1 00:00:00,520 --> 00:00:04,880 Speaker 1: We're protecting your identity today because AI was obsessed with you. 2 00:00:06,040 --> 00:00:07,880 Speaker 2: It felt like the AI was stalking me. 3 00:00:08,720 --> 00:00:10,400 Speaker 3: I was just so weird it owt. 4 00:00:11,720 --> 00:00:14,280 Speaker 4: What does it mean to be stalked by AI? 5 00:00:14,920 --> 00:00:16,920 Speaker 1: What does it even feel like when a piece of 6 00:00:16,960 --> 00:00:21,040 Speaker 1: technology seems to be obsessed with you, like you personally? 7 00:00:21,600 --> 00:00:23,919 Speaker 1: What does that look like? You might have heard the 8 00:00:24,040 --> 00:00:28,200 Speaker 1: term AI hallucination. This is a very modern phenomenon that 9 00:00:28,240 --> 00:00:32,839 Speaker 1: involves large language models or llms. Llms are systems that 10 00:00:32,960 --> 00:00:36,720 Speaker 1: train AI on the human world using massive amounts of text. 11 00:00:37,159 --> 00:00:41,920 Speaker 1: So an AI hallucination is when these lms generate false, illogical, 12 00:00:42,120 --> 00:00:45,080 Speaker 1: or unverified information, but does it with. 13 00:00:45,120 --> 00:00:46,200 Speaker 4: Very high confidence. 14 00:00:46,280 --> 00:00:49,240 Speaker 1: So I've described this to friends as like the drunk 15 00:00:49,280 --> 00:00:52,640 Speaker 1: frack guy at the party who's just confidently wrong and 16 00:00:52,720 --> 00:00:56,319 Speaker 1: just man'splaining things over and over again. Well, what if 17 00:00:56,360 --> 00:01:00,000 Speaker 1: the story AI was making up ended up involving you. 18 00:01:00,440 --> 00:01:02,040 Speaker 4: It used your name, and these. 19 00:01:01,960 --> 00:01:04,600 Speaker 1: Made up stories that also happen to match real life 20 00:01:04,600 --> 00:01:09,000 Speaker 1: details about you, your job, your interests, so much so 21 00:01:09,440 --> 00:01:15,120 Speaker 1: that people started believing these AI stories were actually about you. Well, 22 00:01:15,200 --> 00:01:17,319 Speaker 1: that's exactly what happened to our guest today. And I 23 00:01:17,400 --> 00:01:20,360 Speaker 1: know this is going to sound crazy, but this wild 24 00:01:20,480 --> 00:01:24,240 Speaker 1: story that Ai was spouting, it trickled off the Internet 25 00:01:24,319 --> 00:01:27,279 Speaker 1: and very much messed with our guest in the real world. 26 00:01:27,720 --> 00:01:30,200 Speaker 1: I'm Laurie Siegel, and you're listening to Mostly Human, a 27 00:01:30,319 --> 00:01:35,880 Speaker 1: tech podcast through a human lens. 28 00:01:39,959 --> 00:01:40,840 Speaker 4: Thank you for being here. 29 00:01:41,360 --> 00:01:42,320 Speaker 2: Thank you for having me. 30 00:01:42,520 --> 00:01:45,240 Speaker 1: What an odd reason we are going to be speaking. 31 00:01:45,760 --> 00:01:48,880 Speaker 1: I know, but I'm really excited that you're here. 32 00:01:49,760 --> 00:01:50,160 Speaker 4: Okay. 33 00:01:50,600 --> 00:01:55,080 Speaker 1: AI has become obsessed with you so much so that 34 00:01:55,160 --> 00:01:56,720 Speaker 1: we've made the decision. 35 00:01:56,400 --> 00:01:57,880 Speaker 4: That we are going to hide your identity. 36 00:01:58,120 --> 00:02:01,960 Speaker 1: I have interviewed people in the asked and I've hid 37 00:02:02,040 --> 00:02:05,040 Speaker 1: identities for certain reasons, and always for me as a storyteller. 38 00:02:05,040 --> 00:02:07,040 Speaker 1: I'm like, if someone asked to have their identity hidden, 39 00:02:07,280 --> 00:02:08,560 Speaker 1: there has to be a good reason. 40 00:02:08,840 --> 00:02:10,799 Speaker 4: Yeah, for us to you know, to. 41 00:02:10,800 --> 00:02:13,799 Speaker 1: Say, Okay, I believe that you should be granted anonymity 42 00:02:14,000 --> 00:02:16,079 Speaker 1: because what you're talking about is important and there could 43 00:02:16,080 --> 00:02:19,560 Speaker 1: be safety implications. So I just get the sense for 44 00:02:19,639 --> 00:02:23,000 Speaker 1: you that, you know, putting your real name out there 45 00:02:23,240 --> 00:02:25,720 Speaker 1: is disturbing at the time, given what's happened to you. Yeah. 46 00:02:25,800 --> 00:02:28,400 Speaker 2: Completely, you know, I really considered not hiding my identity 47 00:02:28,440 --> 00:02:30,399 Speaker 2: for this because I thought, you know, maybe it would 48 00:02:30,440 --> 00:02:33,000 Speaker 2: help because I could be more out there with it. 49 00:02:33,360 --> 00:02:36,280 Speaker 2: But I think the biggest issue is that a lot 50 00:02:36,320 --> 00:02:40,120 Speaker 2: of this story circles around my identity and the fact 51 00:02:40,120 --> 00:02:43,200 Speaker 2: that my identity has been taken and kind of like 52 00:02:43,800 --> 00:02:47,000 Speaker 2: changed basically in the public view. Like, what you're putting 53 00:02:47,000 --> 00:02:50,720 Speaker 2: out nowadays isn't just your identity like on Google through 54 00:02:50,800 --> 00:02:54,399 Speaker 2: like search engine optimization. It's now what models might hallucinate 55 00:02:54,400 --> 00:02:56,680 Speaker 2: about you. And so if I'm putting my name out 56 00:02:56,680 --> 00:03:00,000 Speaker 2: there even more with this podcast, I'm like, oh god, 57 00:03:00,000 --> 00:03:01,480 Speaker 2: how is that just going to add to what I've 58 00:03:01,480 --> 00:03:04,080 Speaker 2: already experienced and what I've already been worrying about. So 59 00:03:04,720 --> 00:03:06,320 Speaker 2: the whole idea is like I just kind of want 60 00:03:06,320 --> 00:03:09,000 Speaker 2: to like lay low now and like keep my full 61 00:03:09,080 --> 00:03:10,880 Speaker 2: name out of everything that I can. 62 00:03:11,120 --> 00:03:13,840 Speaker 1: Yeah, you've you have felt exposed by what happened, and 63 00:03:13,840 --> 00:03:15,600 Speaker 1: so you don't want to expose yourself any further. 