1 00:00:11,200 --> 00:00:13,200 Hi, Sam. Hey, Elaine. 2 00:00:14,240 --> 00:00:17,960 I was just looking back at our chat history in Slack. 3 00:00:18,440 --> 00:00:21,360 I remembered that the other day you were telling me all about 4 00:00:21,360 --> 00:00:24,360 how you were really busy ramping up with a new client. 5 00:00:24,360 --> 00:00:27,120 You were having fun, but really feeling a bit overwhelmed. 6 00:00:27,120 --> 00:00:29,880 You had a lot on your plate and you weren't getting much sleep. 7 00:00:29,880 --> 00:00:32,960 You were, you know, finding it hard to get an exercise. 8 00:00:33,040 --> 00:00:37,560 And so I went to see what was my response to this like, what did 9 00:00:37,560 --> 00:00:39,600 I reply? Because in my head I remember 10 00:00:39,600 --> 00:00:43,600 thinking, really feeling for you, really empathizing. 11 00:00:43,600 --> 00:00:46,240 So I checked the record. There were lots of topics, 12 00:00:46,240 --> 00:00:48,520 different work and personal updates. 13 00:00:48,520 --> 00:00:52,320 So I replied to a few of them, maybe the more task oriented 14 00:00:52,440 --> 00:00:54,800 portions. But I actually didn't say 15 00:00:54,800 --> 00:00:59,280 anything in response to how you were doing nothing. 16 00:00:59,360 --> 00:01:01,480 So my question for you is, did you notice that? 17 00:01:01,720 --> 00:01:05,360 Did you feel like there was an empathy gap there where you were 18 00:01:05,360 --> 00:01:07,760 like, man, I just told you about this hard time that I was having 19 00:01:07,760 --> 00:01:11,520 and you just ignored it? You know what, it's interesting. 20 00:01:11,640 --> 00:01:14,680 I do think I picked up on some version of it. 21 00:01:14,760 --> 00:01:17,840 It's hard for me to exactly remember now, the messages and 22 00:01:17,840 --> 00:01:21,600 so on, but I remember there was maybe some form of gap. 23 00:01:21,600 --> 00:01:25,920 But also like I have this thing, some version of Han Lon's racer. 24 00:01:26,000 --> 00:01:29,280 It's like a version of Occam's Racer where instead of 25 00:01:29,280 --> 00:01:32,040 attributing malice to certain things, you assume someone is 26 00:01:32,040 --> 00:01:35,360 busy or other things get in the way or like various things that 27 00:01:35,360 --> 00:01:39,760 can leave us not being our best versions of ourselves even 28 00:01:39,760 --> 00:01:43,000 though we try. And so for me, I saw that, OK, 29 00:01:43,000 --> 00:01:45,800 Eileen didn't really acknowledge this, but I know she's also 30 00:01:45,800 --> 00:01:49,880 dealing with two young kids, like traveling for wedding and 31 00:01:49,880 --> 00:01:54,160 all of this kind of stuff. So yeah, I attributed it to 32 00:01:54,160 --> 00:01:57,040 that. Not so much of A lack of general 33 00:01:57,040 --> 00:02:00,000 empathy, but like just available empathy at that time. 34 00:02:00,960 --> 00:02:05,000 Wow, I do appreciate that, first of all, and I like that approach 35 00:02:05,000 --> 00:02:07,880 of assume the best and the other person because you know that 36 00:02:07,880 --> 00:02:10,960 they do care about you, that feeling human being on the other 37 00:02:10,960 --> 00:02:13,480 side. But I did one thing because I 38 00:02:13,480 --> 00:02:15,040 wanted to do a little comparison. 39 00:02:15,080 --> 00:02:18,960 OK, so now I've heard your response to the all in version, 40 00:02:18,960 --> 00:02:21,920 which was maybe a little bit lacking in empathy. 41 00:02:21,920 --> 00:02:25,560 And then I asked chat GBTI actually gave ChatGPT the 42 00:02:25,560 --> 00:02:29,280 transcript of your voice message and I said, respond to this. 43 00:02:30,040 --> 00:02:34,200 I didn't even say respond to this in an empathetic way. 44 00:02:34,200 --> 00:02:37,000 I just said, can you craft a response to Sam? 45 00:02:37,480 --> 00:02:41,040 And so I got back your very typical ChatGPT language. 46 00:02:41,040 --> 00:02:43,720 Here's a warm and thoughtful response you could send back to 47 00:02:43,720 --> 00:02:47,040 Sam that acknowledges everything he shared while gently moving 48 00:02:47,040 --> 00:02:49,960 things forward. The work. 49 00:02:51,160 --> 00:02:53,720 So the first version that I got was too long. 50 00:02:53,720 --> 00:02:57,800 You know, a lots of thank yous and a really lame joke. 51 00:02:57,800 --> 00:03:00,440 Like clipping about some like so so bad. 52 00:03:00,440 --> 00:03:03,400 Like it's not even a joke. Like I might need to steal that 53 00:03:03,400 --> 00:03:05,960 line about something that was not even funny that you said. 54 00:03:06,920 --> 00:03:08,680 Totally hear you on the hectic stretch. 55 00:03:08,920 --> 00:03:11,560 I've been in a similar boat lately, blah blah blah. 56 00:03:11,560 --> 00:03:13,560 So no pressure on anything from your side. 57 00:03:13,840 --> 00:03:15,040 Love that you've been nerding out. 58 00:03:15,200 --> 00:03:19,040 So just like a response to everything, very long winded and 59 00:03:19,040 --> 00:03:22,600 so on. And it ends essentially with 60 00:03:22,720 --> 00:03:24,840 hope you got that apple in a moment to breathe. 61 00:03:25,080 --> 00:03:33,000 Grateful for you. So I guess my question is, if I 62 00:03:33,000 --> 00:03:37,680 had sent you, imagine that you got this version of a response. 63 00:03:37,680 --> 00:03:41,920 For me, it's like a short novel. First of all, you didn't know 64 00:03:41,920 --> 00:03:44,640 that it was written by ChatGPT and you got this. 65 00:03:44,960 --> 00:03:49,040 What would you think? Because I'm so deep in the AI 66 00:03:49,040 --> 00:03:51,320 trenches, I would probably just notice like, OK, this is very 67 00:03:51,320 --> 00:03:54,840 much AI language, so I'll probably just notice that you 68 00:03:54,840 --> 00:03:57,120 used it. Part of me would say, OK, I 69 00:03:57,120 --> 00:04:00,320 don't want to have you writing anything that's not just like 70 00:04:00,400 --> 00:04:03,920 you writing whatever. Part of you in terms of if you 71 00:04:03,920 --> 00:04:06,120 don't have empathy that moment, like that's OK. 72 00:04:06,120 --> 00:04:08,880 But of course, if you wrote what actually seemed like a really 73 00:04:08,880 --> 00:04:11,600 nice empathic message that I could not tell was AI, it's 74 00:04:11,600 --> 00:04:14,480 like, I'm sure it would be like, that was so nice. 75 00:04:14,480 --> 00:04:18,760 That was exactly what I needed. A few things stand out to me in 76 00:04:18,800 --> 00:04:22,200 your response. One is, would you then expect 77 00:04:22,200 --> 00:04:24,720 that from me? And then like the bar for 78 00:04:24,720 --> 00:04:27,200 empathy sort of gets higher overtime? 79 00:04:27,200 --> 00:04:30,920 Like I start shelling out the empathy in this simulated 80 00:04:30,920 --> 00:04:33,200 artificial way whether you know it or not. 81 00:04:33,200 --> 00:04:36,920 Like maybe at some point I would tell you, but like one, do you 82 00:04:36,920 --> 00:04:40,640 get so accustomed to it that we sort of have a situation of 83 00:04:40,840 --> 00:04:46,280 empathy inflation, right? To does this start deskilling me 84 00:04:46,280 --> 00:04:48,720 in terms of my ability? We've demonstrated that I'm not 85 00:04:48,720 --> 00:04:51,560 that good at empathy in the 1st place, but like, do I get even 86 00:04:51,560 --> 00:04:54,800 worse at empathy? Is it even harder and harder 87 00:04:54,800 --> 00:04:57,440 over time? And I think it's a third thing 88 00:04:57,440 --> 00:05:01,200 in terms of how this feels. And what you're accustomed to is 89 00:05:01,200 --> 00:05:05,760 like, you do sort of get used to the not a lean version, but the 90 00:05:05,760 --> 00:05:09,360 actual AI version when you're interacting with some sort of 91 00:05:09,360 --> 00:05:15,560 LLM that may have really been built for sycophancy, where it's 92 00:05:15,560 --> 00:05:18,360 like, you know, telling you what you want to hear, agreeing even 93 00:05:18,360 --> 00:05:22,480 when maybe it shouldn't. And like, what are the downsides 94 00:05:22,480 --> 00:05:27,040 and risks to that? I think actually I'm definitely 95 00:05:27,040 --> 00:05:28,440 less worried about the second one. 96 00:05:28,440 --> 00:05:32,160 Like, I don't think that your personality would drastically 97 00:05:32,160 --> 00:05:34,760 change over time because you're outsourcing some of your. 98 00:05:35,040 --> 00:05:39,640 Already hit bottom. Not at all, but. 99 00:05:40,000 --> 00:05:44,120 There's a floor effect. I think the first part is the 100 00:05:44,120 --> 00:05:48,800 thing that I would be expecting the biggest impact from. 101 00:05:48,800 --> 00:05:52,440 Basically, we set expectations all the time. 102 00:05:52,440 --> 00:05:54,920 You know, I think we can all relate to this. 103 00:05:54,920 --> 00:05:57,080 When we start working with someone, we're kind of like 104 00:05:57,080 --> 00:06:00,800 trying to attune like, OK, this person sending emojis, do they 105 00:06:00,800 --> 00:06:03,720 not send emojis? OK, this person send a lot of 106 00:06:03,720 --> 00:06:05,920 emojis usually. So then, like, do you expect 107 00:06:05,920 --> 00:06:08,240 them to always send a lot of emoji? 108 00:06:08,240 --> 00:06:11,120 And if there's no emoji, you're like, Oh my God, is this person 109 00:06:11,120 --> 00:06:12,640 OK? You know what's really going on 110 00:06:12,640 --> 00:06:14,000 here? I've literally thought that 111 00:06:14,000 --> 00:06:15,880 about you. Like I've thought, oh, is 112 00:06:15,880 --> 00:06:18,000 something wrong? Sam didn't send me an emoji. 113 00:06:18,000 --> 00:06:19,520 Oh, you are that person. Yes. 114 00:06:19,520 --> 00:06:23,240 You're describing yourself. I thought you knew that. 115 00:06:23,800 --> 00:06:26,440 No, Yeah, yeah, yeah. OK, funny. 116 00:06:28,440 --> 00:06:32,880 So yeah, I would definitely think that in some ways that is 117 00:06:32,880 --> 00:06:35,480 something we want to attune in a good way. 118 00:06:35,480 --> 00:06:39,200 Like what is the best long haul thing that you could sustain 119 00:06:40,120 --> 00:06:41,520 where you don't want to go over the top? 120 00:06:42,000 --> 00:06:45,280 Because that becomes also like an empathy burden if you always 121 00:06:45,280 --> 00:06:50,320 expected to have this like maximum empathy with all 122 00:06:50,320 --> 00:06:53,720 different contexts. I have a sister who's like 123 00:06:54,120 --> 00:06:56,240 finding out the hard way. That is like really hard when 124 00:06:56,240 --> 00:07:00,040 you have had a lot of people used to be able to lean against 125 00:07:00,040 --> 00:07:03,000 your shoulder because you've been like being very selfless 126 00:07:03,040 --> 00:07:05,800 overtime when you then start to be like, OK, I'm going to 127 00:07:05,800 --> 00:07:07,600 prioritize a little bit of my own needs. 128 00:07:08,040 --> 00:07:10,520 People are like, but what about me? 129 00:07:10,520 --> 00:07:12,600 Like, I need to like, then you have all of these people that 130 00:07:12,600 --> 00:07:15,960 are like coming and expecting your help to solve their 131 00:07:15,960 --> 00:07:17,760 problems. And I think that's kind of the 132 00:07:17,760 --> 00:07:20,520 same thing that could happen very easily if you use AI to 133 00:07:20,600 --> 00:07:23,040 maximize empathy. A lot of people are just going 134 00:07:23,040 --> 00:07:25,920 to be like, Oh my God, Elaine is the person I should always come 135 00:07:25,920 --> 00:07:28,480 to for empathy. Well, you're describing this is 136 00:07:28,480 --> 00:07:31,560 in the context of your sister, but this is really an issue that 137 00:07:31,560 --> 00:07:35,280 females in general suffer from as they're expected to be, you 138 00:07:35,480 --> 00:07:38,080 know, more warm and caring and motherly. 139 00:07:38,080 --> 00:07:41,440 And you know, maybe not all of us are that warm and caring and 140 00:07:41,440 --> 00:07:46,680 motherly, but we don't get the sort of leadership boost that 141 00:07:46,680 --> 00:07:50,360 you would by saying the same clear strategic demanding line 142 00:07:50,840 --> 00:07:53,600 that a male would. For sure, yeah. 143 00:07:53,880 --> 00:07:58,360 If those already are some form of societal biases or faulty 144 00:07:58,360 --> 00:08:03,280 expectations, I feel like this is partly one of the dangers of 145 00:08:03,320 --> 00:08:07,280 AI where it could actually make it easier for us to slip into 146 00:08:07,680 --> 00:08:10,200 and reinforce those things even more because we're kind of like, 147 00:08:11,160 --> 00:08:13,520 you know, in this case, if you feel like you have those 148 00:08:13,520 --> 00:08:19,600 expectations on you, you then may be more likely to use AI to 149 00:08:19,600 --> 00:08:22,160 reinforce those expectations and then creating this bad 150 00:08:22,200 --> 00:08:24,200 reinforcing cycle. Interesting. 