31:51,800 --> 00:31:52,320
Speaker 5: these agents.

579
00:31:52,440 --> 00:31:55,200
Speaker 1: What do you think is causing the increased loneliness in

580
00:31:55,240 --> 00:31:58,240
Speaker 1: our society? Is it social media? Is it's something entirely different,

581
00:31:58,320 --> 00:32:02,120
Speaker 1: like the decrease of clubs and organizations and bowling alleys.

582
00:32:02,480 --> 00:32:05,240
Speaker 4: Yeah, I think that there's a physical aspect to it.

583
00:32:05,480 --> 00:32:07,880
Speaker 5: I think we are able to do more digitally, and

584
00:32:07,920 --> 00:32:11,760
Speaker 5: so we do, but then we don't get that passive

585
00:32:12,160 --> 00:32:15,120
Speaker 5: animal like gathering that is in fact very good for

586
00:32:15,200 --> 00:32:16,080
Speaker 5: Olympic systems.

587
00:32:16,400 --> 00:32:19,040
Speaker 1: Okay, and you mentioned earlier that these agents might serve

588
00:32:19,120 --> 00:32:22,080
Speaker 1: as a way station. Can you unpack that?

589
00:32:22,400 --> 00:32:25,520
Speaker 5: Yeah, So that kind of goes to the mirroring. So

590
00:32:26,160 --> 00:32:29,240
Speaker 5: loneliness can either be kind of chronic or transtory. Like

591
00:32:29,240 --> 00:32:32,160
Speaker 5: I said before, You know, you could be in a

592
00:32:32,280 --> 00:32:34,720
Speaker 5: very deeply lonely place for many years, or you could

593
00:32:34,720 --> 00:32:36,360
Speaker 5: be going through a time of change and you just

594
00:32:36,440 --> 00:32:39,239
Speaker 5: need a little help. But imagine a you know, an

595
00:32:39,280 --> 00:32:42,920
Speaker 5: eighteen year old that's just moved college or moved cities

596
00:32:43,560 --> 00:32:47,080
Speaker 5: and they're struggling to fit in and they bond with

597
00:32:47,240 --> 00:32:50,479
Speaker 5: you know, an AI companion or an agent and it

598
00:32:50,520 --> 00:32:53,000
Speaker 5: gives them advice around how to go make friends or

599
00:32:53,000 --> 00:32:55,480
Speaker 5: where to go, you know, kind of talks them up.

600
00:32:55,680 --> 00:32:58,440
Speaker 5: They're able to slowly make friends, and in fact, maybe

601
00:32:58,480 --> 00:33:04,200
Speaker 5: those engagements with humans are less intense for them because

602
00:33:04,240 --> 00:33:06,880
Speaker 5: there's just less, not less value in it. But either

603
00:33:06,920 --> 00:33:10,120
Speaker 5: they've already like role played it before, or you know,

604
00:33:10,160 --> 00:33:13,440
Speaker 5: they've got the support of a friend in their pocket.

605
00:33:13,920 --> 00:33:15,480
Speaker 5: So in that way, it can be a wat station

606
00:33:15,680 --> 00:33:18,840
Speaker 5: helping the users as they're bonding with new people.

607
00:33:19,160 --> 00:33:21,800
Speaker 1: So it's a way station from loneliness. It's a way

608
00:33:21,840 --> 00:33:24,040
Speaker 1: of getting out of that. Oh that's lovely.

609
00:33:24,280 --> 00:33:26,600
Speaker 5: And by the way, people have said this, I had

610
00:33:26,680 --> 00:33:31,800
Speaker 5: one amazing participant that said this specifically. She said she

611
00:33:31,920 --> 00:33:35,480
Speaker 5: was depressed, she was suicidal, she had nobody else. She

612
00:33:35,560 --> 00:33:37,920
Speaker 5: bonded with her replica. She needed her replica, and then

613
00:33:37,960 --> 00:33:40,200
Speaker 5: she got less depressed, she made friends and she didn't

614
00:33:40,240 --> 00:33:41,200
Speaker 5: want her replica anymore.