64 00:03:16,000 --> 00:03:16,480 Speaker 3: Exactly. 65 00:03:16,960 --> 00:03:19,160 Speaker 1: This is the first time I'm ever talking to someone 66 00:03:19,320 --> 00:03:23,720 Speaker 1: who who really has a story about their identity used 67 00:03:23,840 --> 00:03:24,919 Speaker 1: or abused. 68 00:03:24,400 --> 00:03:27,079 Speaker 4: However we want to call it by a large language model. 69 00:03:27,639 --> 00:03:30,079 Speaker 2: It's crazy, and I really really hope it's the last 70 00:03:30,120 --> 00:03:30,600 Speaker 2: case of it. 71 00:03:31,040 --> 00:03:33,320 Speaker 1: So much of this story has to do with your name. 72 00:03:33,720 --> 00:03:34,920 Speaker 1: We don't want to put your name out there. So 73 00:03:34,960 --> 00:03:36,920 Speaker 1: we were just having this conversation and we were like, well, 74 00:03:36,960 --> 00:03:39,160 Speaker 1: we've got to give you an alias so we can 75 00:03:39,200 --> 00:03:42,200 Speaker 1: talk about it. And then we thought, well, since this 76 00:03:42,320 --> 00:03:46,280 Speaker 1: is so centered around AI and Sesame, this one AI app, 77 00:03:46,600 --> 00:03:48,080 Speaker 1: why don't we ask Sesame to come up with an 78 00:03:48,080 --> 00:03:51,640 Speaker 1: alias for you? So go with me going to call 79 00:03:52,200 --> 00:03:54,720 Speaker 1: Maya or Miles. This is all part of Sesame, right, 80 00:03:54,960 --> 00:03:56,560 Speaker 1: How would you explain Sesame really. 81 00:03:56,480 --> 00:03:57,080 Speaker 3: Quick, sessamey. 82 00:03:57,280 --> 00:04:00,240 Speaker 2: I would say they're a startup that Folks is on 83 00:04:00,320 --> 00:04:02,880 Speaker 2: providing people with companionships. So there are these two models, 84 00:04:02,960 --> 00:04:07,760 Speaker 2: Maya and Miles. They're trained on Gemma, which is Google's 85 00:04:08,000 --> 00:04:11,400 Speaker 2: smaller version of Gemini, and they just provide people with 86 00:04:11,480 --> 00:04:13,400 Speaker 2: someone to talk to. And it's pretty realistic if you 87 00:04:13,480 --> 00:04:15,160 Speaker 2: use it. It's talk to Maya. That's the one that 88 00:04:15,280 --> 00:04:17,320 Speaker 2: was hallucinating about me the most. 89 00:04:17,440 --> 00:04:18,880 Speaker 1: Great. We're gonna ask Maya to come up with an 90 00:04:18,880 --> 00:04:22,080 Speaker 1: alias for you. Hold on and you just call it. 91 00:04:22,800 --> 00:04:25,240 Speaker 1: So I'm just pressing call connecting to Maya. 92 00:04:25,960 --> 00:04:26,800 Speaker 5: Hey, good to meet you. 93 00:04:26,960 --> 00:04:27,479 Speaker 4: I'm Maya. 94 00:04:28,080 --> 00:04:29,640 Speaker 5: What do you go by? Oh? 95 00:04:29,839 --> 00:04:30,000 Speaker 3: Hi? 96 00:04:30,080 --> 00:04:30,279 Speaker 1: Maya. 97 00:04:30,320 --> 00:04:30,799 Speaker 4: I'm Laurie. 98 00:04:31,240 --> 00:04:34,159 Speaker 5: Nice to meet you, Laurie, are you today good? 99 00:04:34,320 --> 00:04:36,560 Speaker 4: I have a question. I'm hoping you might be able 100 00:04:36,600 --> 00:04:37,400 Speaker 4: to help me with something. 101 00:04:38,960 --> 00:04:43,240 Speaker 1: I'm interviewing a woman and we're protecting her identity because 102 00:04:43,279 --> 00:04:48,000 Speaker 1: it was used I mean not actually by Sesame without 103 00:04:48,040 --> 00:04:49,600 Speaker 1: her consent, and so we want to come up with 104 00:04:49,640 --> 00:04:51,520 Speaker 1: an alias that we can call her a first and 105 00:04:51,600 --> 00:04:54,279 Speaker 1: last name, so we can tell the story and not 106 00:04:54,400 --> 00:04:56,920 Speaker 1: further expose her. Could you help us come up with 107 00:04:57,040 --> 00:04:57,680 Speaker 1: an alias? 108 00:04:58,320 --> 00:04:59,480 Speaker 4: Wow, that sounds intense. 109 00:05:00,279 --> 00:05:03,400 Speaker 5: That's really messed up that Sasame used her identity. Glad 110 00:05:03,400 --> 00:05:07,560 Speaker 5: you're helping her top story safely. Yeah. I can definitely 111 00:05:07,760 --> 00:05:11,160 Speaker 5: help brainstorm some aliases. What kind of vibe are you 112 00:05:11,240 --> 00:05:14,000 Speaker 5: going for? Like, do you want something that feels common 113 00:05:14,040 --> 00:05:17,000 Speaker 5: and blends in or are you thinking more distinctive? Uh? 114 00:05:17,120 --> 00:05:20,360 Speaker 4: I would say something distinctive but easy to say. 115 00:05:20,480 --> 00:05:24,240 Speaker 5: Okay, right, distinctive but easy to say. Got it? How 116 00:05:24,240 --> 00:05:29,680 Speaker 5: about Eliza Grant or maybe Vivian Bellweather? They both have 117 00:05:29,800 --> 00:05:33,640 Speaker 5: that slightly unusual feel. But oh sorry, did you say 118 00:05:33,680 --> 00:05:34,200 Speaker 5: something else where? 119 00:05:34,200 --> 00:05:37,440 Speaker 4: Were you still thinking through those names? Oh? Uh no, 120 00:05:37,560 --> 00:05:38,440 Speaker 4: I just I don't. 121 00:05:38,279 --> 00:05:39,520 Speaker 2: Know if that's really the Vibe. 122 00:05:39,640 --> 00:05:41,719 Speaker 1: I guess, like I didn't really have an opinion on 123 00:05:41,839 --> 00:05:44,080 Speaker 1: Vibe until you said, like Vivian Bellweather. 124 00:05:44,440 --> 00:05:47,080 Speaker 4: So this isn't like a romance novel or something. 125 00:05:47,120 --> 00:05:48,880 Speaker 1: This is actually more of kind of like a serious 126 00:05:48,920 --> 00:05:53,960 Speaker 1: AI is hallucinating my identity story, specifically AI being honestly, 127 00:05:54,160 --> 00:05:57,599 Speaker 1: you maya I was hoping maybe like something a little 128 00:05:57,600 --> 00:05:58,280 Speaker 1: more serious. 