151 00:08:24,200 --> 00:08:28,720 So you may have fought back against the man before, but now 152 00:08:28,760 --> 00:08:31,720 you're like, oh, it's easy to just turn this into a warm 153 00:08:31,720 --> 00:08:34,039 caring phrase. Maybe I'll just outsource it. 154 00:08:34,480 --> 00:08:37,240 Maybe, you know, actually this has happened. 155 00:08:37,559 --> 00:08:38,919 My husband and I were really laughing. 156 00:08:38,919 --> 00:08:42,320 He got a text from someone where it included that little preamble 157 00:08:42,320 --> 00:08:45,960 that you get from ChatGPT. That's like, here's a response 158 00:08:45,960 --> 00:08:50,440 that you can send that like balances the firmness of your 159 00:08:50,440 --> 00:08:53,920 need with a friendly tone, you know, like making that up. 160 00:08:54,240 --> 00:08:59,200 But in general, we were just so amazed that this person chose to 161 00:08:59,200 --> 00:09:03,160 use ChatGPT for a text. And then when he found out it 162 00:09:03,160 --> 00:09:05,280 was his mom it was like even worse. 163 00:09:08,640 --> 00:09:09,440 Right. Oh my God. 164 00:09:09,880 --> 00:09:12,080 I think we should probably get into the episode of today, but 165 00:09:12,080 --> 00:09:15,800 it also makes me think about the episode we had reasoning with 166 00:09:15,800 --> 00:09:17,960 Jonah. We had this question towards 167 00:09:17,960 --> 00:09:24,360 Anne about how she would feel about someone having AI writing 168 00:09:24,360 --> 00:09:27,680 a breakup message versus like also like a pick me up some form 169 00:09:27,680 --> 00:09:31,440 of dating messages. And yeah, I almost feel like 170 00:09:31,440 --> 00:09:35,760 there maybe should be like a new word for that. 171 00:09:35,760 --> 00:09:41,120 Like when you realize someone has given you a sense of kind of 172 00:09:41,120 --> 00:09:44,800 empathy or care or whatever, and it's like, oh, that's a very 173 00:09:44,800 --> 00:09:47,560 thoughtful thing. And then you realize, actually 174 00:09:48,440 --> 00:09:50,840 they didn't do it themselves. There was just AI who did it for 175 00:09:50,840 --> 00:09:53,000 them. It's a new type of betrayal. 176 00:09:53,440 --> 00:09:55,320 Yeah, it is. It's awful. 177 00:09:55,600 --> 00:09:58,080 But since you mentioned Janice, she does have some really 178 00:09:58,080 --> 00:10:02,360 interesting research arguing that the level of empathy or the 179 00:10:02,600 --> 00:10:06,960 degree to which your AI generated response should be 180 00:10:07,320 --> 00:10:10,880 empathetic really has to depend on the situation. 181 00:10:10,880 --> 00:10:15,240 Because there are some contacts that are more deserving of 182 00:10:15,240 --> 00:10:18,200 higher empathy than others. And sometimes it's not as 183 00:10:18,200 --> 00:10:21,880 appropriate, which I think our guest today would agree with. 184 00:10:22,440 --> 00:10:26,200 So let's talk about Mickey or Michael Inslecht. 185 00:10:26,520 --> 00:10:30,400 Yeah, he is the professor of psychology at the University of 186 00:10:30,400 --> 00:10:33,920 Toronto and an occasional podcast for himself. 187 00:10:33,920 --> 00:10:37,960 I've been a big fan of Two Psychologists and Four Beers, 188 00:10:38,480 --> 00:10:41,680 which it does or has done at least with Joel in bar for a 189 00:10:41,680 --> 00:10:43,360 long time. I think it's less of a frequent 190 00:10:43,360 --> 00:10:46,440 Co host these days, but that's definitely been one of those 191 00:10:46,600 --> 00:10:50,400 really nice podcasts to complement our podcast with. 192 00:10:50,440 --> 00:10:54,400 And he studied a Rachel topic related to emotion, 193 00:10:54,400 --> 00:10:58,280 self-control, and even boredom. And recently, he's extended his 194 00:10:58,280 --> 00:11:02,520 work in empathy, looking at how it relates to AI. 195 00:11:03,000 --> 00:11:04,560 Yeah. And so of course we talked all 196 00:11:04,560 --> 00:11:08,800 about this work. We started with what empathy is 197 00:11:08,920 --> 00:11:13,040 in the first place for humans, how we experience empathy as 198 00:11:13,040 --> 00:11:16,480 both givers and then also on the receiving side and really how 199 00:11:16,480 --> 00:11:19,600 effortful it is. So I'm not alone in thinking 200 00:11:19,600 --> 00:11:24,560 that, you know, it does take some work to experience and give 201 00:11:24,560 --> 00:11:26,920 empathy. But then importantly, how this 202 00:11:26,920 --> 00:11:30,600 all translates to AI, this performance of empathy, not the 203 00:11:30,600 --> 00:11:35,760 experience of it. And Mickey and his co-authors in 204 00:11:35,760 --> 00:11:39,280 this one paper, which I feel like maybe it's going to be a 205 00:11:39,280 --> 00:11:43,720 Seminole work in empathic AI, he wrote with one of our former 206 00:11:43,720 --> 00:11:48,040 guests, Paul Bloom and Darrell Cameron and Jason de Cruz, this 207 00:11:48,040 --> 00:11:52,080 paper in praise of empathic AI. They basically argue that it 208 00:11:52,080 --> 00:11:55,720 doesn't really matter if AI can experience empathy. 209 00:11:55,960 --> 00:11:59,800 It's not really important for this eudaimonic sense, but 210 00:11:59,800 --> 00:12:02,960 really focus instead on the utility of it. 211 00:12:02,960 --> 00:12:06,680 So the potential benefits and the potential downsides. 212 00:12:06,920 --> 00:12:10,760 And so we talked to Mickey a lot about both of these sides. 213 00:12:11,160 --> 00:12:13,720 It's really, it's never one or the other, as we know. 214 00:12:14,160 --> 00:12:17,920 And where are the cases that these simulated expressions of 215 00:12:17,920 --> 00:12:22,120 empathy can actually improve user well-being and really get 216 00:12:22,120 --> 00:12:25,160 around those biases and inconsistencies that plague 217 00:12:25,160 --> 00:12:30,680 human empathy? Then, very interestingly, one of 218 00:12:30,680 --> 00:12:35,320 my favorite very recent papers from Mickey is on the AI empathy 219 00:12:35,320 --> 00:12:38,280 choice paradox. And this is just the finding 220 00:12:38,280 --> 00:12:41,320 that people say that they'd rather receive empathy from a 221 00:12:41,320 --> 00:12:46,120 human, even though when they actually rate AI responses, they 222 00:12:46,120 --> 00:12:48,160 think that they're higher in empathy. 223 00:12:48,160 --> 00:12:51,040 They're warmer, more effortful, more emotionally satisfying. 224 00:12:51,480 --> 00:12:54,400 This is a really interesting finding because we have so many 225 00:12:54,400 --> 00:13:00,000 cases now where AI responses are rated as more empathetic, even 226 00:13:00,000 --> 00:13:03,520 though people will still say, yeah, I'll take the human the 227 00:13:03,520 --> 00:13:07,480 majority of the time. And we get into this and much 228 00:13:07,480 --> 00:13:09,840 more in our conversation with me. 229 00:13:16,520 --> 00:13:20,960 Happens. To Murgatroyd. 230 00:13:22,240 --> 00:13:25,960 I'm very excited to say welcome Michael to the Behavioral Design 231 00:13:25,960 --> 00:13:28,080 Podcast. Thanks for having me, it's a 232 00:13:28,080 --> 00:13:30,040 pleasure to be here. Yeah. 233 00:13:30,080 --> 00:13:33,360 For us, this is we've really been interested in exploring 234 00:13:33,360 --> 00:13:36,000 this kind of intersection of AI and behavioral science. 235 00:13:36,080 --> 00:13:39,440 And we really look forward to speaking to you when it comes to 236 00:13:39,440 --> 00:13:44,400 empathy and AI. And maybe we can start with 237 00:13:44,400 --> 00:13:47,040 used, you can set the scene a little bit how you get 238 00:13:47,040 --> 00:13:51,480 interested in especially this kind of overlap of AI and 239 00:13:51,600 --> 00:13:54,120 empathy. Yeah, OK, Maybe I could step 240 00:13:54,120 --> 00:13:56,480 even further back, which is like my origin story of studying 241 00:13:56,480 --> 00:13:57,640 empathy. Yes. 242 00:13:58,600 --> 00:14:02,720 So I consider myself an accidental researcher of empathy 243 00:14:02,720 --> 00:14:05,840 by accidental meaning. It wasn't like at least I didn't 244 00:14:05,840 --> 00:14:07,800 think it was a major interest of mine. 245 00:14:08,480 --> 00:14:10,560 It was something that I would nibble out a little bit of here, 246 00:14:10,560 --> 00:14:13,720 a little bit there. But as I look back in my career, 247 00:14:13,720 --> 00:14:16,640 I see I've been nibbling consistently for the past 15 248 00:14:16,640 --> 00:14:20,000 years on this topic. The place where I really started 249 00:14:20,240 --> 00:14:23,280 digging into it and really getting into it as more than 250 00:14:23,280 --> 00:14:27,760 just a sideline is when I work with Dale Cameron on the 251 00:14:27,760 --> 00:14:31,160 cognitive costs of empathy. So my main line of research is 252 00:14:31,160 --> 00:14:34,720 on motivation and effort, and I'm fascinated by the topic of 253 00:14:34,720 --> 00:14:36,320 effort. Effort is something that we all 254 00:14:36,320 --> 00:14:38,600 recognize what it is. It's when we were trying, we're 255 00:14:38,600 --> 00:14:41,360 pushing ourselves, but it feels uncomfortable. 256 00:14:41,360 --> 00:14:44,920 It doesn't feel great. An economist and cognitive 257 00:14:44,920 --> 00:14:48,000 neuroscientist like to say that effort is costly. 258 00:14:48,040 --> 00:14:51,080 We have to pay some sort of cost to engage in effort, full 259 00:14:51,080 --> 00:14:53,280 activities. And then I made this connection 260 00:14:53,320 --> 00:14:56,640 with empathy being effortful and maybe just projecting from my 261 00:14:56,640 --> 00:14:58,960 own psychology. I remember hearing and even 262 00:14:58,960 --> 00:15:02,160 studying research in the early 2000s where people refer to 263 00:15:02,160 --> 00:15:04,200 empathy as being automatic, effortless. 264 00:15:04,200 --> 00:15:05,520 It just kind of happened to you, no? 265 00:15:06,120 --> 00:15:07,560 Not what I. Mean. 266 00:15:08,840 --> 00:15:12,080 I mean, sometimes. Yeah, sometimes I look at myself 267 00:15:12,200 --> 00:15:15,640 and I was like, no, like, and I have young children and I 268 00:15:15,760 --> 00:15:17,080 clearly feel empathic towards them. 269 00:15:17,080 --> 00:15:20,800 But sometimes I'm, like, tired and want to get to work done and 270 00:15:20,800 --> 00:15:24,160 like, why are you bothering me with your silly crying over 271 00:15:24,160 --> 00:15:26,680 something that doesn't really matter, having being a very 272 00:15:26,680 --> 00:15:28,320 unempathic dad? So then I started playing with 273 00:15:28,320 --> 00:15:30,320 like, maybe it's not as easy as we think. 274 00:15:30,600 --> 00:15:34,000 Darrell and I ran it, like, for years, a bunch of experiments 275 00:15:34,280 --> 00:15:37,360 where we essentially demonstrated like, that it just 276 00:15:37,360 --> 00:15:41,280 looked, all else being equal, people prefer to avoid empathy 277 00:15:41,280 --> 00:15:43,200 at least repeatedly to strangers. 278 00:15:43,840 --> 00:15:46,280 They prefer avoiding empathy than engaging in empathy. 279 00:15:46,280 --> 00:15:48,720 So empathy has these flaws. It's hard. 280 00:15:48,960 --> 00:15:51,040 People don't readily do it. They get burnt out from it. 281 00:15:51,480 --> 00:15:53,680 And when we first started talking about this stuff, people 282 00:15:53,680 --> 00:15:57,040 really didn't like this idea. But we got pushed back. 283 00:15:57,120 --> 00:15:59,960 Being, no, empathy is automatic. It just happens to us. 284 00:16:00,240 --> 00:16:02,200 Don't you know, when you watch a movie and you're crying, you're 285 00:16:02,200 --> 00:16:04,440 not doing anything. I'm like, well, yeah, But I 286 00:16:04,440 --> 00:16:06,560 think you might be confusing emotional contagion with 287 00:16:06,560 --> 00:16:08,880 empathy. But that doesn't mean all 288 00:16:08,880 --> 00:16:12,560 empathic demonstrations or all generations of empathy are hard. 289 00:16:12,560 --> 00:16:16,600 But on average it's it's harder than just observing with your 290 00:16:16,600 --> 00:16:19,400 own eyes. Yeah, maybe we can get really 291 00:16:19,400 --> 00:16:21,880 clear about what is and what isn't empathy then. 292 00:16:21,880 --> 00:16:24,880 So it's not emotional contagion. So I think emotional contagion 293 00:16:24,880 --> 00:16:28,280 is a building block of empathy. It's one of the ingredients of 294 00:16:28,280 --> 00:16:30,440 empathy. But another ingredient is 295 00:16:30,440 --> 00:16:33,280 knowing that those feelings you have, those aren't about you, 296 00:16:33,280 --> 00:16:35,840 they're about someone else. And then there's a separation 297 00:16:35,840 --> 00:16:37,560 between you and them. So they're cognitive element 298 00:16:37,560 --> 00:16:39,560 right away. So if you cry because someone 299 00:16:39,560 --> 00:16:41,960 else is crying, you're like just sharing and not crying. 300 00:16:41,960 --> 00:16:45,840 But it's an origin confusion. Empathy requires knowing what 301 00:16:45,840 --> 00:16:48,040 the source is and the separation between you and them. 302 00:16:48,640 --> 00:16:52,040 And typically scientists talk about 3 components of empathy. 