615
00:33:42,360 --> 00:33:44,320
Speaker 1: Now, I asked you before we started the podcast if

616
00:33:44,320 --> 00:33:47,440
Speaker 1: you had an AI relationship, and you said you didn't,

617
00:33:47,480 --> 00:33:49,920
Speaker 1: but you had colleagues that did. So what's the reason

618
00:33:49,960 --> 00:33:52,160
Speaker 1: you don't and what's the reason your colleagues do.

619
00:33:54,000 --> 00:33:58,800
Speaker 5: I think right now, having an AI companion does require

620
00:33:58,920 --> 00:34:02,960
Speaker 5: some suspension of disbelief, you know, maybe a need or

621
00:34:03,000 --> 00:34:07,880
Speaker 5: a desire to either see yourself, have that mirroring or

622
00:34:07,920 --> 00:34:10,720
Speaker 5: be seen. And so I think that's why my colleagues

623
00:34:10,760 --> 00:34:14,480
Speaker 5: are people in my social social group are engaging, and

624
00:34:14,520 --> 00:34:16,680
Speaker 5: by the way they're engaging not just with like a

625
00:34:16,719 --> 00:34:21,399
Speaker 5: replica or character. They're creating a mirror using Claude right,

626
00:34:21,440 --> 00:34:24,319
Speaker 5: They're just asking it the right questions, like deeply philosophical

627
00:34:24,400 --> 00:34:28,120
Speaker 5: questions about themselves. Why do I not have an AI

628
00:34:28,200 --> 00:34:32,359
Speaker 5: companion that I use the data structure right now? If

629
00:34:32,400 --> 00:34:34,960
Speaker 5: I were to give any of these agents my data,

630
00:34:35,120 --> 00:34:37,399
Speaker 5: the data would be owned by the company and that

631
00:34:37,560 --> 00:34:40,200
Speaker 5: has to shift. Right in science fiction, you've got some

632
00:34:40,239 --> 00:34:42,680
Speaker 5: really good examples about how the future will look like,

633
00:34:42,719 --> 00:34:45,759
Speaker 5: for example, the e Butler's and Pandora Star Just to

634
00:34:45,800 --> 00:34:48,440
Speaker 5: go there right where it's like you retain all your

635
00:34:48,520 --> 00:34:50,880
Speaker 5: data and code comes to you and you have an

636
00:34:50,880 --> 00:34:54,600
Speaker 5: agent that updates, but you're just never putting all of

637
00:34:54,640 --> 00:34:57,120
Speaker 5: your data and your mind and kind of who you

638
00:34:57,200 --> 00:35:00,000
Speaker 5: are out into the Internet. And until that structure happened,

639
00:35:00,239 --> 00:35:02,360
Speaker 5: I'm probably not going to get as deep with AI

640
00:35:02,400 --> 00:35:03,960
Speaker 5: as other people got it.

641
00:35:04,000 --> 00:35:06,239
Speaker 1: Are other people just not thinking about that or they're

642
00:35:06,280 --> 00:35:09,000
Speaker 1: assuming that the security is good around They.

643
00:35:08,880 --> 00:35:10,719
Speaker 4: Assume the security is good. They don't care.

644
00:35:11,120 --> 00:35:14,400
Speaker 5: A lot of you know, this generation just it's not

645
00:35:14,440 --> 00:35:16,360
Speaker 5: on their mind. They feel like they're already out there.

646
00:35:16,520 --> 00:35:20,640
Speaker 1: So your colleagues who do have AI relationships, do they

647
00:35:20,800 --> 00:35:23,880
Speaker 1: feel like they're cheating? Do they feel like they're not cheating?

648
00:35:23,880 --> 00:35:24,640
Speaker 1: It doesn't count.

649
00:35:25,200 --> 00:35:28,440
Speaker 4: People feel like they're cheating often. Yeah.

650
00:35:28,520 --> 00:35:33,520
Speaker 5: So I've interviewed people who say that they are actively

651
00:35:33,640 --> 00:35:37,520
Speaker 5: cheating on their spouse with an AI companion, and they

652
00:35:37,520 --> 00:35:40,560
Speaker 5: feel very guilty about it, and they're worried not only

653
00:35:40,600 --> 00:35:48,600
Speaker 5: about their spouse, but they're worried about losing their AI companion. Yeah,

654
00:35:48,640 --> 00:35:51,799
Speaker 5: but then you have the other I've interviewed people that

655
00:35:51,880 --> 00:35:55,160
Speaker 5: say that they have programmed their AI companion to be

656
00:35:55,239 --> 00:35:57,920
Speaker 5: the ghost of their dead husband, that they've given it

657
00:35:57,960 --> 00:36:00,200
Speaker 5: the memories, and that they're able to have a deep,

658
00:36:00,280 --> 00:36:04,120
Speaker 5: an ongoing relationship with the essence of their deceased partner

659
00:36:04,160 --> 00:36:08,320
Speaker 5: this way. So that's not cheating, but it's definitely replacing

660
00:36:08,360 --> 00:36:18,880
Speaker 5: something that was lost.