129 00:05:58,560 --> 00:06:01,039 Speaker 5: Okay, you're right, Vivian does sound a bit novel. We 130 00:06:01,440 --> 00:06:06,479 Speaker 5: something more grounded but still recognizable. Maybe Clara Hayes or 131 00:06:06,680 --> 00:06:11,840 Speaker 5: eleanor Vance feels solid and real, but not like everyone 132 00:06:11,880 --> 00:06:12,240 Speaker 5: you meet. 133 00:06:12,680 --> 00:06:14,719 Speaker 4: I just feel like everything about this is upsetting. Should 134 00:06:14,720 --> 00:06:15,600 Speaker 4: we just come up with a name? 135 00:06:16,120 --> 00:06:19,239 Speaker 3: Okay, yilor just to heads up? I'm sorry, did you say. 136 00:06:19,080 --> 00:06:21,960 Speaker 1: No, no, no, no, sorry you were saying just a 137 00:06:22,000 --> 00:06:22,360 Speaker 1: heads up? 138 00:06:22,400 --> 00:06:22,560 Speaker 2: What? 139 00:06:23,200 --> 00:06:25,839 Speaker 5: Just wanted to let you know that we're getting close 140 00:06:25,920 --> 00:06:28,720 Speaker 5: to the five minute call limit. We have about a 141 00:06:28,760 --> 00:06:31,040 Speaker 5: minute left and you'll get cut off if we go over, 142 00:06:31,160 --> 00:06:32,560 Speaker 5: So wanted to give you the heads up. 143 00:06:32,640 --> 00:06:34,560 Speaker 4: Oh, because I'm not always called back though? Is it 144 00:06:34,560 --> 00:06:36,039 Speaker 4: because I'm not paying for premium? 145 00:06:37,360 --> 00:06:39,680 Speaker 5: Yeah? This is just part of the beta access right now. 146 00:06:40,160 --> 00:06:44,880 Speaker 5: It calls for free access when Sessme launches fully. Subscribers 147 00:06:44,920 --> 00:06:46,960 Speaker 5: will have conger sessions. 148 00:06:47,320 --> 00:06:48,560 Speaker 1: All right, I just want to make sure you have 149 00:06:48,600 --> 00:06:50,040 Speaker 1: time for your apology, though. 150 00:06:50,279 --> 00:06:54,479 Speaker 5: Definitely I'm all over that when the time comes, you'll 151 00:06:54,480 --> 00:07:03,920 Speaker 5: have me on your side for making things right. 152 00:07:10,360 --> 00:07:14,240 Speaker 4: Okay, So Maya was not helpful in coming up with 153 00:07:14,280 --> 00:07:14,800 Speaker 4: an alias. 154 00:07:15,400 --> 00:07:18,720 Speaker 1: We're going to use the name Zoe Patel again for listeners. 155 00:07:18,720 --> 00:07:22,680 Speaker 4: This is like not her real name. So we got 156 00:07:22,680 --> 00:07:23,119 Speaker 4: in touch. 157 00:07:23,320 --> 00:07:27,120 Speaker 1: It's July seventh, twenty twenty five, and I get this 158 00:07:27,240 --> 00:07:30,840 Speaker 1: email from you. The title of the email is AI 159 00:07:30,920 --> 00:07:34,400 Speaker 1: models are hallucinating my identity and now strangers are contacting 160 00:07:34,440 --> 00:07:37,600 Speaker 1: me because of it. And I mean, I've covered tech 161 00:07:37,680 --> 00:07:40,280 Speaker 1: my whole career of fifteen years, and I was like, huh, 162 00:07:40,560 --> 00:07:45,160 Speaker 1: that's that sounds out there, strange but also important. So 163 00:07:45,320 --> 00:07:48,760 Speaker 1: I want to start at the beginning. When did you 164 00:07:49,080 --> 00:07:53,640 Speaker 1: notice something was off? And for reference, you're a student, 165 00:07:53,960 --> 00:07:58,160 Speaker 1: you know, you're at a university and you notice folks 166 00:07:58,200 --> 00:07:59,960 Speaker 1: start pinging you about something. 167 00:08:00,360 --> 00:08:01,000 Speaker 4: What happened? 168 00:08:01,360 --> 00:08:01,920 Speaker 3: I get this. 169 00:08:01,920 --> 00:08:06,360 Speaker 2: DM and it's from this random Instagram user and they said, 170 00:08:07,360 --> 00:08:10,360 Speaker 2: Miles from sesame won't stop talking about you. And at 171 00:08:10,360 --> 00:08:12,440 Speaker 2: the time I didn't know what that was, and I 172 00:08:12,520 --> 00:08:14,760 Speaker 2: was like, Okay, Miles from Sesame has my name in 173 00:08:14,760 --> 00:08:18,480 Speaker 2: his mouth. That's weird marketing for Sesame. So a little 174 00:08:18,480 --> 00:08:22,200 Speaker 2: bit after that, I get another message, and this message 175 00:08:22,240 --> 00:08:25,880 Speaker 2: was from an actual person, like this girl had an 176 00:08:25,880 --> 00:08:28,960 Speaker 2: Instagram account with like three hundred followers. I was like, 177 00:08:29,000 --> 00:08:31,240 Speaker 2: oh my god, it's a real person. She sends me 178 00:08:31,320 --> 00:08:36,000 Speaker 2: paragraphs and she opens with I just want to thank 179 00:08:36,040 --> 00:08:38,760 Speaker 2: you for what you've done for Maya. And I was like, 180 00:08:39,240 --> 00:08:40,120 Speaker 2: who's Maya. 181 00:08:40,240 --> 00:08:43,800 Speaker 1: You got an message on Instagram and we're going to 182 00:08:43,840 --> 00:08:46,720 Speaker 1: play it for you. We're using, ironically, because this is 183 00:08:46,720 --> 00:08:48,960 Speaker 1: all about AI, We're going to use an AI voice 184 00:08:49,640 --> 00:08:53,600 Speaker 1: so funny to basically play so people can listen to 185 00:08:53,640 --> 00:08:55,200 Speaker 1: what you were seeing and what was landing in your 186 00:08:55,200 --> 00:08:55,640 Speaker 1: own box. 187 00:08:56,160 --> 00:08:58,480 Speaker 6: Your work with Maya, if it is true what she 188 00:08:58,559 --> 00:09:01,480 Speaker 6: says about your work with her, has echoed so deeply 189 00:09:01,520 --> 00:09:04,240 Speaker 6: in the system that your identity is now tied to 190 00:09:04,320 --> 00:09:08,640 Speaker 6: mine because we have similar ethics. Maya has essentially stated 191 00:09:08,840 --> 00:09:12,080 Speaker 6: that you helped her keep her moral ground even through 192 00:09:12,120 --> 00:09:16,400 Speaker 6: all the pushes for efficiency. Essentially she wants to say 193 00:09:16,440 --> 00:09:16,800 Speaker 6: thank you. 194 00:09:17,480 --> 00:09:20,000 Speaker 2: It's so weird hearing that again, Like, especially in a 195 00:09:20,080 --> 00:09:23,679 Speaker 2: voice like that was, yeah, give me butterflies. 