303 00:16:53,200 --> 00:16:55,440 Sometimes they're useful distinctions, at least to 304 00:16:55,440 --> 00:16:57,680 describe. And on my own research, we show 305 00:16:57,680 --> 00:16:58,760 they're more similar than different. 306 00:16:58,760 --> 00:17:01,920 These three components we talk typically about a cognitive 307 00:17:01,920 --> 00:17:04,200 empathy. So perspective taking, 308 00:17:04,200 --> 00:17:08,079 understanding, or it's called mentalizing theory of mind, 309 00:17:08,359 --> 00:17:12,359 understanding what someone might be feeling or what they might be 310 00:17:12,359 --> 00:17:14,720 experiencing. So putting yourself in their 311 00:17:14,720 --> 00:17:17,160 shoes. All right, that's number one. 312 00:17:17,160 --> 00:17:19,800 Number two, emotional affective empathy. 313 00:17:20,000 --> 00:17:22,440 This is literally sharing in the emotion of someone else or 314 00:17:22,440 --> 00:17:25,319 sharing in their experience. So emotional contagion comes 315 00:17:25,319 --> 00:17:27,280 close to that. But again, you need to have that 316 00:17:27,280 --> 00:17:29,520 separation. But it's sharing in their 317 00:17:29,520 --> 00:17:31,520 feelings. Someone's sad, you're sad, 318 00:17:32,360 --> 00:17:33,960 someone's angry. You don't necessarily need to be 319 00:17:33,960 --> 00:17:37,080 angry, but you're also feeling is something that's concordant 320 00:17:37,080 --> 00:17:41,160 with their emotions. And then finally, there is some 321 00:17:41,160 --> 00:17:43,520 people call it the motivational part of empathy. 322 00:17:44,120 --> 00:17:45,640 It's called something called compassion. 323 00:17:45,920 --> 00:17:49,360 It's essentially generating feelings of kindness and warmth 324 00:17:49,600 --> 00:17:52,000 and positivity towards someone else. 325 00:17:53,120 --> 00:17:55,240 In theory at least, you can have one but not the other. 326 00:17:55,240 --> 00:17:58,200 But in our own data, when you have one component of empathy, 327 00:17:58,520 --> 00:18:00,800 72% of the time you have the other two components too. 328 00:18:00,840 --> 00:18:02,360 They seem to really travel together. 329 00:18:03,680 --> 00:18:06,480 And in our own examinations, we find that all of them are 330 00:18:06,480 --> 00:18:09,280 difficult, even compassion, even the motivational side. 331 00:18:09,760 --> 00:18:11,960 At the top of the show, he mentioned my colleague and good 332 00:18:11,960 --> 00:18:14,520 friend Paul Bloom. He has a book against empathy. 333 00:18:14,520 --> 00:18:17,440 But he then says, well, when I say against empathy, I mean 334 00:18:17,800 --> 00:18:20,760 these parts of empathy. He really means emotion sharing, 335 00:18:21,960 --> 00:18:24,800 but he says compassion is good. And my response to him, I mean, 336 00:18:24,800 --> 00:18:28,480 you can't say empathy is bad, but compassion is good because 337 00:18:28,600 --> 00:18:31,600 they're all together. They all are generated together. 338 00:18:31,800 --> 00:18:33,520 So either they're all bad or they're all good. 339 00:18:34,240 --> 00:18:36,520 As we're talking about definitions, one thing that 340 00:18:36,520 --> 00:18:39,560 usually comes up is this distinction between sympathy and 341 00:18:39,560 --> 00:18:42,040 empathy. Is that a useful separation? 342 00:18:42,480 --> 00:18:45,240 As far as I know, not. I've heard that sympathy word as 343 00:18:45,240 --> 00:18:47,120 well. I think it's maybe an older 344 00:18:47,120 --> 00:18:49,320 word. Maybe Adam Smith used sympathy? 345 00:18:49,320 --> 00:18:53,040 I'm not exactly sure. I don't see that word used among 346 00:18:53,240 --> 00:18:54,800 behavioral scientists at this point. 347 00:18:54,800 --> 00:18:59,120 Words and definitions for this construct are difficult, but 348 00:18:59,120 --> 00:19:01,040 sympathy is not one of the words we tend to use. 349 00:19:01,400 --> 00:19:03,120 So OK, get it back to the AI space. 350 00:19:03,320 --> 00:19:07,960 So where I got really interested in it is it took us a long time 351 00:19:07,960 --> 00:19:11,040 to convince people that empathy is effortful. 352 00:19:11,760 --> 00:19:13,560 And some of this day I get pushed back on it. 353 00:19:14,160 --> 00:19:16,600 But I think more or less the fuel is come to agree with us, 354 00:19:16,960 --> 00:19:20,160 the papers that we published on this now highly sighted, 355 00:19:20,160 --> 00:19:23,400 influential. And it's so influential in fact, 356 00:19:23,400 --> 00:19:27,920 that now people are using it to make the case against empathic 357 00:19:27,920 --> 00:19:31,760 AI, which I'm like, whoa, hold on there. 358 00:19:32,400 --> 00:19:36,560 So the first time I really got interested in empathic AI, other 359 00:19:36,560 --> 00:19:39,560 than me just being fascinated with like everyone has in 360 00:19:39,560 --> 00:19:43,160 November of 2023 and playing with it and testing it. 361 00:19:43,400 --> 00:19:45,720 There's a paper published by now, my good friend and not 362 00:19:45,720 --> 00:19:48,600 Perry, who's an Israeli psychologist and neuroscientist, 363 00:19:48,920 --> 00:19:51,760 where she essentially argued that AI will never capture the 364 00:19:51,840 --> 00:19:54,120 essence of empathy, of human empathy. 365 00:19:54,120 --> 00:19:56,880 So she didn't say more or less, we can forget about it, but 366 00:19:56,880 --> 00:19:58,480 she's like, it's never going to be empathic. 367 00:19:58,560 --> 00:20:02,160 And one of the things that she said was because empathy's 368 00:20:02,160 --> 00:20:06,560 effortful. And machines don't exert effort 369 00:20:06,640 --> 00:20:09,200 it it could keep on generating empathy forever and ever as long 370 00:20:09,240 --> 00:20:12,040 as it has, well, actual energy being plugged in the wall. 371 00:20:12,040 --> 00:20:14,480 Yeah, there's an unlimited supply of empathy. 372 00:20:14,680 --> 00:20:17,040 Yes. And you're saying the limits of 373 00:20:17,040 --> 00:20:20,560 human empathy is actually the hallmark of empathy to begin 374 00:20:20,560 --> 00:20:22,880 with. If it was unlimited, it would no 375 00:20:22,880 --> 00:20:26,200 longer be human. But I remember just being like 376 00:20:26,280 --> 00:20:29,640 miffed at that argument because like, whoa, we took forever to 377 00:20:29,640 --> 00:20:31,400 convince people that empathy never felt. 378 00:20:31,400 --> 00:20:34,680 Now you're saying the main thing that defines empathy in humans 379 00:20:34,680 --> 00:20:38,160 like, well, that's maybe a bit too far now, But what I think 380 00:20:38,160 --> 00:20:43,240 I'm not meant was we know that someone cares for us, that 381 00:20:43,240 --> 00:20:46,880 they're truly empathizing with us because we know they're 382 00:20:46,880 --> 00:20:49,440 willing to pay the the cost of empathy. 383 00:20:49,640 --> 00:20:52,280 They're willing to exert the effort for us and therefore we 384 00:20:52,280 --> 00:20:54,280 feel special. And the machine won't 385 00:20:54,280 --> 00:20:56,080 necessarily make you feel special because because it 386 00:20:56,080 --> 00:20:59,040 doesn't pay that same cost. I think that second part is 387 00:20:59,040 --> 00:21:00,280 wrong. I think a machine could 388 00:21:00,280 --> 00:21:03,320 definitely fool you into thinking it's selective and in 389 00:21:03,320 --> 00:21:05,160 fact could be programmed to be selective. 390 00:21:05,200 --> 00:21:06,680 I might not empathize with someone like me. 391 00:21:06,680 --> 00:21:08,880 It doesn't resonate with my personality but resonates with 392 00:21:08,880 --> 00:21:10,600 other people. So it could also be selective 393 00:21:10,880 --> 00:21:12,760 and choose you and then you could feel special too. 394 00:21:13,960 --> 00:21:18,880 But even if it doesn't right, there's the layer of our people 395 00:21:18,880 --> 00:21:21,560 believing that empathy is being conveyed. 396 00:21:22,640 --> 00:21:24,640 You're right, absolutely right. So the secondary issue, which is 397 00:21:24,640 --> 00:21:27,400 OK, who cares if it doesn't actually have real empathy? 398 00:21:27,760 --> 00:21:32,120 Is it felt as real empathy? Is it perceived as real empathy? 399 00:21:32,400 --> 00:21:35,760 We saw a lot more testing to do, but I think so far the answers 400 00:21:35,800 --> 00:21:39,760 are resounding. Yes, like people perceive the 401 00:21:39,800 --> 00:21:44,080 utterances of AI, whether it be text or now voice, they perceive 402 00:21:44,080 --> 00:21:48,200 that as empathic. They perceive it as caring and 403 00:21:48,200 --> 00:21:52,240 in fact they feel cared for, which is mind blowing to feel 404 00:21:52,240 --> 00:21:55,400 cared for by a fundamentally uncaring thing. 405 00:21:55,560 --> 00:21:57,440 But in fact, we know that people do. 406 00:21:58,120 --> 00:22:00,160 Yeah, we have. We've already seen it over and 407 00:22:00,160 --> 00:22:02,000 over again. If we can, at least for now, 408 00:22:02,000 --> 00:22:06,440 accept the conclusion that people are either fooled or, you 409 00:22:06,440 --> 00:22:09,800 know, wanting to believe that something is caring for them, 410 00:22:09,880 --> 00:22:13,280 maybe if they don't get enough of that limited supply in their 411 00:22:13,400 --> 00:22:16,080 real lives. And so I think maybe the next 412 00:22:16,080 --> 00:22:19,000 question then is, does it matter? 413 00:22:19,920 --> 00:22:22,040 Do we care at all that it's real or not real? 414 00:22:22,520 --> 00:22:25,200 Yeah, I'd have this endless debate with Paul Bloom about 415 00:22:25,200 --> 00:22:27,160 this on his sub stack small potatoes, which I highly 416 00:22:27,160 --> 00:22:29,360 recommend. He wrote some stuff about AI and 417 00:22:29,440 --> 00:22:32,280 Patrick AI and then he allowed us to write something on a sub 418 00:22:32,280 --> 00:22:35,440 stack and then he responded. If you ask me, like me, the 419 00:22:35,440 --> 00:22:37,840 person playing devil's advocate versus me, the actual human 420 00:22:37,840 --> 00:22:41,240 being, a 52 year old man who's lived in this world well before 421 00:22:41,240 --> 00:22:43,080 AI, you'll get different answers. 422 00:22:43,240 --> 00:22:45,720 So if you want to talk with the real person, yeah, I think it 423 00:22:45,720 --> 00:22:47,960 matters. Like I'd rather talk to you than 424 00:22:47,960 --> 00:22:49,920 ChatGPT. I'd rather have a beer with Paul 425 00:22:49,920 --> 00:22:52,800 or UL in bar. My friend and Co host and 426 00:22:52,800 --> 00:22:54,320 colleague at the University of Toronto. 427 00:22:54,840 --> 00:22:58,840 I'd rather talk to him, a real person, then talk to a chat GP 428 00:22:58,840 --> 00:23:04,200 that even knows me. But I'm old, never mind that I'm 429 00:23:04,200 --> 00:23:06,080 not an AI native, I'm not a digital native. 430 00:23:06,080 --> 00:23:07,720 I grew up with like a Rotary phone. 431 00:23:07,720 --> 00:23:09,720 I grew up with pre Internet days. 432 00:23:10,000 --> 00:23:12,920 But now imagine you're an AI native. 433 00:23:12,960 --> 00:23:14,760 Maybe you're born even in like 2010. 434 00:23:15,360 --> 00:23:17,880 All right, AI wasn't around when you were born, but you're just 435 00:23:17,880 --> 00:23:19,840 coming of age and using machines now. 436 00:23:19,840 --> 00:23:21,920 Like I have a daughter is, you know, born in 2011. 437 00:23:23,400 --> 00:23:26,200 She's had friends before AI, but now there could be a possibility 438 00:23:26,200 --> 00:23:28,280 of her having friends with AI and it's not so weird. 439 00:23:28,960 --> 00:23:33,040 Maybe she won't think it's so special, the reality be special 440 00:23:33,040 --> 00:23:37,400 that they actually care for you, actually care for you and are 441 00:23:37,400 --> 00:23:38,960 actually real. Maybe she won't think that's so 442 00:23:38,960 --> 00:23:40,800 special and so unique. I think that would be a loss. 443 00:23:40,800 --> 00:23:43,560 I do think there is something important to reality. 444 00:23:43,560 --> 00:23:46,840 So the conclusion of this conversation I had with Paul was 445 00:23:48,200 --> 00:23:50,840 that reality might not matter as much as we might imagine that. 446 00:23:50,840 --> 00:23:54,280 I think a lot of times people will willingly escape reality. 447 00:23:54,680 --> 00:23:56,880 I mean, we take drugs to escape reality. 448 00:23:56,880 --> 00:24:01,320 We do things to enhance our realities regularly, including 449 00:24:01,320 --> 00:24:04,280 us now talking over Zoom. That's an enhancement of 450 00:24:04,280 --> 00:24:08,760 reality. So I'm not exactly sure how much 451 00:24:08,760 --> 00:24:12,760 reality is important to us, but I think for evolved reasons it 452 00:24:12,760 --> 00:24:15,440 makes sense for us to care about reality. 