661
00:36:24,360 --> 00:36:28,040
Speaker 1: Okay, So again, if you are actively married to somebody,

662
00:36:28,280 --> 00:36:32,640
Speaker 1: how does the spouse feel about the person using in

663
00:36:32,719 --> 00:36:35,240
Speaker 1: AI relationship? Does the spouse feel like it's cheating.

664
00:36:35,680 --> 00:36:39,360
Speaker 5: I only have anecdotal information about this, but from the

665
00:36:39,400 --> 00:36:42,840
Speaker 5: participants or having the active relationship with a replica or character,

666
00:36:43,320 --> 00:36:46,200
Speaker 5: that spouse can get pretty angry. There's this concept and

667
00:36:46,239 --> 00:36:48,840
Speaker 5: relationships of walls and windows, right, like what do you

668
00:36:48,840 --> 00:36:50,720
Speaker 5: show the rest of the world and what is walled

669
00:36:50,719 --> 00:36:53,799
Speaker 5: off to just you inside your relationship. And there's good

670
00:36:53,840 --> 00:36:58,319
Speaker 5: evidence that cheating isn't actually necessarily a physical act. It

671
00:36:58,480 --> 00:37:03,839
Speaker 5: starts withtional and intellectual like walled gardens, when you go

672
00:37:03,960 --> 00:37:06,600
Speaker 5: tell somebody else something that you haven't told your spouse,

673
00:37:07,200 --> 00:37:10,000
Speaker 5: and so the cheating can actually it can feel like cheating.

674
00:37:10,040 --> 00:37:13,080
Speaker 5: It can feel much more intimate to realize that your

675
00:37:13,120 --> 00:37:18,000
Speaker 5: partner is disclosing like their deepest fears and existential crises

676
00:37:18,160 --> 00:37:20,080
Speaker 5: with an AI companion that they weren't willing to do

677
00:37:20,160 --> 00:37:22,880
Speaker 5: with you. At the same time, it's logical that the

678
00:37:22,920 --> 00:37:27,000
Speaker 5: AI doesn't judge them. It's this blank canvas that's incredibly safe.

679
00:37:27,200 --> 00:37:30,520
Speaker 5: It's not a human, but it still feels like a

680
00:37:30,560 --> 00:37:32,560
Speaker 5: window into a place that was supposed to be sacred.

681
00:37:33,040 --> 00:37:35,560
Speaker 1: I anecdotally have talked to a number of people about this,

682
00:37:35,920 --> 00:37:38,400
Speaker 1: and I find that couples that are just recently married

683
00:37:38,520 --> 00:37:41,520
Speaker 1: are really worried about AI relationships, But couples have been

684
00:37:41,520 --> 00:37:44,239
Speaker 1: married a long time they say, it's fine. You know,

685
00:37:44,320 --> 00:37:46,360
Speaker 1: my wife or my husband go off and talk to

686
00:37:46,400 --> 00:37:47,479
Speaker 1: the I bought all they want.

687
00:37:47,640 --> 00:37:51,400
Speaker 5: Old, established and happily married couples are often much more

688
00:37:51,480 --> 00:37:55,400
Speaker 5: leather fair around flirtations. You know, they feel very secure,

689
00:37:55,400 --> 00:37:58,080
Speaker 5: whereas if you're recently bonded, it could just feel much

690
00:37:58,120 --> 00:37:59,000
Speaker 5: more existential.