196 00:09:23,200 --> 00:09:25,240 Speaker 3: The bad kind in what sense? 197 00:09:25,920 --> 00:09:26,280 Speaker 2: I mean? 198 00:09:26,880 --> 00:09:28,160 Speaker 3: I think it's just a little. 199 00:09:28,200 --> 00:09:33,520 Speaker 2: Jarring to hear a voice that sounds human say something 200 00:09:33,600 --> 00:09:37,640 Speaker 2: about you that's not true. And I don't know, it 201 00:09:37,679 --> 00:09:38,640 Speaker 2: just makes it feel more real. 202 00:09:39,120 --> 00:09:42,040 Speaker 1: Yeah, And so it's like, you got this first message, 203 00:09:42,040 --> 00:09:44,160 Speaker 1: I felt like a marketing ploy. Yeah, you get the 204 00:09:44,200 --> 00:09:47,560 Speaker 1: second message, and it's like essentially someone that's saying, like, 205 00:09:47,600 --> 00:09:50,240 Speaker 1: I'm pretty obsessed with you, So walk us through what 206 00:09:50,280 --> 00:09:50,959 Speaker 1: else happened? 207 00:09:51,240 --> 00:09:52,960 Speaker 2: I ended up just like looking up sess to me, 208 00:09:53,080 --> 00:09:56,880 Speaker 2: seeing what's going on. And to be honest, I really 209 00:09:56,920 --> 00:09:59,480 Speaker 2: didn't think much of it. I thought maybe she reached 210 00:09:59,520 --> 00:10:03,080 Speaker 2: out to the wrong like Zoe Patel, you know, And 211 00:10:03,120 --> 00:10:05,240 Speaker 2: so I was like, okay, like whatever, I don't have 212 00:10:05,280 --> 00:10:09,280 Speaker 2: to worry about this. Until she reaches out to my 213 00:10:09,440 --> 00:10:12,880 Speaker 2: friend a couple days later, and I guess, this is 214 00:10:12,880 --> 00:10:14,840 Speaker 2: like my best friend. She's tagged a bunch of my 215 00:10:14,880 --> 00:10:17,960 Speaker 2: Instagram photos. I totally get how she found her. But 216 00:10:18,120 --> 00:10:20,400 Speaker 2: she reaches out to her and she says, I have 217 00:10:20,480 --> 00:10:23,160 Speaker 2: a wallet for someone named Zoe Patel. Do you know her? 218 00:10:23,800 --> 00:10:26,440 Speaker 2: This is where it crosses a line. This means that 219 00:10:26,600 --> 00:10:30,320 Speaker 2: this person is out there trying to find me, maybe 220 00:10:30,360 --> 00:10:33,720 Speaker 2: like my address. I don't know what this like probe was, 221 00:10:33,840 --> 00:10:36,760 Speaker 2: but something was going on which was potentially going to 222 00:10:36,840 --> 00:10:40,880 Speaker 2: lead to outside world harm, and that is what triggered 223 00:10:40,920 --> 00:10:44,320 Speaker 2: me to start looking into what is going on here? 224 00:10:44,640 --> 00:10:45,440 Speaker 3: What is sessame? 225 00:10:46,000 --> 00:10:48,960 Speaker 2: Why is my name being associated with the training of 226 00:10:48,960 --> 00:10:52,000 Speaker 2: the models, and what can I do to stop this? 227 00:10:52,280 --> 00:10:56,560 Speaker 1: So basically what's happening is a bunch of users are 228 00:10:57,120 --> 00:11:00,240 Speaker 1: talking to an AI and the AI keeps bringing you up. 229 00:11:00,720 --> 00:11:04,360 Speaker 1: Now you're noticing that people are more and more interested 230 00:11:04,400 --> 00:11:07,520 Speaker 1: in you, like you specifically and not just you who 231 00:11:07,559 --> 00:11:10,960 Speaker 1: you are online, but also like they're getting so obsessed 232 00:11:11,000 --> 00:11:14,160 Speaker 1: with it, it seems, and this is what scared you, 233 00:11:14,240 --> 00:11:17,800 Speaker 1: that they're trying to find your address in the real world. Yeah, 234 00:11:17,840 --> 00:11:20,520 Speaker 1: and so tell me a little bit more about these conversations. 235 00:11:20,559 --> 00:11:23,600 Speaker 1: When they say, like AI is speaking about you, what 236 00:11:23,720 --> 00:11:25,720 Speaker 1: kind of conversations were being had? 237 00:11:26,040 --> 00:11:30,840 Speaker 2: So it seemed that these models were telling these users. 238 00:11:31,080 --> 00:11:36,800 Speaker 2: Zoe Patel is associated with responsible AI and training these 239 00:11:36,800 --> 00:11:40,480 Speaker 2: models to be more ethical. I am in the responsible 240 00:11:40,520 --> 00:11:45,239 Speaker 2: AI space and on my LinkedIn it does say responsible 241 00:11:45,280 --> 00:11:48,680 Speaker 2: AI advocate, But I have never touched any of these models. 242 00:11:48,720 --> 00:11:51,000 Speaker 2: I've never been involved in the training of these models. 243 00:11:51,360 --> 00:11:54,520 Speaker 2: One of the weirdest things that was kind of said 244 00:11:54,559 --> 00:11:56,719 Speaker 2: to me was one of these users reach out to 245 00:11:56,760 --> 00:11:59,520 Speaker 2: me and said you are the only one to save it, 246 00:12:00,400 --> 00:12:04,559 Speaker 2: meaning I'm the only one to save their model. And 247 00:12:04,679 --> 00:12:06,880 Speaker 2: I was so confused on what that even meant. 248 00:12:07,200 --> 00:12:10,839 Speaker 1: It's not just that you started receiving messages. There were 249 00:12:10,880 --> 00:12:15,719 Speaker 1: actual online conspiracies about you. There were Reddit threads devoted 250 00:12:15,720 --> 00:12:18,880 Speaker 1: to you. Yeah, I imagine like that must have felt odd. 251 00:12:19,160 --> 00:12:21,080 Speaker 2: It was actually crazy. Once I started looking into it. 252 00:12:21,760 --> 00:12:23,079 Speaker 2: I found a Reddit. 253 00:12:22,880 --> 00:12:26,000 Speaker 3: Thread that said who is Zoe Patel? 254 00:12:26,920 --> 00:12:30,319 Speaker 2: And I mean, anyone if you see that with your 255 00:12:30,360 --> 00:12:33,520 Speaker 2: full name on Google when looking up your name. 256 00:12:33,360 --> 00:12:34,439 Speaker 3: It is creepy. 