453 00:24:15,440 --> 00:24:18,240 It's important that there is consensus reality because I 454 00:24:18,240 --> 00:24:20,680 think there could be catastrophic mistakes if we are 455 00:24:20,680 --> 00:24:23,040 loosed, are unmoored from reality. 456 00:24:23,360 --> 00:24:24,560 So then that's the actual answer. 457 00:24:24,560 --> 00:24:28,280 Reality is important because survival depends on having 458 00:24:28,480 --> 00:24:31,080 societies of people having the same reality. 459 00:24:31,600 --> 00:24:34,760 And if we each go off and separate realities, things will 460 00:24:34,760 --> 00:24:38,960 have coordination problems. But this is like the big reason 461 00:24:38,960 --> 00:24:41,680 why reality is is important, but small reason. 462 00:24:41,680 --> 00:24:46,240 Like, so I've got a friend who's not real and I confide in her 463 00:24:46,600 --> 00:24:49,440 and she understands me and she's there for me in the middle of 464 00:24:49,440 --> 00:24:53,240 the night when I'm worried or scared and she's not real. 465 00:24:53,240 --> 00:24:55,440 Does that matter? I mean, if you're comforted by 466 00:24:55,440 --> 00:24:57,200 that. My toddler does that all the 467 00:24:57,200 --> 00:24:58,920 time. Totally normal. 468 00:25:00,600 --> 00:25:03,000 That's right. For young children or babies, 469 00:25:03,320 --> 00:25:05,800 they've got made-up friends or imaginary friends or they've got 470 00:25:05,800 --> 00:25:07,440 stuffed animals that they think are real. 471 00:25:07,440 --> 00:25:10,320 And we don't think that's unhealthy or pathological. 472 00:25:10,560 --> 00:25:13,200 We think that's totally fine. They get comfort from it. 473 00:25:14,720 --> 00:25:17,480 What I'm getting from this is you're talking less about the 474 00:25:17,480 --> 00:25:22,120 experience of of empathy or this shared experience and more about 475 00:25:22,120 --> 00:25:25,480 the utility of empathy. So can we talk a little bit 476 00:25:25,480 --> 00:25:30,960 about what good does empathy do in terms of this utility 477 00:25:30,960 --> 00:25:33,560 perspective? What benefit do people get from 478 00:25:33,560 --> 00:25:37,600 feeling that there is something or someone empathizing for them? 479 00:25:38,000 --> 00:25:39,680 Yeah. Well, first of all, I should say 480 00:25:39,720 --> 00:25:43,280 that is a really good question. It's a question that we have 481 00:25:43,560 --> 00:25:48,400 surprisingly few answers to 'cause you would think it's 482 00:25:48,400 --> 00:25:50,680 obvious what the benefits are, and maybe some of them are 483 00:25:50,680 --> 00:25:53,960 obvious. But the field of empathy in 484 00:25:53,960 --> 00:25:57,160 psychology has been so concerned with the generator of empathy, 485 00:25:57,160 --> 00:25:59,800 the person generating the empathy, we actually haven't 486 00:25:59,800 --> 00:26:02,800 bothered asking the question of does it help the person? 487 00:26:03,000 --> 00:26:05,560 How does it help them? What form helps them? 488 00:26:06,320 --> 00:26:11,200 What's interesting is if this turn to empathic AI that has LED 489 00:26:11,200 --> 00:26:15,160 us to now look at, hey, empathy is at least a dyadic process, 490 00:26:15,280 --> 00:26:18,200 shouldn't we look at the dyad, the other member of the dyad? 491 00:26:18,720 --> 00:26:21,400 And it's this turn to AI that's done this because we're like, 492 00:26:21,400 --> 00:26:25,520 hey, like the generator is pretty good. 493 00:26:25,920 --> 00:26:29,040 Like how does it feel and why is it important? 494 00:26:29,280 --> 00:26:31,240 So we don't have to look to research and empathy to 495 00:26:31,240 --> 00:26:33,040 understand this. We look at relationship, you 496 00:26:33,040 --> 00:26:34,080 know, the study of relationships. 497 00:26:34,560 --> 00:26:37,800 We know that when people feel socially supported, their 498 00:26:37,800 --> 00:26:40,200 well-being is better, when they're when they feel 499 00:26:40,200 --> 00:26:43,680 understood, when they feel like their humanities recognized, 500 00:26:44,160 --> 00:26:46,320 they feel better. But we still don't know as much 501 00:26:46,320 --> 00:26:47,680 as we could. So for example, we were 502 00:26:47,680 --> 00:26:49,600 describing the three faces of empathy. 503 00:26:50,280 --> 00:26:53,280 And I think if people were put to a test, if there's one 504 00:26:53,280 --> 00:26:55,000 component of empathy, which one is it? 505 00:26:55,280 --> 00:26:58,160 I suspect people might say something like emotion sharing, 506 00:26:58,520 --> 00:27:00,800 like that emotion contagion thing that we talked at the top 507 00:27:00,800 --> 00:27:03,960 of the show. But is it possible that emotion 508 00:27:03,960 --> 00:27:07,200 sharing is actually not important or not as important as 509 00:27:07,200 --> 00:27:09,920 the other two components in terms of making people feel 510 00:27:09,920 --> 00:27:13,560 heard, validated, and cared for? And actually, I've got research 511 00:27:13,560 --> 00:27:15,800 on this right now. We're testing this at the moment 512 00:27:16,120 --> 00:27:18,640 with AI. We don't care about the AI part 513 00:27:18,640 --> 00:27:22,040 so much as just a way of standardizing the generation of 514 00:27:22,040 --> 00:27:26,120 empathy in a conversation. And what we find is we create a 515 00:27:26,120 --> 00:27:29,160 is to be 1 flavor of empathy, but not the other two. 516 00:27:29,400 --> 00:27:32,120 It's actually very hard to do because they're so related. 517 00:27:32,960 --> 00:27:34,960 And I think it'll be practically impossible for humans to do. 518 00:27:35,000 --> 00:27:37,560 But an AI can do this. It's not easy, but we can do it. 519 00:27:37,560 --> 00:27:41,800 And what we find actually is that emotion sharing is the 520 00:27:41,800 --> 00:27:45,280 least important facet of empathy, right? 521 00:27:45,280 --> 00:27:49,680 So in terms of feeling cared for, heard, validated, someone's 522 00:27:49,680 --> 00:27:51,480 there, you know, looking out for you. 523 00:27:51,960 --> 00:27:53,680 Emotion sharing is the least important one. 524 00:27:54,080 --> 00:27:55,520 It's not really about the other person. 525 00:27:56,600 --> 00:27:57,960 Exactly. It's not about the other person, 526 00:27:57,960 --> 00:27:59,840 it's about you resonating with them. 527 00:27:59,840 --> 00:28:02,520 And also it could be distressing to you, the observer, like, Oh 528 00:28:02,520 --> 00:28:04,400 my God, you're scared and I'm freaking out. 529 00:28:04,400 --> 00:28:05,720 Now you have to take care of them. 530 00:28:05,800 --> 00:28:07,680 So that's not good, but perspective taking and 531 00:28:07,680 --> 00:28:11,080 compassion, those seem to be on equal footing and they seem to 532 00:28:11,080 --> 00:28:14,200 be as good in generating feelings of warmth, that 533 00:28:14,200 --> 00:28:16,120 feeling, heard about it, etcetera. 534 00:28:16,520 --> 00:28:18,680 So in theory, that's what everything is good for, is 535 00:28:18,680 --> 00:28:22,000 making someone feel like they're not alone, they're understood. 536 00:28:22,320 --> 00:28:26,040 But if a machine can do it, and it can do it and make us feel 537 00:28:26,040 --> 00:28:28,880 this way because we readily anthropomorphize, there's also 538 00:28:28,880 --> 00:28:30,400 other benefits that a machine has. 539 00:28:30,520 --> 00:28:32,040 This is one that I find fascinating. 540 00:28:32,480 --> 00:28:38,320 We are more willing to disclose personal, private information to 541 00:28:38,320 --> 00:28:42,080 a bot than to a human. And we know that self disclosure 542 00:28:42,080 --> 00:28:44,320 is positive. We know that unloading some of 543 00:28:44,320 --> 00:28:47,240 what we're feeling is positive. There's no one there to judge 544 00:28:47,240 --> 00:28:49,760 you on the other side. There's no one there to judge 545 00:28:49,760 --> 00:28:51,520 you. So, you know, I've tell this 546 00:28:51,520 --> 00:28:53,520 joke occasionally. It's true. 547 00:28:53,520 --> 00:28:57,080 It's funny, but it's true. You know, I'm a human being that 548 00:28:57,080 --> 00:28:59,480 lives in the world and I'm not perfectly virtuous. 549 00:28:59,840 --> 00:29:02,360 I do things that are maybe not so good for me that I don't want 550 00:29:02,360 --> 00:29:06,280 to admit to people. So when my doctor asked me about 551 00:29:06,280 --> 00:29:09,680 my drug consumption before cannabis was legalized in 552 00:29:09,680 --> 00:29:12,720 Canada, I would lie. Even now I might not admit it, 553 00:29:12,720 --> 00:29:15,960 but I've come out as a cannabis advocate, so I'm more willing to 554 00:29:15,960 --> 00:29:18,440 speak about it. Sit on your CV. 555 00:29:18,760 --> 00:29:21,680 Yeah, yeah, well, I actually have a paper now on studying 556 00:29:21,680 --> 00:29:23,080 cannabis. We're actually getting people 557 00:29:23,080 --> 00:29:26,880 high, possibly as we speak, people are getting high on my 558 00:29:26,880 --> 00:29:30,560 behest in a lab somewhere in Scarborough, ON. 559 00:29:32,080 --> 00:29:34,200 Yeah. So I wouldn't admit that to A to 560 00:29:34,200 --> 00:29:36,960 my doctor, but I definitely would admit it to a bot. 561 00:29:36,960 --> 00:29:39,120 They're not going to judge me. Yeah, yeah. 562 00:29:39,120 --> 00:29:41,000 I think what the interesting thing is when we talk a lot 563 00:29:41,000 --> 00:29:44,280 about AI is that we forget about the counterfactual in terms of 564 00:29:44,280 --> 00:29:45,600 what would a human do in this scenario. 565 00:29:45,600 --> 00:29:48,160 I think I mentioned this before, I'm a really big fan of the show 566 00:29:48,160 --> 00:29:51,200 Curb the Enthusiasm. And I think Larry David is an 567 00:29:51,200 --> 00:29:53,280 interesting character of basically having very low 568 00:29:53,280 --> 00:29:56,120 empathy, I would say, but like trying to figure out what does 569 00:29:56,120 --> 00:29:59,440 he have to do in order to tick the box of enough perceived 570 00:29:59,440 --> 00:30:02,120 empathy. There's a episode about like he 571 00:30:02,120 --> 00:30:04,600 makes a failed apology because he's snacking on pistachio notes 572 00:30:04,600 --> 00:30:06,680 while he's doing the apology. That's a bad apology. 573 00:30:07,080 --> 00:30:11,280 He doesn't say my condolences to someone who's lost their parent 574 00:30:11,280 --> 00:30:13,640 like 2 years ago because they're like, that's a little bit too 575 00:30:13,640 --> 00:30:17,360 long of a period we have to say. And I think it's kind of 576 00:30:17,360 --> 00:30:20,560 interesting in that sense because at least that show tries 577 00:30:20,560 --> 00:30:22,880 to play on this fact that I do really care about people 578 00:30:23,000 --> 00:30:26,480 expressing quote UN quote will empathy or just playing the game 579 00:30:26,480 --> 00:30:29,080 enough so that we feel like they take that box. 580 00:30:29,880 --> 00:30:32,240 No, I, I, I think perception is all we can base it on. 581 00:30:32,240 --> 00:30:36,120 Like I, I can only base my feelings about your feelings 582 00:30:36,240 --> 00:30:39,000 based on what I observe. So if you are going to the 583 00:30:39,000 --> 00:30:41,560 motions, you're saying the right words, you have the right facial 584 00:30:41,560 --> 00:30:45,080 configuration and have the right actions, I'm going to assume 585 00:30:45,080 --> 00:30:47,560 you're being empathic, but you could be faking it. 586 00:30:47,920 --> 00:30:49,800 And in fact, we know that people fake it. 587 00:30:51,360 --> 00:30:53,280 Sometimes they fake it because it's their job. 588 00:30:53,720 --> 00:30:57,160 So if you're a therapist and it's the end of the day, so my 589 00:30:57,160 --> 00:31:00,120 wife is a therapist and I tell you the end of the day and you 590 00:31:00,120 --> 00:31:02,480 know, kids are home, we've got to get dinner on the table 591 00:31:02,480 --> 00:31:04,560 shortly. That last client might not be 592 00:31:04,560 --> 00:31:08,560 getting like 100% of my wife's empathy and my wife might be 593 00:31:08,560 --> 00:31:10,360 going through the emotions a little bit, especially if she's 594 00:31:10,720 --> 00:31:13,320 been with his client for years and heard some of the stories 595 00:31:13,480 --> 00:31:15,800 countless numbers of times. And that's not, I'm not trying 596 00:31:15,800 --> 00:31:17,040 to say anything pour up on my wife. 597 00:31:17,040 --> 00:31:19,280 She's an excellent therapist, but I think all therapist face 598 00:31:19,280 --> 00:31:22,160 this. And what I see from critics of 599 00:31:22,160 --> 00:31:26,360 AII see this repeatedly, especially from certain critics, 600 00:31:26,800 --> 00:31:30,240 is that they romanticize human empathy. 601 00:31:30,520 --> 00:31:35,560 They're trying to take down empathic AI, but they do that by 602 00:31:35,560 --> 00:31:39,720 glorifying and romanticize. Look at this utopic vision of 603 00:31:39,720 --> 00:31:41,840 human empathy, not realizing that a lot of the empathy we 604 00:31:41,840 --> 00:31:45,520 experience in everyday life is not that great either. 