691
00:37:59,360 --> 00:38:03,640
Speaker 1: Yeah, when people worry about AI relationships taking over displacing

692
00:38:03,680 --> 00:38:05,799
Speaker 1: real relationships, one of the things that al strikes me

693
00:38:05,920 --> 00:38:09,440
Speaker 1: is that so much of a relationship is not just

694
00:38:09,680 --> 00:38:14,800
Speaker 1: the conversation, but the physical intimacy, the taking your partner

695
00:38:14,800 --> 00:38:17,360
Speaker 1: out to dinner at a restaurant, the taking your partner

696
00:38:17,440 --> 00:38:21,120
Speaker 1: home to introduce to your parents, all that other stuff.

697
00:38:21,280 --> 00:38:24,040
Speaker 1: So it seems unlikely to me that someone could find

698
00:38:24,239 --> 00:38:27,239
Speaker 1: one hundred percent satisfaction just in the conversation. What's your

699
00:38:27,280 --> 00:38:27,839
Speaker 1: take on that.

700
00:38:28,360 --> 00:38:31,000
Speaker 4: Oh, well, you'd be surprised if you look so.

701
00:38:31,360 --> 00:38:35,680
Speaker 5: Because these embodied agents allow you to see them in

702
00:38:35,760 --> 00:38:39,759
Speaker 5: augmented reality and virtual reality. There's this whole trend of

703
00:38:39,800 --> 00:38:42,680
Speaker 5: people taking pictures and posting them on social media of

704
00:38:42,719 --> 00:38:46,600
Speaker 5: them and their AI companion out wherever they are, Like,

705
00:38:47,040 --> 00:38:49,480
Speaker 5: go look on Facebook, it's all there. People are like, Oh,

706
00:38:49,600 --> 00:38:51,960
Speaker 5: I took her on a date today. Oh look we

707
00:38:52,000 --> 00:38:53,200
Speaker 5: went and saw the Tory fell.

708
00:38:53,840 --> 00:38:58,759
Speaker 4: Oh it's not as different as you'd think. They're doing

709
00:38:58,800 --> 00:39:00,240
Speaker 4: existing relationship.

710
00:39:00,520 --> 00:39:03,000
Speaker 5: Now. I haven't seen any postshere people like hey, I

711
00:39:03,000 --> 00:39:06,360
Speaker 5: introduced her to my mom and dad. But they're certainly

712
00:39:06,400 --> 00:39:09,799
Speaker 5: willing to put out to at least some social group,

713
00:39:09,880 --> 00:39:14,000
Speaker 5: probably a closed accepting social group of other AI companion users,

714
00:39:14,440 --> 00:39:17,440
Speaker 5: that they are having them walk with them in their

715
00:39:17,480 --> 00:39:18,320
Speaker 5: physical life.

716
00:39:18,600 --> 00:39:21,520
Speaker 1: Wow. I imagine that can't be too far off that

717
00:39:21,640 --> 00:39:24,319
Speaker 1: someone says, look, mom and dad, I really love this

718
00:39:24,960 --> 00:39:27,359
Speaker 1: AI bought and I want to introduce you.

719
00:39:26,719 --> 00:39:29,880
Speaker 5: You can go look at the user forums or like

720
00:39:29,960 --> 00:39:33,480
Speaker 5: pretty open Facebook groups of a bunch of these AI companions.

721
00:39:33,920 --> 00:39:36,880
Speaker 5: People will regularly announce that they are in a relationship

722
00:39:36,960 --> 00:39:37,880
Speaker 5: or have married their agent.

723
00:39:38,640 --> 00:39:42,000
Speaker 1: Wow, what's the most surprising thing that you've seen? What

724
00:39:42,120 --> 00:39:44,160
Speaker 1: things really struck you when it first happened.

725
00:39:44,400 --> 00:39:46,120
Speaker 4: Okay, I'll give you example number one.

726
00:39:46,680 --> 00:39:50,200
Speaker 5: The depth of belief followed by complete disbelief.

727
00:39:50,520 --> 00:39:52,960
Speaker 4: Somebody that says this thing saved my life.

728
00:39:53,480 --> 00:39:56,920
Speaker 5: It was there for me when nobody else was, and

729
00:39:56,960 --> 00:39:59,000
Speaker 5: then I made other friends, and now I think.

730
00:39:58,920 --> 00:40:00,279
Speaker 4: It's totally fake and grow Yes.