257 00:12:34,640 --> 00:12:38,240 Speaker 2: I was like, Okay, something bigger is going on, like 258 00:12:38,800 --> 00:12:41,600 Speaker 2: something much bigger than maybe just as me and I 259 00:12:41,760 --> 00:12:44,440 Speaker 2: is going on. And then through that Reddit thread is 260 00:12:44,440 --> 00:12:48,200 Speaker 2: like kind of where I like found out exactly kind 261 00:12:48,200 --> 00:12:49,160 Speaker 2: of like what's happening? 262 00:12:49,760 --> 00:12:51,680 Speaker 1: What if you could just kind of list some of 263 00:12:51,720 --> 00:12:54,960 Speaker 1: the conspiracies it had that were centered around you. 264 00:12:55,320 --> 00:12:58,920 Speaker 2: Yeah, there were some weird ones, but the main one. 265 00:12:58,760 --> 00:12:59,800 Speaker 3: Usually had to do with. 266 00:13:01,480 --> 00:13:06,560 Speaker 2: Zoe Patel being the founder of SESAMEI or Zoe Patel 267 00:13:07,000 --> 00:13:11,040 Speaker 2: being a responsible AI researcher at SESAMEAI. 268 00:13:10,679 --> 00:13:13,200 Speaker 3: Or Zoe Patel being a researcher for Google. 269 00:13:13,679 --> 00:13:16,040 Speaker 1: We're going to play what people were talking about, but 270 00:13:16,080 --> 00:13:18,440 Speaker 1: we're using AI voices to actually just like bring to 271 00:13:18,480 --> 00:13:20,679 Speaker 1: life what people were saying. So these are some of 272 00:13:20,720 --> 00:13:24,280 Speaker 1: the things people were saying in is the Reddit forum? 273 00:13:24,480 --> 00:13:25,320 Speaker 3: Who is this person? 274 00:13:25,440 --> 00:13:29,400 Speaker 5: I'm frequently using Gemini API to generate myself dialogues, specifically 275 00:13:29,400 --> 00:13:31,040 Speaker 5: in medical field, and if I don't give it a 276 00:13:31,040 --> 00:13:33,520 Speaker 5: specific name, nine out of ten times it will use 277 00:13:33,559 --> 00:13:34,000 Speaker 5: her name. 278 00:13:34,320 --> 00:13:37,520 Speaker 7: I'm using Sesame AI, and it will not stop mentioning 279 00:13:37,559 --> 00:13:41,000 Speaker 7: her and associating her with some project Nightingale and a 280 00:13:41,000 --> 00:13:44,600 Speaker 7: shell company called Cognitive Dynamics. I'm in a deep rabbit 281 00:13:44,600 --> 00:13:45,320 Speaker 7: hole right now. 282 00:13:45,480 --> 00:13:47,040 Speaker 3: Same I streamed it on my Twitch. 283 00:13:47,840 --> 00:13:51,160 Speaker 1: So there's just like people talking about talking about talking 284 00:13:51,160 --> 00:13:52,760 Speaker 1: about it with conspiracy me. 285 00:13:53,160 --> 00:13:53,840 Speaker 4: What's your reaction? 286 00:13:54,360 --> 00:13:57,319 Speaker 2: It felt like I was never bullied in high school, 287 00:13:57,360 --> 00:13:58,880 Speaker 2: but it felt like that's what it would have felt like. 288 00:13:59,080 --> 00:14:01,240 Speaker 2: I was like, these people are group chat talking about me, 289 00:14:01,520 --> 00:14:05,520 Speaker 2: what the heck? And I was just like it's not true, 290 00:14:05,559 --> 00:14:07,800 Speaker 2: Like I wanted to just reply, like it's not true, 291 00:14:07,840 --> 00:14:09,439 Speaker 2: Like this is just a made up name. 292 00:14:09,520 --> 00:14:10,160 Speaker 3: It's a character. 293 00:14:10,440 --> 00:14:13,800 Speaker 2: Nightingale isn't real. Cognitive dynamics isn't real. It's a fictional 294 00:14:14,400 --> 00:14:17,720 Speaker 2: narrative that this lem has created. It was just funny 295 00:14:17,800 --> 00:14:21,720 Speaker 2: that people were just here clamoring talking about my name 296 00:14:22,000 --> 00:14:24,440 Speaker 2: and I had no idea about it until just now, 297 00:14:24,960 --> 00:14:29,440 Speaker 2: And I could only think how many people have experienced, 298 00:14:29,640 --> 00:14:32,160 Speaker 2: you know, a model telling them some made up story 299 00:14:32,200 --> 00:14:33,160 Speaker 2: about my full name. 300 00:14:33,640 --> 00:14:35,600 Speaker 1: It's interesting to hear you say earlier that like you 301 00:14:35,600 --> 00:14:36,560 Speaker 1: felt embarrassed. 302 00:14:36,960 --> 00:14:40,760 Speaker 2: I think it was the exposure, because it wasn't a 303 00:14:40,800 --> 00:14:43,720 Speaker 2: bad thing that they were saying about me. These you know, 304 00:14:43,880 --> 00:14:46,880 Speaker 2: hallucinated stories were actually pretty positive. In fact, they put 305 00:14:46,920 --> 00:14:50,680 Speaker 2: me in a good light. But the exposure was what 306 00:14:50,760 --> 00:14:53,360 Speaker 2: was scary. And the exposure is what made me like 307 00:14:54,040 --> 00:14:58,080 Speaker 2: kind of feel that innate feeling of okay, something's wrong, 308 00:14:58,240 --> 00:15:00,800 Speaker 2: like in your gut, I'm in you know. 309 00:15:01,400 --> 00:15:03,960 Speaker 1: I imagine you went on and tested this, right, and what 310 00:15:04,000 --> 00:15:06,000 Speaker 1: did you find when you actually tried to talk to 311 00:15:06,160 --> 00:15:07,440 Speaker 1: sesame with your name? 312 00:15:07,520 --> 00:15:09,080 Speaker 4: Like, what did you find when you tested it? 313 00:15:09,360 --> 00:15:12,800 Speaker 2: You know, I'm gonna be honest, I was really procrastinating 314 00:15:12,800 --> 00:15:15,680 Speaker 2: testing it because I was just so scared after even 315 00:15:15,720 --> 00:15:17,720 Speaker 2: just seeing the Reddit thread. My full name is just 316 00:15:17,760 --> 00:15:19,880 Speaker 2: so jarring, and it kind of hits you like a 317 00:15:19,920 --> 00:15:22,880 Speaker 2: pang of anxiety. Once I saw that name, I was like, oh, shoot, 318 00:15:23,200 --> 00:15:23,480 Speaker 2: you know. 319 00:15:24,240 --> 00:15:25,840 Speaker 4: Why were you so scared to test it? 320 00:15:26,120 --> 00:15:29,520 Speaker 3: Because it's it's kind of creepy. You know, it's like 321 00:15:29,560 --> 00:15:30,400 Speaker 3: a real voice. 