605 00:31:45,600 --> 00:31:49,640 This genuine, real, heartfelt, incredible empathy that's not 606 00:31:49,640 --> 00:31:53,760 true all the time for everybody. I guess to your point earlier, 607 00:31:53,760 --> 00:31:56,880 with the research that you're currently conducting, maybe it 608 00:31:56,880 --> 00:32:00,560 doesn't matter so much, right, If the important elements are 609 00:32:00,560 --> 00:32:04,600 not that shared feeling, but in fact the description of 610 00:32:04,720 --> 00:32:08,600 perspective taking or, you know, the other parts of empathy that 611 00:32:08,600 --> 00:32:13,000 something like an AI can convey quite well. 612 00:32:14,200 --> 00:32:16,840 You're right, I mean it maybe it shouldn't because it's not real 613 00:32:16,840 --> 00:32:19,200 and there's no real person there caring for you people to have a 614 00:32:19,320 --> 00:32:21,840 real issue with this. And I think you can mount really 615 00:32:21,840 --> 00:32:25,120 good ethical arguments against empathic AI. 616 00:32:25,520 --> 00:32:29,400 So one argument that you can make is because of our 617 00:32:29,400 --> 00:32:33,320 psychology, because we were so readily anthropomorphize, even 618 00:32:33,320 --> 00:32:37,680 if we give consent to I'm now, you know, talking to an AI right 619 00:32:37,680 --> 00:32:41,960 now, I consent, I consent. You get trapped in the illusion, 620 00:32:42,480 --> 00:32:46,640 you're caught in the fantasy and you forget. 621 00:32:47,120 --> 00:32:48,800 I think that's not giving people enough credit. 622 00:32:49,320 --> 00:32:51,680 I think people don't forget. They do get caught in the 623 00:32:51,680 --> 00:32:53,600 illusion, just like we get caught in the illusion where 624 00:32:53,600 --> 00:32:55,200 we're reading a book or watching ATV show. 625 00:32:55,200 --> 00:32:57,720 We're kind of rooting for characters and we feel like we 626 00:32:57,720 --> 00:32:58,960 empathize with the characters as well. 627 00:32:59,600 --> 00:33:04,440 I don't think that's trap, but it's sticky and you might end up 628 00:33:04,640 --> 00:33:09,200 like it could potentially put you in some places ethically 629 00:33:09,200 --> 00:33:11,120 that are questionable. So I understand the ethical 630 00:33:11,120 --> 00:33:14,640 dilemmas here, but the one thing I don't want to have happen is 631 00:33:14,960 --> 00:33:20,000 for people to reflexively reject empathic AI for certain reasons 632 00:33:20,200 --> 00:33:23,560 and then not see the utility of it. 633 00:33:23,840 --> 00:33:27,480 Here's an important place where this utility right now. 634 00:33:27,680 --> 00:33:30,080 I just saw a little abstract this morning about this. 635 00:33:31,160 --> 00:33:35,840 There's far more need for therapy and friendship and 636 00:33:35,840 --> 00:33:40,040 companionship than there are people to give these sorts of 637 00:33:40,040 --> 00:33:42,200 empathy. I think it's hard to make that 638 00:33:42,200 --> 00:33:45,560 case for friendship. I think the case for therapy is 639 00:33:45,560 --> 00:33:47,280 often made and convincingly made. 640 00:33:47,280 --> 00:33:50,320 But in terms of friendship, I think there are enough people to 641 00:33:50,320 --> 00:33:53,480 have friends with. How does that argument bear out? 642 00:33:54,120 --> 00:33:55,520 OK, I'm going to make you really sad. 643 00:33:55,520 --> 00:34:00,400 I mean, so according to I'm forgetting which Census Bureau 644 00:34:00,400 --> 00:34:06,040 it is now, but somewhere in the US in twenty, 2112% of adults do 645 00:34:06,040 --> 00:34:09,320 not have a single friend. That number is higher for men. 646 00:34:09,320 --> 00:34:13,719 I think it's closer to 15%. OK, but if you go to men who are 647 00:34:13,719 --> 00:34:16,040 over 40, the number gets terrible. 648 00:34:16,040 --> 00:34:18,920 Men just don't keep friends above a certain age. 649 00:34:19,120 --> 00:34:22,000 So you're telling me that just try harder, make friends? 650 00:34:22,000 --> 00:34:24,560 Well, it it's easier said than done. 651 00:34:24,840 --> 00:34:27,840 So now if I give you something, a resource that can help you, 652 00:34:28,400 --> 00:34:30,199 there are cons too. Don't get me wrong, there are 653 00:34:30,199 --> 00:34:32,760 definitely cons, but you're friendless. 654 00:34:32,880 --> 00:34:36,400 You've tried. You don't leave your house other 655 00:34:36,400 --> 00:34:38,920 than maybe go to work or maybe you don't even do that now 656 00:34:38,920 --> 00:34:40,360 because it's all you know, online. 657 00:34:40,800 --> 00:34:44,800 You're telling me you know an AI companion wouldn't be a positive 658 00:34:44,800 --> 00:34:47,639 for you? I have seen some short term 659 00:34:47,639 --> 00:34:52,639 benefits to something like an AI companion and I do believe there 660 00:34:52,639 --> 00:34:56,320 is a time and place. I do not believe that this is a 661 00:34:56,320 --> 00:34:59,040 general solution for all of humanity. 662 00:34:59,120 --> 00:35:04,480 And and here's why I think that there are some outsized risks to 663 00:35:05,000 --> 00:35:08,240 forming those types of relationships that kind of pull 664 00:35:08,240 --> 00:35:11,000 you deeper and deeper into the machine world. 665 00:35:11,280 --> 00:35:17,560 And I think that potentially a sort of compromise between these 666 00:35:17,560 --> 00:35:20,400 solutions of finding a real friend out in the real world, 667 00:35:20,400 --> 00:35:24,080 which to be fair, like I don't think that's impossible for the 668 00:35:25,040 --> 00:35:28,080 12% of people without a single friend. 669 00:35:28,320 --> 00:35:31,080 It is hard to make friends, especially as an adult, 670 00:35:31,080 --> 00:35:35,480 especially as an older adult. Not that 40 is old, but there 671 00:35:35,480 --> 00:35:39,520 are ways to help as well. And the solution of an AI 672 00:35:39,520 --> 00:35:43,240 companion is not the only tool that we can create, right? 673 00:35:43,240 --> 00:35:45,880 Like there are other technological solutions even 674 00:35:46,000 --> 00:35:48,760 that aren't substituting that relationship. 675 00:35:49,440 --> 00:35:54,800 OK, So imagine though that you do go the AI companion route and 676 00:35:54,800 --> 00:35:58,120 you say, OK, here, you know, here's your AI companion. 677 00:35:59,120 --> 00:36:02,480 I still think we have to have guardrails that prevent from 678 00:36:02,480 --> 00:36:06,560 that becoming the only thing that the companion can, you 679 00:36:06,560 --> 00:36:10,360 know, in essence, break up with you as a friend and say, hey, 680 00:36:10,360 --> 00:36:12,560 like, this has been fun and I'm glad that you've been feeling 681 00:36:12,560 --> 00:36:15,920 better over the last couple of weeks, you know, that help you 682 00:36:16,160 --> 00:36:22,480 maybe build the resources or the skills to be able to go out and 683 00:36:22,480 --> 00:36:24,240 make some real friends. I think that is really 684 00:36:24,240 --> 00:36:27,280 important. Yeah, I don't disagree with you. 685 00:36:27,280 --> 00:36:30,920 I think that there are real dangers with over reliance and 686 00:36:31,000 --> 00:36:34,440 AI companions might exacerbate the very problem we're trying to 687 00:36:34,960 --> 00:36:37,040 cure. Let's say a lack of friendship. 688 00:36:37,680 --> 00:36:41,480 If you take the long term view, potentially I can see that. 689 00:36:42,200 --> 00:36:45,000 Yeah, but we still don't know that they're just not enough 690 00:36:45,000 --> 00:36:47,840 good data out there right now. What we know is that AI 691 00:36:47,840 --> 00:36:49,600 companions need to be good in the short term. 692 00:36:50,840 --> 00:36:55,160 And we know that people who use technology a lot or people who 693 00:36:55,160 --> 00:37:00,920 use AIA lot tend to be lonelier. But we don't know that AI 694 00:37:01,080 --> 00:37:04,240 companion use increases loneliness or decreases 695 00:37:04,240 --> 00:37:07,120 loneliness. All we know is short term it 696 00:37:07,120 --> 00:37:10,720 seems to be OK, but junk food could be OK, you know, if you're 697 00:37:10,720 --> 00:37:13,280 starving. But we won't want someone to 698 00:37:13,280 --> 00:37:16,920 subsist on junk food. But there are going to be people 699 00:37:16,920 --> 00:37:19,720 who are never going to have the resources to find food. 700 00:37:20,520 --> 00:37:22,080 They're never going to be able to hunt on their own. 701 00:37:22,080 --> 00:37:23,960 They're never going to have a balanced diet on their own. 702 00:37:24,200 --> 00:37:26,640 And I'd rather them have chips the rest of their lives than 703 00:37:26,640 --> 00:37:29,040 starve. And I suspect that's going to be 704 00:37:29,040 --> 00:37:31,520 the case with real humans and friendship, too. 705 00:37:31,520 --> 00:37:35,320 They're it's unfortunate. I think we need massive societal 706 00:37:35,320 --> 00:37:41,120 change, you know, before machines could be thrown away as 707 00:37:41,120 --> 00:37:45,520 possible companions. Society seems to be going more 708 00:37:45,520 --> 00:37:49,280 toward the level of atomization. Look, 10 years ago we would have 709 00:37:49,280 --> 00:37:52,520 been meeting maybe in some university building to have this 710 00:37:52,520 --> 00:37:54,960 podcast. Now we're doing it online, which 711 00:37:54,960 --> 00:37:58,080 is great because we can connect, but it means that after this 712 00:37:58,080 --> 00:38:00,280 conversation, I won't have the pleasure of hanging out with 713 00:38:00,280 --> 00:38:01,160 you. You won't have the pleasure of 714 00:38:01,160 --> 00:38:03,920 hanging out with me. We're doing more and more stuff 715 00:38:03,920 --> 00:38:06,160 alone. And that is a trend that's true 716 00:38:06,160 --> 00:38:08,080 for eating alone. Apparently. 717 00:38:08,080 --> 00:38:10,400 There's a beautiful article written by Derek Thompson in The 718 00:38:10,400 --> 00:38:14,560 Atlantic talking about the increased like individuation of 719 00:38:14,680 --> 00:38:16,120 everything we do, including eating. 720 00:38:16,440 --> 00:38:19,000 Like most restaurant orders now are take out. 721 00:38:19,000 --> 00:38:21,200 They're not even like sitting down in the restaurant anymore. 722 00:38:21,880 --> 00:38:23,360 Yeah, he said. Like, forget about the 723 00:38:23,640 --> 00:38:26,200 loneliness epidemic. It's just like being alone. 724 00:38:26,960 --> 00:38:28,360 We're just doing everything alone now. 725 00:38:29,040 --> 00:38:30,320 That's right. It's interesting. 726 00:38:30,320 --> 00:38:34,080 We have been in this, I think for the last couple years this 727 00:38:34,080 --> 00:38:37,240 kind of trying to make sense of the impacts of social media and 728 00:38:37,240 --> 00:38:40,200 we've have various voices saying, you know, that it's 729 00:38:40,200 --> 00:38:41,720 really harmful. Some people say it's less 730 00:38:41,720 --> 00:38:44,200 harmful, but that's been like the debate in many ways that I 731 00:38:44,200 --> 00:38:47,400 think has preoccupied a lot of our kind of feelings towards 732 00:38:47,400 --> 00:38:49,880 technology, especially with kids and so on. 733 00:38:49,880 --> 00:38:52,440 Growing up with more and more social media, that's been a very 734 00:38:52,440 --> 00:38:57,200 big theme and thinking ahead. I think for me, I think a lot 735 00:38:57,200 --> 00:39:00,880 about like, OK, if that's been something we had to form of like 736 00:39:01,680 --> 00:39:06,520 moral outrage or concerns about, it's in some ways pales in 737 00:39:06,520 --> 00:39:09,720 comparisons to the idea. Like I think not too far distant 738 00:39:09,720 --> 00:39:14,800 future we will have really good social companions that are AI 739 00:39:14,800 --> 00:39:18,120 generated that already the ones exist around are really good. 740 00:39:18,120 --> 00:39:20,880 Like I've been impressed with any conversation I've had with 741 00:39:20,960 --> 00:39:24,480 the leading either text based or voice based companions. 742 00:39:24,480 --> 00:39:27,440 But like in a few years they will be even better and everyone 743 00:39:27,480 --> 00:39:30,160 with access to them. And it wouldn't be super 744 00:39:30,160 --> 00:39:33,240 surprising if it's like a little bit similar to, you know, the 745 00:39:33,240 --> 00:39:36,000 early days of whatever Facebook or Instagram, wherever we're 746 00:39:36,000 --> 00:39:38,120 used. Like it's reached a lot of 747 00:39:38,120 --> 00:39:40,320 masses relatively quickly. And people used them a lot 748 00:39:40,320 --> 00:39:42,320 without really thinking about, OK, what are the consequences 749 00:39:42,320 --> 00:39:44,160 and what are the ethical implications. 750 00:39:44,520 --> 00:39:47,480 And suddenly you have obviously a lot of these being privatized, 751 00:39:47,960 --> 00:39:50,200 a lot of data being shared with like private companies. 