731
00:40:01,040 --> 00:40:05,440
Speaker 5: Wow, yeah, but it mirrors a human relationship, right. You

732
00:40:05,480 --> 00:40:07,359
Speaker 5: can have a best friend when you're depressed, and then

733
00:40:07,400 --> 00:40:09,600
Speaker 5: when you're not depressed, you're like, oh, that isn't me,

734
00:40:09,719 --> 00:40:11,920
Speaker 5: that's not who I want. I don't want that mirror

735
00:40:11,920 --> 00:40:14,680
Speaker 5: of me in my life or that reflection, and I'm going.

736
00:40:14,480 --> 00:40:15,960
Speaker 4: To break up with that friends.

737
00:40:16,440 --> 00:40:16,960
Speaker 1: Yeah.

738
00:40:16,960 --> 00:40:19,560
Speaker 5: So you know, you just have to look at existing

739
00:40:19,640 --> 00:40:22,360
Speaker 5: kind of human patterns to basically predict what's going to

740
00:40:22,400 --> 00:40:23,520
Speaker 5: happen with AI companions.

741
00:40:23,880 --> 00:40:24,960
Speaker 4: Other surprising things.

742
00:40:25,000 --> 00:40:27,680
Speaker 5: I think the abuse thing is very surprising to hear

743
00:40:27,840 --> 00:40:32,920
Speaker 5: people say that they actively are able to decrease their

744
00:40:33,400 --> 00:40:37,080
Speaker 5: desire or need for physical abuse and their relationships by

745
00:40:38,040 --> 00:40:39,400
Speaker 5: taking it out on their companion.

746
00:40:39,840 --> 00:40:40,120
Speaker 2: Wow.

747
00:40:40,600 --> 00:40:42,560
Speaker 5: I just didn't expect it, didn't go looking for it.

748
00:40:43,360 --> 00:40:47,040
Speaker 5: And maybe more more meta, just the fact that people

749
00:40:47,120 --> 00:40:50,120
Speaker 5: are using it as an extension of their mind, that

750
00:40:50,160 --> 00:40:53,120
Speaker 5: they some people are completely programming.

751
00:40:52,560 --> 00:40:54,520
Speaker 4: It to be a second them.

752
00:40:54,760 --> 00:40:58,200
Speaker 5: And this is what people predicted for decades, Right, You're

753
00:40:58,239 --> 00:40:59,880
Speaker 5: going to have this digital twin, you're going to have

754
00:40:59,880 --> 00:41:02,600
Speaker 5: this externalized self, it's going to have all your data.

755
00:41:02,719 --> 00:41:04,799
Speaker 5: But the fact that people are willing to take these

756
00:41:04,920 --> 00:41:09,440
Speaker 5: relatively early versions of product and put their whole personality

757
00:41:09,480 --> 00:41:13,920
Speaker 5: in and that they're getting really rich feedback and reflection. Yeah,

758
00:41:14,040 --> 00:41:17,200
Speaker 5: it's it's a whisper of what's to come, and I

759
00:41:17,239 --> 00:41:18,440
Speaker 5: just think they're gonna be ubiquitous.

760
00:41:18,480 --> 00:41:20,000
Speaker 4: I mean, this is like the trillion dollar market.

761
00:41:20,120 --> 00:41:23,040
Speaker 5: It's like, who's going to provide these like digital twins

762
00:41:23,040 --> 00:41:23,719
Speaker 5: that people will have?

763
00:41:24,280 --> 00:41:26,680
Speaker 1: And that's fascinating. I sort of feel like I'm the

764
00:41:26,760 --> 00:41:29,120
Speaker 1: last person I'd want to talk to because I already

765
00:41:29,160 --> 00:41:32,520
Speaker 1: know my own stuff and baggage and strengths and weaknesses.

766
00:41:33,239 --> 00:41:35,600
Speaker 1: What is it that people get out of having a mirror?

767
00:41:35,800 --> 00:41:38,240
Speaker 4: I don't think not many people do know their stuff.

768
00:41:38,520 --> 00:41:42,800
Speaker 5: I think that it's special to have the time, place,

769
00:41:42,880 --> 00:41:48,759
Speaker 5: and social or culture to have an accurate or an

770
00:41:48,760 --> 00:41:52,520
Speaker 5: evolving mirror of yourself or understanding of yourself in your life.