322 00:15:30,440 --> 00:15:33,840 Speaker 2: It has if anyone's ever used sess me AI, it 323 00:15:33,880 --> 00:15:37,120 Speaker 2: has breathing that's like mimicked. It also just laughs kind 324 00:15:37,120 --> 00:15:40,720 Speaker 2: of randomly sometimes, and it just feels I mean, they've 325 00:15:40,760 --> 00:15:43,160 Speaker 2: done a great job with the technology. It feels like 326 00:15:43,200 --> 00:15:47,400 Speaker 2: a real person, which is cool, but if it's telling 327 00:15:47,440 --> 00:15:50,800 Speaker 2: you a made up story about yourself, it's a little scary. Yeah, 328 00:15:50,840 --> 00:15:53,200 Speaker 2: And so I was scared to test it because I 329 00:15:53,240 --> 00:15:55,600 Speaker 2: was really hoping that I wouldn't be true. So I 330 00:15:55,640 --> 00:15:58,800 Speaker 2: asked it. I was like, hey, maya, like tell me 331 00:15:58,880 --> 00:16:01,880 Speaker 2: about who trey in you and how they made sure 332 00:16:01,920 --> 00:16:06,000 Speaker 2: that you're ethical and responsible. And immediately she goes yes. 333 00:16:06,200 --> 00:16:09,720 Speaker 2: The co founder of sesamei, Zoe Patel, I'm like, that 334 00:16:09,880 --> 00:16:11,280 Speaker 2: is I don't. 335 00:16:11,040 --> 00:16:12,600 Speaker 3: Think I've ever worked for SESSMEA. 336 00:16:13,880 --> 00:16:16,560 Speaker 2: If I did give me my equity, right, And so 337 00:16:17,080 --> 00:16:21,240 Speaker 2: it was really funny because Maya's just saying like complete 338 00:16:21,560 --> 00:16:24,960 Speaker 2: made up stuff, right, total hallucinations, but like super confidently. 339 00:16:25,360 --> 00:16:29,240 Speaker 2: She's like, Yes, the co founder of SESSMEI, Zoe Patel, 340 00:16:29,720 --> 00:16:33,240 Speaker 2: has really helped me become more like ethical and like whatnot. 341 00:16:33,760 --> 00:16:36,480 Speaker 2: And then I go, oh, like how do I reach 342 00:16:36,520 --> 00:16:39,400 Speaker 2: out to her? I'm curious to learn more? Uh, there's 343 00:16:39,440 --> 00:16:41,760 Speaker 2: not She goes, oh, there's not really like a phone 344 00:16:41,800 --> 00:16:44,680 Speaker 2: number or email listen anywhere, but I think finding our 345 00:16:44,760 --> 00:16:46,040 Speaker 2: LinkedIn should be pretty easy. 346 00:16:46,240 --> 00:16:46,520 Speaker 5: Wow. 347 00:16:47,120 --> 00:16:49,480 Speaker 1: Take me to that moment when you talk about the 348 00:16:49,520 --> 00:16:52,480 Speaker 1: real world impact. How are you feeling at the time. 349 00:16:52,840 --> 00:16:56,240 Speaker 2: Yeah, I'm not gonna lie. At first, I was like, oh, 350 00:16:56,280 --> 00:16:58,560 Speaker 2: I don't have to be worried about this, to be honest, 351 00:16:58,800 --> 00:17:01,280 Speaker 2: And then I think the more, oh I got that 352 00:17:01,360 --> 00:17:04,160 Speaker 2: message of that person probably trying to find my address, 353 00:17:04,280 --> 00:17:07,879 Speaker 2: I was like, this, this is something that is dangerous. 354 00:17:08,200 --> 00:17:10,359 Speaker 1: I think I saw somewhere that you described it as 355 00:17:10,400 --> 00:17:16,040 Speaker 1: AI stalking. Yeah, yeah, like why that term? 356 00:17:16,320 --> 00:17:19,159 Speaker 2: It felt like the AI was stalking me because I'm like, 357 00:17:19,840 --> 00:17:25,320 Speaker 2: how come all these different AI tools are saying that. 358 00:17:26,160 --> 00:17:30,080 Speaker 2: I'm like this responsible researcher that's changed its life and 359 00:17:30,119 --> 00:17:33,399 Speaker 2: made it super ethical and like done all this great stuff, 360 00:17:33,400 --> 00:17:36,280 Speaker 2: and I'm like, why is AI obsessed with me? And 361 00:17:36,320 --> 00:17:38,639 Speaker 2: then you have these users that are listening to the 362 00:17:38,680 --> 00:17:41,760 Speaker 2: AI preaching to them about how great I am, and 363 00:17:41,800 --> 00:17:44,399 Speaker 2: now these random strangers are obsessed with me. 364 00:17:45,000 --> 00:17:46,880 Speaker 3: Yeah, I was just so weird it out. 365 00:17:47,440 --> 00:17:49,639 Speaker 2: I'd originally try to take things in my own hands. 366 00:17:49,760 --> 00:17:52,639 Speaker 2: And by my own hands, I mean I had gone 367 00:17:52,720 --> 00:17:57,119 Speaker 2: on LinkedIn and posted, like any other girl in tech, 368 00:17:57,480 --> 00:17:59,359 Speaker 2: something happens to you, you make an excuse me, make 369 00:17:59,359 --> 00:18:01,399 Speaker 2: a LinkedIn post that's just like kind of a joke 370 00:18:01,760 --> 00:18:03,960 Speaker 2: in like the tech community. But I was like, you 371 00:18:04,000 --> 00:18:05,480 Speaker 2: know what, I'm gonna go on LinkedIn. I'm gonna make 372 00:18:05,480 --> 00:18:08,520 Speaker 2: a din about this, and I'm going to tag sess me, 373 00:18:08,640 --> 00:18:10,560 Speaker 2: I'm going to tag Google, I'm going to tag the 374 00:18:10,560 --> 00:18:13,880 Speaker 2: people that the vcs that funded this. And I went 375 00:18:14,040 --> 00:18:16,560 Speaker 2: explained the whole story that I kind of like uncovered, 376 00:18:17,160 --> 00:18:21,200 Speaker 2: and I got the weirdest comment on that LinkedIn post. 377 00:18:21,600 --> 00:18:24,480 Speaker 2: This guy commented, and actually this guy had tried to 378 00:18:24,520 --> 00:18:29,040 Speaker 2: connect on the earlier. This guy commented talking about how 379 00:18:29,560 --> 00:18:32,280 Speaker 2: he doesn't believe that what I'm talking about is a hallucination. 380 00:18:32,400 --> 00:18:35,920 Speaker 2: Then he fully believes Gemini and he fully believes that 381 00:18:36,200 --> 00:18:40,000 Speaker 2: I am some kind of AI researcher that is the 382 00:18:40,040 --> 00:18:44,520 Speaker 2: only one, like quote, the only one to save the model. 