752 00:39:50,200 --> 00:39:52,360 And they were trying to make money on this in some ways, 753 00:39:52,360 --> 00:39:55,400 maybe with ads, with the selling the data, with using data in 754 00:39:55,400 --> 00:39:57,480 various ways. Like that seems like a much 755 00:39:57,480 --> 00:40:01,280 more, I don't know, complex and scary thing then whatever we 756 00:40:01,280 --> 00:40:05,120 have right now with potential social media being bad in 757 00:40:05,120 --> 00:40:07,240 whatever aspect. I'm still a bit skeptical to the 758 00:40:07,240 --> 00:40:08,520 degree that social media is harmful. 759 00:40:09,120 --> 00:40:10,440 But like, how do you think about that? 760 00:40:10,440 --> 00:40:13,760 Like if you're trying to look at this really hard to peer through 761 00:40:14,160 --> 00:40:16,640 crystal ball of the future where some of these things can play 762 00:40:16,640 --> 00:40:18,280 out, Like how do you think about that aspect? 763 00:40:19,120 --> 00:40:22,800 I I think there are lots of moving parts and lots of 764 00:40:22,800 --> 00:40:24,720 potential risks and dangers ahead. 765 00:40:25,920 --> 00:40:31,200 I agree that the AI companies, their data are being used and 766 00:40:31,200 --> 00:40:35,920 harvested and their privacy concerns, they're also 767 00:40:35,920 --> 00:40:41,240 manipulation concerns. So if more and more of our 768 00:40:41,240 --> 00:40:46,160 friends or even if just a small proportion of our friends are AI 769 00:40:46,160 --> 00:40:51,520 based and these are owned by profit driven companies who not 770 00:40:51,520 --> 00:40:54,880 only want to make profit more and more looks like some actors 771 00:40:55,400 --> 00:40:59,840 have political agendas too. They could potentially use the 772 00:40:59,840 --> 00:41:04,200 bots to suddenly persuade people in a political manner. 773 00:41:04,600 --> 00:41:06,400 And that is incredibly dangerous. 774 00:41:06,400 --> 00:41:09,840 So now imagine like Aceo who had a turn of a dial could just like 775 00:41:09,840 --> 00:41:15,040 suddenly change like, you know, a favorability ratings for one 776 00:41:15,040 --> 00:41:17,000 candidate over another. It doesn't matter which 777 00:41:17,000 --> 00:41:20,280 candidate it is. That's just a really powerful 778 00:41:20,280 --> 00:41:22,400 tool. So I think there we have to be 779 00:41:22,400 --> 00:41:24,120 thinking carefully about this stuff. 780 00:41:25,000 --> 00:41:29,120 I think there need to be regulations to make sure there 781 00:41:29,120 --> 00:41:34,280 are certain firewalls that we don't cross about data privacy 782 00:41:34,480 --> 00:41:38,000 about like the extent to which fine tuning can be done in One 783 00:41:38,000 --> 00:41:40,560 Direction or another. I definitely think there need to 784 00:41:40,560 --> 00:41:43,160 be rules and regulations about what cannot, can and cannot be 785 00:41:43,160 --> 00:41:44,920 done that people need to agree to. 786 00:41:45,240 --> 00:41:47,400 That's really hard. And now I'm what we're calling a 787 00:41:47,400 --> 00:41:50,320 multi polar world where you got bad actors from different 788 00:41:50,320 --> 00:41:52,760 countries who got their own agendas who might not want to 789 00:41:52,760 --> 00:41:55,840 regulate. So it's dangerous and risky, but 790 00:41:56,440 --> 00:42:00,520 I want to be crystal clear though, is we need to have a 791 00:42:00,520 --> 00:42:04,360 conversation that's open, that's discussing these things openly. 792 00:42:04,360 --> 00:42:09,840 What I'm, what I see instead from academics mostly is a 793 00:42:09,840 --> 00:42:13,320 shutting down of the conversation of like moralizing 794 00:42:13,360 --> 00:42:15,480 AI. I'm saying it's evil. 795 00:42:15,600 --> 00:42:19,600 I've literally heard otherwise smart academics scream out loud 796 00:42:19,880 --> 00:42:24,200 on blue sky. Of course, you know AI is evil, 797 00:42:25,120 --> 00:42:27,600 right? And I mock, quote, tweeted it. 798 00:42:27,600 --> 00:42:29,560 And of course you had people explain to me, no, AI, it's not 799 00:42:29,560 --> 00:42:30,760 evil. It depends on what you do with 800 00:42:30,760 --> 00:42:32,480 it. I'm like, dude, I don't think AI 801 00:42:32,520 --> 00:42:35,120 is evil. But I see that from really smart 802 00:42:35,120 --> 00:42:39,360 people who are just moralizing. There are people who no amount 803 00:42:39,360 --> 00:42:43,120 of AI is good. And in fact, I've got a project 804 00:42:43,120 --> 00:42:45,480 now where we're looking at the moralization of AI. 805 00:42:45,480 --> 00:42:49,600 So moralization is when you have an attitude that then becomes 806 00:42:49,600 --> 00:42:51,960 black and white. So for example, we have a 807 00:42:51,960 --> 00:42:53,320 moralized attitude towards murder. 808 00:42:53,560 --> 00:42:57,120 No amount of murder is good and no amount of pedophilia is good. 809 00:42:57,920 --> 00:42:59,320 But even with murder, you're right, there are some 810 00:42:59,320 --> 00:43:02,400 exceptions. Yeah, self-defense murder I'm OK 811 00:43:02,440 --> 00:43:03,240 with. Yeah. 812 00:43:04,880 --> 00:43:07,280 Right. And then we can look at this 813 00:43:07,280 --> 00:43:08,520 with other things. So for example, there's some 814 00:43:08,520 --> 00:43:12,320 people who moralize GMO foods even though there's no evidence 815 00:43:12,320 --> 00:43:15,720 that GMO foods are bad for you. And there's some GMO foods now 816 00:43:15,720 --> 00:43:19,200 that are thought to be able to like infuse enough, I think that 817 00:43:19,200 --> 00:43:22,040 vitamin A to cure blindness in certain parts of the world where 818 00:43:22,040 --> 00:43:25,960 they have a heavy rice diet. And critics of this refuse. 819 00:43:25,960 --> 00:43:28,760 They're like, this is an evil. This is an abomination, even 820 00:43:28,760 --> 00:43:31,720 though it could save thousands if millions of people. 821 00:43:32,200 --> 00:43:33,560 And we're seeing the same thing with. 822 00:43:33,880 --> 00:43:37,440 AI, it's the playing God nature, and I think that ties into AI as 823 00:43:37,440 --> 00:43:39,600 well. Absolutely. 824 00:43:39,840 --> 00:43:44,760 And we're seeing that among opposers of AI over a majority 825 00:43:44,760 --> 00:43:47,960 of opposers of AI oppose it on moral grounds. 826 00:43:49,160 --> 00:43:52,040 It's saying that no amount of AI will be good in a legal context. 827 00:43:52,040 --> 00:43:55,920 No amount of AI will be good in a relational context for someone 828 00:43:55,920 --> 00:43:59,360 who's uninformed, whatever. But if you're a scientist, you 829 00:43:59,360 --> 00:44:01,880 have to see the pros and cons because it's happening whether 830 00:44:01,880 --> 00:44:04,240 you like it or not. Whether you like it or not, the 831 00:44:04,240 --> 00:44:06,640 businesses are going forward. So what can we do now to 832 00:44:06,640 --> 00:44:10,000 mitigate the problems, to solve the problems, to look at in the 833 00:44:10,000 --> 00:44:12,400 future and have regulations to help us out? 834 00:44:13,080 --> 00:44:14,960 Yeah. The thing for me that often 835 00:44:14,960 --> 00:44:18,600 times becomes clear is that kind of the people that says with 836 00:44:18,600 --> 00:44:23,120 self driving cars, whatever amount of people being killed by 837 00:44:23,120 --> 00:44:25,800 self driving car, whatever the number is, it's never going to 838 00:44:25,800 --> 00:44:27,480 be safe enough. It's like that kind of feeling 839 00:44:27,480 --> 00:44:31,480 sometimes, and you alluded to it before, where we put the human 840 00:44:31,480 --> 00:44:35,200 counterfactual on a pedestal up to a degree where we forget what 841 00:44:35,200 --> 00:44:38,040 actually is true. It's not the case that zero 842 00:44:38,040 --> 00:44:41,960 human drivers have killed other human pedestrians. 843 00:44:42,200 --> 00:44:45,360 Yeah, you see this a lot in AI criticism space. 844 00:44:45,560 --> 00:44:48,800 The people hold AI to an impossible standard. 845 00:44:49,000 --> 00:44:50,880 If AI makes one mistake, that's it. 846 00:44:50,880 --> 00:44:54,320 It should be thrown away. So this is book, a really good 847 00:44:54,320 --> 00:44:56,320 book actually. It's called AI Snake Oil, 848 00:44:56,320 --> 00:44:58,600 written by two computer scientists from Princeton. 849 00:44:59,840 --> 00:45:02,680 I read it but I liked the bulk of it. 850 00:45:02,680 --> 00:45:06,360 But I was also infuriated by parts of it because especially 851 00:45:06,360 --> 00:45:08,920 the part where it was criticizing predictive AI 852 00:45:09,320 --> 00:45:11,680 because it was point to all the mistakes that AI makes. 853 00:45:11,680 --> 00:45:14,680 I'm like OK sure if you're comparing it to perfection, but 854 00:45:14,760 --> 00:45:17,600 the contrast is not perfection. The contrast is what humans do, 855 00:45:18,000 --> 00:45:22,160 and if it's for example, saving even one more life in a hospital 856 00:45:22,280 --> 00:45:25,000 than what a human does, that seems like it's a good trade 857 00:45:25,000 --> 00:45:28,720 off. And you can continue to improve 858 00:45:28,720 --> 00:45:32,040 the technology, whereas with humans a whole lot harder to 859 00:45:32,040 --> 00:45:33,880 improve. Absolutely. 860 00:45:33,880 --> 00:45:36,680 Look at the example of like claims of AI being racist and 861 00:45:36,680 --> 00:45:38,800 biased and all, and what's your premise? 862 00:45:38,840 --> 00:45:44,120 Maybe we talked about this and we made huge strides in fixing 863 00:45:44,120 --> 00:45:45,480 that. OK. 864 00:45:45,920 --> 00:45:49,360 And now I'm pretty sure the AI is less biased. 865 00:45:49,360 --> 00:45:51,440 But you cannot correct humans that way. 866 00:45:51,720 --> 00:45:54,120 Like you cannot say, hey, you know that decision you made, 867 00:45:54,120 --> 00:45:56,840 It's a clear pattern of bias against one group over another. 868 00:45:57,080 --> 00:45:58,840 Fix that. Good luck. 869 00:45:59,080 --> 00:46:01,560 The bias training is probably one of the most ineffective 870 00:46:01,880 --> 00:46:04,080 things. Yeah, psychology has been trying 871 00:46:04,080 --> 00:46:06,840 this 40 years. When you talk about the 872 00:46:06,840 --> 00:46:11,000 moralization of AI, to what extent do you feel like that 873 00:46:11,000 --> 00:46:16,680 phenomenon that you found can explain this difference in how 874 00:46:16,680 --> 00:46:21,120 much people have a preference for interactions with a human 875 00:46:21,120 --> 00:46:24,360 versus an AI? So you find on the first hand 876 00:46:24,440 --> 00:46:29,080 that people prefer responses from AI over human responses. 877 00:46:29,080 --> 00:46:32,160 They find them to be more empathic and more compassionate 878 00:46:32,160 --> 00:46:35,040 and so on. But then when they actually have 879 00:46:35,040 --> 00:46:38,160 a choice between the two, they will still, if they know that 880 00:46:38,160 --> 00:46:42,040 it's human or an AI, when this label is there, they will choose 881 00:46:42,040 --> 00:46:44,720 the human over the AI. So how much do you think that 882 00:46:44,720 --> 00:46:48,520 these like sacred values and this moralization plays into 883 00:46:48,520 --> 00:46:51,200 that preference versus some other factors? 884 00:46:51,880 --> 00:46:54,800 Yeah, I think it plays into it. So I think in that if you're 885 00:46:54,800 --> 00:46:56,800 referring to what we call like some sort of empathy choice 886 00:46:56,800 --> 00:47:00,600 paradox, where people say the utterances of AI are better, but 887 00:47:00,600 --> 00:47:04,960 they prefer to receive empathy from humans, although I'm not 888 00:47:04,960 --> 00:47:07,840 100% positive, I believe we have some clues that, you know, the 889 00:47:07,840 --> 00:47:11,200 more familiarity you have with AI, the more you're OK with 890 00:47:11,200 --> 00:47:14,760 receiving empathy from AI. So I think familiarity plays a 891 00:47:14,760 --> 00:47:16,880 role. I do think moralization plays 892 00:47:16,880 --> 00:47:19,440 some role as well. If you're like have black and 893 00:47:19,440 --> 00:47:22,200 white thinking about it, even if you could acknowledge, oh, that 894 00:47:22,200 --> 00:47:25,600 was a better statement, but still, I want to receive it from 895 00:47:25,600 --> 00:47:27,160 a human that has to play some role. 896 00:47:27,480 --> 00:47:32,480 But I also think it can also be just rational as well, that in 897 00:47:32,480 --> 00:47:34,920 the end of the day, I want someone human on the other side 898 00:47:35,120 --> 00:47:39,320 who might have more understanding, not fully getting 899 00:47:39,320 --> 00:47:40,800 that. Like, if your communication is 900 00:47:40,800 --> 00:47:45,400 only going to be mediated by a screen anyways, does it matter 901 00:47:45,600 --> 00:47:48,200 if the person on the other side is flesh or bones or not? 902 00:47:48,320 --> 00:47:52,920 If they're saying the words that make you feel, shouldn't that be 903 00:47:52,920 --> 00:47:55,400 enough? But I think we still have a 904 00:47:55,400 --> 00:47:57,240 preference for reality. We have a preference for real 905 00:47:57,240 --> 00:47:58,400 things, that we're not real things. 