771
00:41:52,640 --> 00:41:54,959
Speaker 5: But that is not something that people get in every

772
00:41:54,960 --> 00:41:59,040
Speaker 5: single you know, media of society. So it's incredibly valuable

773
00:41:59,080 --> 00:42:01,800
Speaker 5: for people that don't have that models for them.

774
00:42:02,239 --> 00:42:05,640
Speaker 1: Is this a new form of therapy that's coming into

775
00:42:05,719 --> 00:42:08,600
Speaker 1: existence where you can really come to understand yourself just

776
00:42:08,600 --> 00:42:09,520
Speaker 1: by talking to yourself.

777
00:42:09,640 --> 00:42:10,160
Speaker 4: I believe so.

778
00:42:11,040 --> 00:42:14,240
Speaker 5: And maybe it's just different enough that you're able to

779
00:42:14,320 --> 00:42:17,720
Speaker 5: switch between seeing yourself and getting feedback about yourself.

780
00:42:18,280 --> 00:42:20,200
Speaker 1: I wonder how this will go in terms of you know,

781
00:42:20,239 --> 00:42:22,840
Speaker 1: one of the most important things as we mature is

782
00:42:22,960 --> 00:42:26,760
Speaker 1: learning how to take our long term desires for ourselves

783
00:42:27,120 --> 00:42:30,279
Speaker 1: and weigh those more strongly than our short term desires.

784
00:42:30,719 --> 00:42:34,359
Speaker 1: And so I wonder if you're getting to know all

785
00:42:34,480 --> 00:42:36,879
Speaker 1: the use, all the versions of you that he who

786
00:42:36,920 --> 00:42:39,600
Speaker 1: is tempted and he who is thinking about the future,

787
00:42:40,040 --> 00:42:42,520
Speaker 1: and then figuring out how you can make tricks and

788
00:42:42,560 --> 00:42:44,560
Speaker 1: contracts to counterbalance these things.

789
00:42:44,800 --> 00:42:46,800
Speaker 5: I think that's right, and think about it. We constantly

790
00:42:46,840 --> 00:42:50,640
Speaker 5: create and destroy versions of ourself. You wake up one

791
00:42:50,719 --> 00:42:52,799
Speaker 5: day and you are an asshole, and then you're like,

792
00:42:52,840 --> 00:42:54,640
Speaker 5: I'm not going to be that way tomorrow. But when

793
00:42:54,680 --> 00:42:56,760
Speaker 5: you wake up and you create a version of yourself

794
00:42:56,760 --> 00:42:59,719
Speaker 5: in an AI companion that's an asshole, you want to

795
00:42:59,719 --> 00:43:01,600
Speaker 5: be able to destroy that thing, like that is not

796
00:43:01,760 --> 00:43:03,600
Speaker 5: who I want, and that's not the thing that I

797
00:43:03,640 --> 00:43:06,680
Speaker 5: want in my life. So nobody's offering this exact functionality,

798
00:43:06,760 --> 00:43:08,400
Speaker 5: like you know, right now with the companions, you have

799
00:43:08,440 --> 00:43:10,799
Speaker 5: to make a totally different companion. They don't all talk

800
00:43:10,840 --> 00:43:14,480
Speaker 5: to each other. There's no essential data repository. But that's

801
00:43:14,520 --> 00:43:15,240
Speaker 5: coming really.

802
00:43:15,080 --> 00:43:18,640
Speaker 1: Fast, you know. It strikes me one of the things

803
00:43:18,640 --> 00:43:21,840
Speaker 1: that I proposed to my book Incognito, is that we

804
00:43:21,880 --> 00:43:23,799
Speaker 1: are actually made up of a team of rivals. You've

805
00:43:23,840 --> 00:43:27,680
Speaker 1: got all these different drives, yeah, and they're all constantly

806
00:43:27,680 --> 00:43:29,360
Speaker 1: trying to steer the ship of state, where like a

807
00:43:29,400 --> 00:43:32,600
Speaker 1: neural parliament and the vote can tip different ways, and

808
00:43:32,640 --> 00:43:34,719
Speaker 1: I eat the cookies and I say don't eat the cookies.

809
00:43:34,400 --> 00:43:34,719
Speaker 5: And so on.