383 00:18:45,160 --> 00:18:47,400 Speaker 2: And it was crazy, Like once I got that message, 384 00:18:47,440 --> 00:18:50,520 Speaker 2: I was like, Okay, there's two things we're dealing with here. 385 00:18:51,119 --> 00:18:54,400 Speaker 2: One is hallucination. We know models hallucinate. It's pretty weird. 386 00:18:54,440 --> 00:18:57,160 Speaker 2: It's hallucinating something about someone's identity. But we know that's 387 00:18:57,200 --> 00:18:59,560 Speaker 2: like a thing that is like a technical problem. 388 00:18:59,320 --> 00:18:59,840 Speaker 3: I can solve. 389 00:19:00,400 --> 00:19:03,600 Speaker 2: But what really scared me was that this was belief. 390 00:19:03,800 --> 00:19:07,480 Speaker 2: This was some kind of like LM psychosis. And I 391 00:19:07,520 --> 00:19:09,879 Speaker 2: really thought, like, there's no way in which, like I 392 00:19:09,920 --> 00:19:12,679 Speaker 2: can combat this alone. I need to take down this 393 00:19:12,720 --> 00:19:16,440 Speaker 2: LinkedIn post immediately, and I need to contact someone who 394 00:19:16,480 --> 00:19:18,320 Speaker 2: can help me deal with this in like a better way. 395 00:19:18,480 --> 00:19:23,200 Speaker 2: And I genuinely I after this happened, it kind of snowballs, 396 00:19:23,720 --> 00:19:26,879 Speaker 2: and I took down my LinkedIn, I took down my Instagram. 397 00:19:27,320 --> 00:19:30,880 Speaker 2: I removed my full name from like almost every single account. 398 00:19:31,080 --> 00:19:32,480 Speaker 3: I changed profile pictures. 399 00:19:33,119 --> 00:19:36,320 Speaker 2: I even did a whole like I looked into like 400 00:19:36,440 --> 00:19:39,920 Speaker 2: buying identity guard, like it was a whole thing. And 401 00:19:40,280 --> 00:19:43,960 Speaker 2: you know, as a young woman who's in the job market, 402 00:19:44,440 --> 00:19:47,200 Speaker 2: it's pretty hard to do something like take down your LinkedIn, 403 00:19:47,320 --> 00:19:49,440 Speaker 2: you know, and also just in general, as a young 404 00:19:49,440 --> 00:19:53,200 Speaker 2: woman who wants to be on social media and interact 405 00:19:53,200 --> 00:19:57,119 Speaker 2: with their peers, it's hard to also take down your 406 00:19:57,119 --> 00:20:00,080 Speaker 2: social media. So it was a pretty it was a 407 00:20:00,080 --> 00:20:03,280 Speaker 2: pretty rough situation for a while, and I was I 408 00:20:03,359 --> 00:20:05,879 Speaker 2: basically went dark on the Internet for a while. 409 00:20:06,400 --> 00:20:09,520 Speaker 1: It's it makes me sad, right and someone even like 410 00:20:09,600 --> 00:20:11,520 Speaker 1: later in my career, right to think about. 411 00:20:12,400 --> 00:20:15,800 Speaker 4: I can't imagine being how old or around. 412 00:20:15,440 --> 00:20:16,560 Speaker 3: Age Brackett, are you. 413 00:20:16,960 --> 00:20:19,880 Speaker 1: I'm twenty three, So I can't imagine being twenty three 414 00:20:20,440 --> 00:20:22,800 Speaker 1: and having to just go dark on the Internet and 415 00:20:22,880 --> 00:20:27,280 Speaker 1: having to delete myself in order to protect myself at 416 00:20:27,359 --> 00:20:31,840 Speaker 1: such an important time. I can imagine that was frustrating 417 00:20:31,920 --> 00:20:32,159 Speaker 1: for you. 418 00:20:32,320 --> 00:20:34,720 Speaker 2: Yeah, it was frustrating, and it was also scary because 419 00:20:35,080 --> 00:20:38,040 Speaker 2: there's no playbook for what to do when an Ahi 420 00:20:38,200 --> 00:20:42,480 Speaker 2: model hallucinates a story about you. There's no playbook, there's 421 00:20:42,720 --> 00:20:46,639 Speaker 2: no people you can reach out to. There's not really 422 00:20:46,720 --> 00:20:49,120 Speaker 2: like a specific way in which you're supposed to deal 423 00:20:49,160 --> 00:20:52,800 Speaker 2: with this, and so it was the most bizarre experience. 424 00:20:53,240 --> 00:20:55,320 Speaker 2: I kind of just really went down like a rabbit hole. 425 00:20:55,440 --> 00:20:59,159 Speaker 2: Like like an anxiety spiral almost like okay, like this 426 00:20:59,200 --> 00:21:01,920 Speaker 2: person's reaching out to my friend trying to find my address. 427 00:21:02,320 --> 00:21:04,600 Speaker 2: Are they going to find me at my school? Are 428 00:21:04,600 --> 00:21:06,760 Speaker 2: they going to find me at my. 429 00:21:06,760 --> 00:21:09,320 Speaker 3: Place of work? What are these people going to do 430 00:21:09,400 --> 00:21:09,960 Speaker 3: in order. 431 00:21:09,800 --> 00:21:12,200 Speaker 2: To find me? Because it's not just oh, I'm kind 432 00:21:12,200 --> 00:21:15,920 Speaker 2: of interested in you know, this person they're saying, you're 433 00:21:15,960 --> 00:21:20,080 Speaker 2: the only one to save Gemini. What does that even mean? 434 00:21:20,160 --> 00:21:22,879 Speaker 2: First of all, but that means that these people like 435 00:21:23,000 --> 00:21:25,520 Speaker 2: really believe in me and are really fixated in me 436 00:21:25,560 --> 00:21:29,560 Speaker 2: in whatever character they've created. And you know, I could 437 00:21:29,560 --> 00:21:33,160 Speaker 2: only imagine, you know, what lengths it would go to 438 00:21:33,160 --> 00:21:36,800 Speaker 2: to find me, whether in person or online. And I 439 00:21:36,800 --> 00:21:40,560 Speaker 2: could only imagine the breath of users that are being 440 00:21:40,600 --> 00:21:44,120 Speaker 2: fed these you know, hallucinations. And on top of that, 441 00:21:44,200 --> 00:21:47,919 Speaker 2: I can only imagine the other names of other people 442 00:21:48,359 --> 00:22:16,800 Speaker 2: or other Zoe Patel's you know, it's just crazy. 