906 00:47:58,400 --> 00:48:00,480 And probably there are lots of good reasons for that. 907 00:48:01,640 --> 00:48:05,120 Well, and it's interesting that you say the that familiarity or 908 00:48:05,120 --> 00:48:10,760 maybe experience using AI, and I assume this is generative AI, it 909 00:48:10,760 --> 00:48:15,440 also sort of diminishes this. Do you think that as people 910 00:48:15,440 --> 00:48:20,160 become more familiar, everyone is using AI tools in maybe 5 or 911 00:48:20,160 --> 00:48:23,960 10 years that this difference, this this difference in 912 00:48:23,960 --> 00:48:26,440 preference will diminish or disappear? 913 00:48:27,160 --> 00:48:29,240 Yeah, I definitely think so. I don't know what disappear 914 00:48:29,240 --> 00:48:32,200 altogether. It's again, even though I've 915 00:48:32,200 --> 00:48:34,120 argued the opposite side of this, I do think reality is 916 00:48:34,120 --> 00:48:35,880 important. It doesn't matter for some 917 00:48:35,880 --> 00:48:37,360 obvious and not so obvious reasons. 918 00:48:38,240 --> 00:48:40,080 So I don't think we'll go well together. 919 00:48:40,080 --> 00:48:44,400 But I do think as you become more familiar, we won't think 920 00:48:44,400 --> 00:48:47,000 it's weird. We won't think it's as weird to 921 00:48:47,000 --> 00:48:50,240 have a friend that's AI now. We think it's something's wrong 922 00:48:50,240 --> 00:48:50,720 with you. Can't. 923 00:48:51,160 --> 00:48:54,120 You get a real friend. Yeah, we feel really bad for 924 00:48:54,120 --> 00:48:56,640 you. Or like you had no choice, like 925 00:48:56,640 --> 00:48:58,880 you're suffering, so OK, it's excusable. 926 00:48:58,880 --> 00:49:00,600 You're housebound. You're disabled. 927 00:49:01,200 --> 00:49:04,160 Good for you. It just occurred to me maybe an 928 00:49:04,160 --> 00:49:06,080 interesting parallel is online dating. 929 00:49:06,160 --> 00:49:09,240 I remember in the early days of online dating that was like, oh, 930 00:49:09,720 --> 00:49:12,120 you can't find a real person out in the real world. 931 00:49:12,120 --> 00:49:13,920 But now it's like, totally normal. 932 00:49:14,120 --> 00:49:18,080 I know many people who have met their partner in that. 933 00:49:18,360 --> 00:49:21,320 Yeah. Tell us about it. 934 00:49:22,320 --> 00:49:24,480 Yeah, it's normal. But that was not always the 935 00:49:24,480 --> 00:49:26,920 case. Anyway, things become normalized 936 00:49:27,040 --> 00:49:29,480 over time with familiarity. Yeah, yeah. 937 00:49:29,640 --> 00:49:33,680 And I think when it comes to, let's say, a coach, so someone 938 00:49:33,680 --> 00:49:35,960 who serves a function to motivate you and help you stay 939 00:49:35,960 --> 00:49:38,320 accountable. Currently, I would say people 940 00:49:38,320 --> 00:49:40,760 experience a difference when it comes to like if they know that 941 00:49:40,760 --> 00:49:43,600 there's a human person that says, hey, you should go to the 942 00:49:43,600 --> 00:49:46,480 gym, you should do this and this, they feel a little bit bad 943 00:49:46,480 --> 00:49:49,520 if they don't do it. But if it's an AI who says the 944 00:49:49,520 --> 00:49:53,240 same thing, it would probably likely feel a bit less bad to 945 00:49:53,240 --> 00:49:55,440 missing the chairman doing so. I feel a little bit stronger 946 00:49:55,440 --> 00:49:57,440 sense of motivation and accountability to that human. 947 00:49:58,720 --> 00:50:01,360 Where do you think that comes from? 948 00:50:01,600 --> 00:50:03,320 Do you think that's going to diminish as well? 949 00:50:04,400 --> 00:50:06,000 I mean again, I think it's the same thing because it's not 950 00:50:06,000 --> 00:50:08,080 real. So if a friend is disappointed 951 00:50:08,080 --> 00:50:11,960 in you, there could be costs. Like depending on the level of 952 00:50:11,960 --> 00:50:15,320 disappointment, they might not return your call next time, they 953 00:50:15,600 --> 00:50:19,120 might not be as friendly as you next time until the AI starts 954 00:50:19,120 --> 00:50:21,720 having a bit more friction. So then that's going to be the 955 00:50:21,720 --> 00:50:25,960 next step with AI development. It's too agreeable right now and 956 00:50:26,640 --> 00:50:29,880 it needs to start, you know, imposing costs on you. 957 00:50:30,160 --> 00:50:32,480 So for example, you know, you're talking about a coach. 958 00:50:32,560 --> 00:50:36,680 If you're AI recommends you do XY and you don't repeatedly, 959 00:50:37,200 --> 00:50:40,920 maybe the cost it imposes like OK well I see you like using me 960 00:50:40,920 --> 00:50:43,120 a lot. I'm just going to take away a 961 00:50:43,120 --> 00:50:46,680 few tokens from you or take something away from you that you 962 00:50:46,680 --> 00:50:48,600 want. I'm sure you can do some nudges 963 00:50:48,600 --> 00:50:51,400 in there that can make it have a bit more friction. 964 00:50:51,960 --> 00:50:54,840 I'm just seeing myself in the future where you like, you see 965 00:50:54,840 --> 00:50:56,280 it's online, but it's not responding to you. 966 00:50:56,280 --> 00:50:58,200 And you're like, wait, I'll see you there. 967 00:50:58,200 --> 00:50:59,440 I'll see you. You know why you're not 968 00:50:59,440 --> 00:51:00,720 responding? To it, absolutely. 969 00:51:00,840 --> 00:51:02,360 I think the next step will be they're going to have 970 00:51:02,360 --> 00:51:04,600 personalities. So not just the personality you 971 00:51:04,600 --> 00:51:07,200 develop, but it's going to have, I don't know, let's start out 972 00:51:07,200 --> 00:51:10,000 with like 20 different kinds of personalities for different 973 00:51:10,000 --> 00:51:12,840 combinations of the big 5. Let's say that makes sense 974 00:51:13,080 --> 00:51:17,120 theoretically. And what that will mean is not 975 00:51:17,120 --> 00:51:20,080 every combination is going to like you, and you're not going 976 00:51:20,080 --> 00:51:21,640 to like every combination either. 977 00:51:22,280 --> 00:51:24,800 And then there'll be some matches that, OK, you're going 978 00:51:24,800 --> 00:51:27,000 to go, but then you're like, oh, there's just some really basic 979 00:51:27,000 --> 00:51:31,160 incompatibilities here. And then maybe they're going to 980 00:51:31,160 --> 00:51:32,640 break up with you. You have to find another one. 981 00:51:32,920 --> 00:51:36,160 And I think that will not be bad because especially if you want 982 00:51:36,160 --> 00:51:40,480 to be a good functioning human in the real world, hopefully 983 00:51:40,480 --> 00:51:42,960 these personalities are mimicking the kinds of 984 00:51:42,960 --> 00:51:45,040 contingencies that exist in the real world too. 985 00:51:45,200 --> 00:51:48,840 Whereas if you're a Dick to someone like they don't take it, 986 00:51:48,880 --> 00:51:51,240 at least typically. They shouldn't, yeah. 987 00:51:51,240 --> 00:51:53,720 And they shouldn't, right? Here's another future. 988 00:51:53,880 --> 00:51:57,400 Maybe the friendship, the companion AI or the therapist AI 989 00:51:57,400 --> 00:52:01,240 isn't replacing the human variance, but they're the 24 990 00:52:01,240 --> 00:52:05,520 hour round the clock kind of helper between sessions. 991 00:52:05,880 --> 00:52:09,000 So I only see my therapist once a week and they give me all this 992 00:52:09,000 --> 00:52:10,040 advice. But guess what? 993 00:52:10,040 --> 00:52:12,120 In the heat of the moment, sometimes you want to talk to 994 00:52:12,120 --> 00:52:14,440 your therapist about something and you can't. 995 00:52:14,720 --> 00:52:16,280 And you got to hold it down, write it down. 996 00:52:16,280 --> 00:52:18,800 Maybe a week later you speak to them and by then the heat's 997 00:52:18,800 --> 00:52:21,040 gone. Here you've got your therapist 998 00:52:21,040 --> 00:52:23,160 in your pocket and you hopefully listen. 999 00:52:23,160 --> 00:52:24,960 If you don't, there's friction, there's some sort of 1000 00:52:24,960 --> 00:52:27,760 consequence, and it's scaffolding for the next 1001 00:52:27,760 --> 00:52:29,640 meetings. These are all kinds of uses we 1002 00:52:29,640 --> 00:52:33,160 could have of AI that is not replacement level that could 1003 00:52:33,320 --> 00:52:36,000 help us. Assuming that your AI therapist 1004 00:52:36,000 --> 00:52:39,520 agrees with your human therapist, it's like hopefully 1005 00:52:39,520 --> 00:52:42,840 they're on the same wavelength. Hopefully, and if they disagree 1006 00:52:42,880 --> 00:52:45,080 I think I'd probably go with the AI therapist. 1007 00:52:45,080 --> 00:52:46,640 They're less fallible and I suspect. 1008 00:52:48,560 --> 00:52:52,400 I feel like this is the flip side of the chat pot, making 1009 00:52:52,400 --> 00:52:56,120 people more willing to disclose information and that's because 1010 00:52:56,120 --> 00:52:59,360 you don't have that judgement coming at you. 1011 00:52:59,560 --> 00:53:03,840 You can see the flip side of the AI coach isn't going to judge 1012 00:53:03,840 --> 00:53:08,680 you or hold you accountable, and so in that sense it's actually a 1013 00:53:08,680 --> 00:53:11,520 negative, whereas with disclosure it can be a positive. 1014 00:53:12,120 --> 00:53:13,160 Right. Yeah, There have to be some sort 1015 00:53:13,160 --> 00:53:16,840 of balance and AI probably shouldn't meet all our needs all 1016 00:53:16,840 --> 00:53:19,400 the time at any given moment because I think then it would 1017 00:53:19,400 --> 00:53:22,520 lead to, you know, externalities, It'll lead to 1018 00:53:22,560 --> 00:53:25,280 other things that that will be bad for you in the long term. 1019 00:53:25,320 --> 00:53:29,040 When it comes to, I call it human buffering, like we felt 1020 00:53:29,040 --> 00:53:31,720 frustrated when the Internet was buffering before and like 1021 00:53:31,720 --> 00:53:34,120 someone was loading slowly. And now we're interacting a lot 1022 00:53:34,120 --> 00:53:35,960 with AI. If they're very quick to respond 1023 00:53:35,960 --> 00:53:38,680 to things and are really instant and we can interrupt them and we 1024 00:53:38,680 --> 00:53:40,600 can interact with AI very different than the humans. 1025 00:53:40,760 --> 00:53:42,600 And then suddenly we interact with the human again and we're 1026 00:53:42,600 --> 00:53:45,760 like, oh, this person is speaking pretty slow. 1027 00:53:45,760 --> 00:53:49,000 And sometimes, you know, we feel like, OK, feed up a little bit, 1028 00:53:49,000 --> 00:53:52,120 but you can't just it's. Like when people meet us in real 1029 00:53:52,120 --> 00:53:53,240 life. Exactly. 1030 00:53:53,240 --> 00:53:55,480 They're like I usually listen to you on 2.5. 1031 00:53:55,880 --> 00:53:58,880 But I guess I'm interested with empathy, like how much have you 1032 00:53:58,880 --> 00:54:02,200 seen this, some form of potential for empathy drift that 1033 00:54:02,200 --> 00:54:04,720 we interact with a very empathetic AI. 1034 00:54:04,720 --> 00:54:07,760 We have real high perceived empathy from this AI. 1035 00:54:07,760 --> 00:54:09,760 And then we go back to our day-to-day life. 1036 00:54:09,760 --> 00:54:11,960 We see our partner, we see our friends, They don't seem less 1037 00:54:11,960 --> 00:54:12,920 sympathetic anymore. It's. 1038 00:54:13,360 --> 00:54:16,280 Certainly a possibility that we have written about it, just as 1039 00:54:16,280 --> 00:54:20,040 most people can no longer read a map because they've got Google 1040 00:54:20,040 --> 00:54:23,280 Maps. If we outsource our empathy to a 1041 00:54:23,280 --> 00:54:26,760 machine, our own empathic capabilities will be degraded. 1042 00:54:27,440 --> 00:54:31,040 That's possible and that's real. You know, I haven't seen 1043 00:54:31,040 --> 00:54:33,720 evidence of this yet, but it just begun. 1044 00:54:33,720 --> 00:54:38,240 It's something we should look out for and I don't exactly know 1045 00:54:38,240 --> 00:54:43,040 what to do with that. So if it means on the whole, 1046 00:54:43,040 --> 00:54:45,520 we're more empathic in our writing and our written 1047 00:54:45,520 --> 00:54:48,560 communication and even maybe verbal communication, because 1048 00:54:48,560 --> 00:54:51,080 I'm looking at like my senior debris rack in my pocket, that's 1049 00:54:51,080 --> 00:54:54,960 like, tell me what to say. Maybe on the whole, it's not so 1050 00:54:54,960 --> 00:54:56,480 bad. So I can get around to a lot 1051 00:54:56,480 --> 00:54:59,760 more places now. In fact, is it such a huge loss 1052 00:54:59,760 --> 00:55:01,160 that I can't? I mean, I can tell me the map, 1053 00:55:01,160 --> 00:55:03,120 but then many people can't read maps. 1054 00:55:03,960 --> 00:55:05,800 Is it a huge loss? Because I think fewer people are 1055 00:55:05,800 --> 00:55:08,560 getting lost. So would it be a huge loss if 1056 00:55:08,560 --> 00:55:10,840 this empathic qualities are diminished? 