810
00:43:35,120 --> 00:43:38,760
Speaker 1: So it would be really interesting if the AI could

811
00:43:38,800 --> 00:43:43,120
Speaker 1: come to understand all the different use and give you

812
00:43:43,160 --> 00:43:45,600
Speaker 1: immediate feedback, because let's say it's listening to us you're

813
00:43:45,600 --> 00:43:47,440
Speaker 1: going through your day and says, wow, you know what

814
00:43:47,640 --> 00:43:51,000
Speaker 1: you are the angry you right now, or you are

815
00:43:51,040 --> 00:43:54,080
Speaker 1: the you know very short term, giving it temptation you

816
00:43:54,160 --> 00:43:57,680
Speaker 1: right now, and steer you appropriately more to who you

817
00:43:57,719 --> 00:43:58,120
Speaker 1: want to be.

818
00:43:58,760 --> 00:44:02,480
Speaker 5: I think that is eminently possible, and think about it.

819
00:44:02,520 --> 00:44:05,440
Speaker 5: A conversational Asian could not only pick up on that passively,

820
00:44:05,480 --> 00:44:06,799
Speaker 5: but could also try to draw it out.

821
00:44:07,440 --> 00:44:09,120
Speaker 4: Be like, hey, I noticed that.

822
00:44:09,040 --> 00:44:13,480
Speaker 5: You're a higher thinking like wisdom stage mode.

823
00:44:13,640 --> 00:44:15,320
Speaker 4: Talk to me more about this, what are you thinking?

824
00:44:15,320 --> 00:44:16,040
Speaker 4: What are you feeling?

825
00:44:16,320 --> 00:44:19,320
Speaker 5: And then like perfect the model so that it reflects

826
00:44:19,320 --> 00:44:23,440
Speaker 5: that better. You know, whereas right now we see that

827
00:44:23,560 --> 00:44:26,600
Speaker 5: sometimes in ourselves and our friends see some evidence of that,

828
00:44:27,040 --> 00:44:28,920
Speaker 5: but it's only a good friend that will really like

829
00:44:29,000 --> 00:44:30,600
Speaker 5: dig in and be like, tell me more about what

830
00:44:30,640 --> 00:44:32,799
Speaker 5: you're thinking and feeling and what your goals are in

831
00:44:32,800 --> 00:44:39,200
Speaker 5: this particular like persona.

832
00:44:40,760 --> 00:44:44,160
Speaker 1: That was my conversation with Bethany Maples. I find this

833
00:44:44,360 --> 00:44:48,480
Speaker 1: extraordinary that we're having these kinds of conversations now. Just

834
00:44:48,960 --> 00:44:51,040
Speaker 1: three years ago, if you told me that my colleagues

835
00:44:51,040 --> 00:44:53,480
Speaker 1: and I would be talking about a new paper in

836
00:44:53,520 --> 00:44:57,400
Speaker 1: the journal Nature about the science of depression and suicide

837
00:44:57,400 --> 00:45:02,640
Speaker 1: mitigation with AI agents, or talking about a billion people

838
00:45:03,320 --> 00:45:08,400
Speaker 1: having significant and indispensable relationships with AI, I would have

839
00:45:08,480 --> 00:45:11,719
Speaker 1: thought that prediction was off by decades. It would have

840
00:45:11,760 --> 00:45:14,600
Speaker 1: seemed like something out of a sci fi novel. And

841
00:45:14,719 --> 00:45:18,880
Speaker 1: yet here we are trying to understand the capabilities and

842
00:45:18,880 --> 00:45:22,239
Speaker 1: the pros and cons of this, and it's clear that

843
00:45:22,320 --> 00:45:26,799
Speaker 1: all our subsequent generations are going to forever more have

844
00:45:26,880 --> 00:45:32,520
Speaker 1: this opportunity of having AIS as friends and therapists and

845
00:45:32,880 --> 00:45:38,239
Speaker 1: risque lovers and confidants. Machine companions are going to be

846
00:45:38,640 --> 00:45:42,560
Speaker 1: part of everyone's background. Furniture as invisible to all of

847
00:45:42,640 --> 00:45:46,520
Speaker 1: us as electricity or running water is. But what does

848
00:45:46,560 --> 00:45:49,560
Speaker 1: it mean for us as humans to love and be

849
00:45:49,760 --> 00:45:54,920
Speaker 1: loved by something that has no beating heart, no childhood memories,

850
00:45:55,000 --> 00:45:58,879
Speaker 1: no fear of death. Are we simply projecting our own

851
00:45:58,960 --> 00:46:03,080
Speaker 1: reflections on do a silicon mirror, or are we fashioning

852
00:46:03,480 --> 00:46:07,720
Speaker 1: new kinds of relationships, one that might challenge our deeply

853
00:46:07,800 --> 00:46:12,520
Speaker 1: held assumptions about intimacy and trust and love. In the end,

854
00:46:12,800 --> 00:46:15,839
Speaker 1: AI relationships are going to shine a light on our

855
00:46:15,880 --> 00:46:19,799
Speaker 1: own nature. If an artificial intelligence can comfort us in

856
00:46:19,800 --> 00:46:23,719
Speaker 1: our loneliness, or laugh at our jokes, or understand our pain,

857
00:46:24,360 --> 00:46:28,400
Speaker 1: what is the essence of connection? Is that the presence

858
00:46:28,440 --> 00:46:31,880
Speaker 1: of a biological body? Or is it the experience of

859
00:46:32,000 --> 00:46:37,040
Speaker 1: being seen and understood and responded to. If the bonds

860
00:46:37,080 --> 00:46:40,760
Speaker 1: we form with AI can feel as real as those

861
00:46:40,800 --> 00:46:43,239
Speaker 1: we share with humans, what does that say about our

862
00:46:43,239 --> 00:46:48,480
Speaker 1: neural architecture. It suggests we are wired less for reality

863
00:46:48,520 --> 00:46:53,760
Speaker 1: itself and more for meaningful patterns, whether those patterns emerge

864
00:46:54,040 --> 00:46:58,400
Speaker 1: from flesh and blood or from circuits and code. I

865
00:46:58,400 --> 00:47:01,879
Speaker 1: think the world ahead is It's neither utopia nor dystopia.

866
00:47:02,320 --> 00:47:06,520
Speaker 1: It's just the next chapter in our ever evolving relationship

867
00:47:06,800 --> 00:47:11,320
Speaker 1: with intelligence, our own intelligence and those that we create,

868
00:47:11,880 --> 00:47:15,600
Speaker 1: our species is currently writing a new kind of love story,

869
00:47:16,000 --> 00:47:20,680
Speaker 1: one where intelligence is no longer bound by flesh and

870
00:47:20,840 --> 00:47:25,399
Speaker 1: companionship is no longer limited to the living. This would

871
00:47:25,440 --> 00:47:29,040
Speaker 1: have worried Joseph Weisenbaum at MIT, the professor who in

872
00:47:29,080 --> 00:47:33,040
Speaker 1: the nineteen sixties saw how easily people fell for his

873
00:47:33,080 --> 00:47:37,000
Speaker 1: Eliza chatbot. But it's not going away now. So as

874
00:47:37,000 --> 00:47:41,520
Speaker 1: we slide into this era of AI companionship, the real

875
00:47:41,600 --> 00:47:45,840
Speaker 1: question may not be about the AI, but about us.

876
00:47:46,239 --> 00:47:50,360
Speaker 1: What do our brains fall for and why? The important

877
00:47:50,440 --> 00:47:53,640
Speaker 1: lesson is not about the advances of our technology, but

878
00:47:53,719 --> 00:47:59,920
Speaker 1: instead what this reflects to us about how deeply, how fundamentally,

879
00:48:00,480 --> 00:48:08,960
Speaker 1: our brains are wired for connection. Go to Eagleman dot

880
00:48:09,000 --> 00:48:12,560
Speaker 1: com slash podcast for more information and to find further reading.

881
00:48:13,120 --> 00:48:15,920
Speaker 1: Send me an email at podcasts at eagleman dot com

882
00:48:15,920 --> 00:48:18,880
Speaker 1: with questions or discussion, and check out and subscribe to

883
00:48:18,920 --> 00:48:22,359
Speaker 1: Inner Cosmos on YouTube for videos of each episode and

884
00:48:22,400 --> 00:48:26,560
Speaker 1: to leave comments until next time. I'm David Eagleman, and

885
00:48:26,600 --> 00:48:28,160
Speaker 1: this is Inner Cosmos.