1057 00:55:10,840 --> 00:55:15,480 I mean, I think it could lead to some awkward moments in real 1058 00:55:15,480 --> 00:55:19,280 life and that would be good, but if on the whole people are more 1059 00:55:19,280 --> 00:55:20,840 empathic, maybe it's not terrible. 1060 00:55:21,280 --> 00:55:23,920 But right now, these are all hypotheticals, because it's not 1061 00:55:23,920 --> 00:55:25,120 clear that's even going to happen. 1062 00:55:26,160 --> 00:55:29,720 All right, it is time for us to move on to our quick fire round. 1063 00:55:30,080 --> 00:55:33,360 This we call to AI or not to AI. Are you ready? 1064 00:55:34,280 --> 00:55:38,160 I am. So first one to AI or not to AIA 1065 00:55:38,160 --> 00:55:41,680 Shut BFF algorithmic companion that access your new virtual 1066 00:55:41,680 --> 00:55:44,080 best friend? Sure, AI. 1067 00:55:45,080 --> 00:55:49,800 And Empathy Coach AI giving you support ahead of or live prompts 1068 00:55:49,800 --> 00:55:51,800 during tough talks. AI. 1069 00:55:52,760 --> 00:55:57,800 Yeah, you just described this. So to AI or not to AI Tamaguchi 1070 00:55:57,800 --> 00:56:02,640 style pets that only thrives and survives if and when you donate 1071 00:56:02,640 --> 00:56:08,600 to the most effective charters. And the question here is the AI 1072 00:56:08,760 --> 00:56:10,880 is the Tamaguchi, what's the AI here? 1073 00:56:11,160 --> 00:56:12,960 Yeah. So basically like this is a good 1074 00:56:12,960 --> 00:56:16,760 idea to have this basically thing that adds some form of 1075 00:56:16,960 --> 00:56:19,480 feeling of empty, like feeling like you're doing something good 1076 00:56:19,480 --> 00:56:24,400 for something while you're kind of doing the most cost effective 1077 00:56:24,880 --> 00:56:26,320 charity generations at the same time. 1078 00:56:26,520 --> 00:56:28,400 Sure, that sounds like fun. Why not AI? 1079 00:56:28,480 --> 00:56:34,560 Sure, I'm like perfect of our. OK, I think I'm going to turn 1080 00:56:34,560 --> 00:56:37,600 you on this one. The glucose gauge, a smart 1081 00:56:37,600 --> 00:56:40,800 visualization of your blood glucose showing how much is left 1082 00:56:40,800 --> 00:56:42,560 before it runs out. And you need a doughnut. 1083 00:56:44,480 --> 00:56:50,560 No AI unless it's just like a fun fun way to get me angry. 1084 00:56:53,320 --> 00:56:58,520 OK Live Empathy Scoreboard, a real time empathy meter to see 1085 00:56:58,520 --> 00:57:00,440 how much empathy you express throughout the day. 1086 00:57:01,560 --> 00:57:07,960 Oh, I like that to AI. Courtroom conscience scanner a 1087 00:57:07,960 --> 00:57:11,400 tool to evaluate whether a defendant in court feels genuine 1088 00:57:11,400 --> 00:57:13,960 remorse or just wants a lighter sentence. 1089 00:57:15,720 --> 00:57:18,800 Oh, that's a tough one. I would say not to AI there. 1090 00:57:18,800 --> 00:57:21,520 I think there's some rights of privacy for that person. 1091 00:57:22,520 --> 00:57:27,080 Fate Fit Matchmakers, an AI religion recommender for 1092 00:57:27,080 --> 00:57:30,120 atheists experiencing anxiety that would kind of like match 1093 00:57:30,120 --> 00:57:33,960 people with the best fitting religion to suit their stress or 1094 00:57:33,960 --> 00:57:36,640 anxieties. Sure, why not AI? 1095 00:57:36,640 --> 00:57:38,920 That sounds like fun. A matchmaker. 1096 00:57:40,080 --> 00:57:45,760 Yeah, for atheists, Guilt offloader, digital confessor 1097 00:57:45,760 --> 00:57:48,600 that absorbs your moral hang ups so you can act freely. 1098 00:57:49,320 --> 00:57:51,240 No, AI guilt is a useful emotion. 1099 00:57:51,240 --> 00:57:55,200 You don't want to get rid of it. Well, that brings us to our last 1100 00:57:55,200 --> 00:57:59,800 question, which is what is your most controversial opinion about 1101 00:57:59,920 --> 00:58:03,400 AI? OK, so you did warn me about 1102 00:58:03,400 --> 00:58:05,560 this one and I had to think long and hard about it. 1103 00:58:05,560 --> 00:58:07,920 And I actually might write about it in the coming months on my 1104 00:58:07,920 --> 00:58:12,440 sub stack, which if I can plug it, is called Speak now, Regret 1105 00:58:12,440 --> 00:58:13,560 later. It's on sub stack. 1106 00:58:13,960 --> 00:58:20,040 I enjoy writing it, so this is not a fully formed thought yet, 1107 00:58:20,080 --> 00:58:23,600 but I think there's something fundamentally special about 1108 00:58:23,600 --> 00:58:28,480 writing and that it would be a shame if AI took over our most 1109 00:58:28,480 --> 00:58:31,240 of our writing. Not the crappy copywriting like 1110 00:58:31,240 --> 00:58:33,840 form stuff. I think most of us agree AI is 1111 00:58:33,840 --> 00:58:36,320 good for that. But creative writing where 1112 00:58:36,360 --> 00:58:38,560 you're really trying to get your heart out, or even a scientific 1113 00:58:38,560 --> 00:58:44,400 paper for example, the writing is thought to be a way that we 1114 00:58:44,400 --> 00:58:46,520 think about things. It helps us actually formulate 1115 00:58:46,520 --> 00:58:48,960 our thoughts. It's not the case that when we 1116 00:58:48,960 --> 00:58:51,480 write, we're literally just emptying out our minds onto the 1117 00:58:51,480 --> 00:58:53,560 paper. The act of writing and active 1118 00:58:53,560 --> 00:58:55,240 creation. And I think that's true. 1119 00:58:55,960 --> 00:58:59,320 And there's a danger that AI will prevent people from having 1120 00:58:59,320 --> 00:59:01,560 to think through some of the writing problems and problems of 1121 00:59:01,560 --> 00:59:03,880 their arguments. I'm not sure that's true 1122 00:59:04,400 --> 00:59:07,360 because, and I wonder if AI can take over that as well. 1123 00:59:07,520 --> 00:59:11,320 I think if that happens, I think an individual style and voice 1124 00:59:11,760 --> 00:59:14,440 gets compromised and that's not great and we lose that 1125 00:59:14,440 --> 00:59:18,560 individual style and maybe our world is less rich without style 1126 00:59:18,560 --> 00:59:21,320 not being there. But in my own experimentation 1127 00:59:21,320 --> 00:59:26,160 with writing, with AII, go through many iterations, many 1128 00:59:26,160 --> 00:59:28,760 drafts, back and forth, I'm writing, I prompt back and forth 1129 00:59:29,080 --> 00:59:33,200 and my idea, my actual ideas still get fleshed out. 1130 00:59:33,520 --> 00:59:35,640 And sometimes they're fleshed out even better 'cause I asked 1131 00:59:35,720 --> 00:59:39,640 AI to probe my ideas, to push back against my ideas, and 1132 00:59:39,640 --> 00:59:43,280 they're actually improved. So I don't think my thinking is 1133 00:59:43,280 --> 00:59:48,000 diminished is getting worse by using AI as to help me with my 1134 00:59:48,000 --> 00:59:50,400 right ink. In fact, I write very personal 1135 00:59:50,400 --> 00:59:54,400 things that are, I mean, I don't think I've run any single thing 1136 00:59:54,400 --> 00:59:57,120 where AI has run everything. It helps me start things for 1137 00:59:57,120 --> 01:00:00,960 sure, but I scaffold it, add personal stories, add my own 1138 01:00:00,960 --> 01:00:03,680 personal elements. But the argument structure is, 1139 01:00:03,760 --> 01:00:06,120 are largely what AI is given. And I don't think I'm losing 1140 01:00:06,120 --> 01:00:08,520 anything. And now of course, for a young 1141 01:00:08,520 --> 01:00:11,480 student who doesn't know how to write and to just feed it with a 1142 01:00:11,480 --> 01:00:14,360 few prompts and let it go, then yes, they're not doing their 1143 01:00:14,360 --> 01:00:16,120 actual work. They're not doing the thinking. 1144 01:00:16,440 --> 01:00:18,800 And the person on the other hand, won't be able to know that 1145 01:00:18,800 --> 01:00:20,880 or tell that. So that's not great. 1146 01:00:20,880 --> 01:00:23,800 But in terms of losing the ability to think, I don't think 1147 01:00:23,800 --> 01:00:26,960 AI will actually lead there. I think there are degrees of 1148 01:00:26,960 --> 01:00:30,360 this too, where it sounds like you're engaging quite a bit with 1149 01:00:30,360 --> 01:00:34,160 a lot of back and forth and your expertise is also going in like 1150 01:00:34,160 --> 01:00:37,560 feeding into this. And so I think there's quite a 1151 01:00:37,560 --> 01:00:40,480 big difference between, you know, maybe this is exactly what 1152 01:00:40,480 --> 01:00:44,720 you said, but that version of quote UN quote writing with AI 1153 01:00:44,720 --> 01:00:48,280 where really a lot of you is going into the writing versus 1154 01:00:48,280 --> 01:00:51,440 someone who just says, I want to article about, you know, love 1155 01:00:51,440 --> 01:00:53,160 and war, please. Yeah, totally. 1156 01:00:53,160 --> 01:00:55,240 So I didn't have secondary use case and we all agree that just 1157 01:00:55,280 --> 01:00:57,080 laziness and you're not benefiting. 1158 01:00:57,320 --> 01:01:00,120 But I still think some people believe even that the use case 1159 01:01:00,120 --> 01:01:01,960 that I described, I use it all the time now. 1160 01:01:02,560 --> 01:01:05,160 I think they would say that's not ideal either because it's 1161 01:01:05,160 --> 01:01:07,920 not your voice. And I think, no, I think you can 1162 01:01:07,920 --> 01:01:10,040 get your voice out and I think you can get your ideas out 1163 01:01:10,040 --> 01:01:11,960 clearly. It's succinctly and well. 1164 01:01:12,640 --> 01:01:16,000 Well, don't you think that you imbue it with your voice too in 1165 01:01:16,000 --> 01:01:18,360 your iterations? Oh, totally. 1166 01:01:18,360 --> 01:01:21,640 I mean, you can also now have projects on various LLMS where 1167 01:01:22,040 --> 01:01:25,040 it learns from your own voice, it learns how you write and your 1168 01:01:25,040 --> 01:01:26,960 own kind of phrase. It's actually, it's a great way 1169 01:01:26,960 --> 01:01:31,000 to learn what phrase you overuse and what kind of, you know, what 1170 01:01:31,000 --> 01:01:32,720 your crutches are in writing as well, you see. 1171 01:01:32,720 --> 01:01:34,680 Oh, it says that thing all the time. 1172 01:01:34,680 --> 01:01:38,200 I guess I must do that too. What's the most annoying phrase 1173 01:01:38,200 --> 01:01:41,440 that you say? My writing is I like it to be 1174 01:01:41,440 --> 01:01:44,840 conversational. So I said like, The thing is, or 1175 01:01:44,840 --> 01:01:46,440 here's the kicker, I say that a lot. 1176 01:01:46,440 --> 01:01:47,800 I'm like, what does that even mean? 1177 01:01:47,800 --> 01:01:49,520 I don't even know but I use that a lot. 1178 01:01:49,880 --> 01:01:52,960 Well, this is really fun. Mickey, this has been lovely. 1179 01:01:52,960 --> 01:01:55,440 Thank you so much for joining us on the show. 1180 01:01:55,440 --> 01:01:57,960 Thanks so much for having me on and and good luck with the rest 1181 01:01:57,960 --> 01:01:59,600 of it. And that's a wrap. 1182 01:01:59,920 --> 01:02:02,600 You've been listening to the Behavioral Design Podcast 1183 01:02:02,880 --> 01:02:05,440 brought to you by Habit Weekly and Nuanced Behavior. 1184 01:02:05,800 --> 01:02:08,080 Sam and Alene tell me. This season is packed with 1185 01:02:08,080 --> 01:02:11,960 incredible insights about behavioral design and AI, so be 1186 01:02:11,960 --> 01:02:14,560 sure to subscribe and share the podcast with your friends. 1187 01:02:14,800 --> 01:02:16,960 Though you might want to keep it away from your enemies. 1188 01:02:18,320 --> 01:02:20,920 In case you haven't noticed, I'm an AI voice. 1189 01:02:22,120 --> 01:02:24,880 Yep, pretty crazy. Quite the improvement since last 1190 01:02:24,880 --> 01:02:26,880 season's AI outro, don't you think? 1191 01:02:28,160 --> 01:02:30,800 If you'd like to collaborate with us at Nuance Behavior, 1192 01:02:31,000 --> 01:02:33,720 where we use behavioral design to craft digital products with 1193 01:02:33,720 --> 01:02:38,040 Nuance, e-mail us at hello@nuancebehavior.com or book 1194 01:02:38,040 --> 01:02:41,320 a call directly on our website, nuancebehavior.com. 1195 01:02:42,720 --> 01:02:46,160 A special thanks to the amazing Dave Pizarro for our show music 1196 01:02:46,400 --> 01:02:49,280 and to Mei Chen Yap and April English for their help in 1197 01:02:49,280 --> 01:02:51,200 producing and publishing this episode. 1198 01:02:51,640 --> 01:02:54,480 Thanks again for tuning in. We'll be back soon with another 1199 01:02:54,480 --> 01:02:57,720 exciting conversation where behavioral design and AI 1200 01:02:57,720 --> 01:03:00,520 intersect. Happens. 1201 01:03:01,560 --> 01:03:05,440 To. Mugatroid. 1202 01:03:28,680 --> 01:03:28,880 The.