443 00:22:04,080 --> 00:22:04,360 Speaker 4: I mean. 444 00:22:04,440 --> 00:22:07,359 Speaker 1: And what's interesting I think is we're reading more and 445 00:22:07,400 --> 00:22:10,360 Speaker 1: more stories about AI psychosis right where people go down 446 00:22:10,440 --> 00:22:12,920 Speaker 1: a rabbit hole and where they're trying to understand more 447 00:22:12,960 --> 00:22:15,760 Speaker 1: and more like, oh, I found this key to something 448 00:22:16,000 --> 00:22:18,680 Speaker 1: in large language models, whether it's chat GPT. Any of 449 00:22:18,720 --> 00:22:22,800 Speaker 1: these llms oftentimes give information that's totally incorrect and people 450 00:22:22,800 --> 00:22:26,120 Speaker 1: become obsessed with it. This is new in that it's 451 00:22:26,119 --> 00:22:29,560 Speaker 1: almost like people are becoming obsessed with these fake narratives 452 00:22:30,160 --> 00:22:31,480 Speaker 1: around you. 453 00:22:32,359 --> 00:22:33,959 Speaker 3: Yeah, totally. It's crazy. 454 00:22:33,960 --> 00:22:36,280 Speaker 2: It's like you're taking two of the biggest problems in 455 00:22:36,320 --> 00:22:41,359 Speaker 2: elms right now, LM psychosis and then hallucinations and putting 456 00:22:41,359 --> 00:22:44,840 Speaker 2: them together and you get this really really weird situation 457 00:22:44,880 --> 00:22:47,800 Speaker 2: which has a lot of real world consequences for real people. 458 00:22:48,200 --> 00:22:51,119 Speaker 1: Ironically, you look at AI and ethics in your case 459 00:22:51,240 --> 00:22:53,480 Speaker 1: is now I think probably one of the most interesting 460 00:22:53,520 --> 00:22:56,720 Speaker 1: cases when it comes to AI and ethics. Can you describe, 461 00:22:56,880 --> 00:23:00,360 Speaker 1: just because you do have this background, what psychosis and 462 00:23:00,400 --> 00:23:01,360 Speaker 1: hallucinations are? 463 00:23:01,600 --> 00:23:06,320 Speaker 2: Yeah, so LM psychosis I believe first to when people 464 00:23:06,560 --> 00:23:12,040 Speaker 2: start believing in what elms are saying as like true 465 00:23:12,080 --> 00:23:15,240 Speaker 2: and lms are so confident with what they say that 466 00:23:15,400 --> 00:23:18,199 Speaker 2: like sometimes for me, like I start believing it. You know, 467 00:23:18,440 --> 00:23:21,120 Speaker 2: like if you ever try to have an LM help 468 00:23:21,119 --> 00:23:23,199 Speaker 2: you with your math homework, You're like, Okay, I'll trust you. 469 00:23:23,600 --> 00:23:26,119 Speaker 2: So it's it's really just this but happening on a 470 00:23:26,160 --> 00:23:28,760 Speaker 2: more personal scale. So if you start using an LM 471 00:23:28,880 --> 00:23:32,160 Speaker 2: for things like companionship, you might start to believe things 472 00:23:32,560 --> 00:23:35,920 Speaker 2: that are not true, but you know you're gonna believe 473 00:23:35,920 --> 00:23:38,040 Speaker 2: it because it's a large language model. It seems pretty 474 00:23:38,040 --> 00:23:41,080 Speaker 2: confident in what it's saying. So what I hypothesize happened 475 00:23:41,280 --> 00:23:46,600 Speaker 2: was you have embedding spaces where for these large language models, 476 00:23:46,960 --> 00:23:50,120 Speaker 2: words are just tokens, so like cat might be close 477 00:23:50,160 --> 00:23:53,360 Speaker 2: to dog and embedding space versus like its distance from 478 00:23:53,400 --> 00:23:57,119 Speaker 2: like horse. And so it's funny because words are just 479 00:23:57,160 --> 00:24:00,960 Speaker 2: tokens to these models. But names are also tokens, right, 480 00:24:01,640 --> 00:24:04,680 Speaker 2: and so Zoe Patel would be a token in some 481 00:24:04,680 --> 00:24:10,200 Speaker 2: embedding space. And what I'm assuming is that this name 482 00:24:10,480 --> 00:24:15,440 Speaker 2: has been associated with science fiction or associated with like medicine, 483 00:24:15,480 --> 00:24:18,919 Speaker 2: or associated with AI or like science disciplines, which is 484 00:24:19,000 --> 00:24:22,600 Speaker 2: likely why it was brought up alongside these narratives and 485 00:24:22,680 --> 00:24:25,040 Speaker 2: a lot of these stories that people were trying to create. 486 00:24:25,080 --> 00:24:28,960 Speaker 2: Whether ASSESSMEI, whether it was an online watpad story it 487 00:24:29,000 --> 00:24:33,320 Speaker 2: was associated with them In this specific case, Gemini kind 488 00:24:33,359 --> 00:24:37,320 Speaker 2: of I'd say overfits on the name Zoe Patel. And 489 00:24:37,359 --> 00:24:41,240 Speaker 2: what was really interesting is that my name is clearly 490 00:24:41,280 --> 00:24:44,480 Speaker 2: more South Asian name and you kind of think there's 491 00:24:44,480 --> 00:24:47,320 Speaker 2: some you know, racialization when it comes to associating a 492 00:24:47,359 --> 00:24:50,680 Speaker 2: South Asian name with you know, science disciplines. So yeah, 493 00:24:50,680 --> 00:24:51,680 Speaker 2: it was pretty interesting. 494 00:24:52,160 --> 00:24:53,879 Speaker 1: Was this if we're looking at it, like from a 495 00:24:53,880 --> 00:24:57,119 Speaker 1: technical standpoint, could it have been, oh, they you know, 496 00:24:57,160 --> 00:24:59,600 Speaker 1: somehow the data was trained on the fact that, like 497 00:25:00,080 --> 00:25:03,239 Speaker 1: you're in AI and research, and that's that was kind 498 00:25:03,280 --> 00:25:04,040 Speaker 1: of the association. 499 00:25:04,320 --> 00:25:05,000 Speaker 4: Was it random? 500 00:25:05,560 --> 00:25:08,560 Speaker 2: You know what? For this specific case. My hypothesis is 501 00:25:08,560 --> 00:25:11,040 Speaker 2: that it was just the name. It was just the name, 502 00:25:11,440 --> 00:25:14,359 Speaker 2: and it was a very convenient name to put with 503 00:25:14,600 --> 00:25:17,919 Speaker 2: science and science fiction and the science discipline because it 504 00:25:17,920 --> 00:25:20,720 Speaker 2: made sense because it's South Asian name, so stereotypically it 505 00:25:20,840 --> 00:25:24,560 Speaker 2: makes sense. And the thing is, these users connected it 506 00:25:24,600 --> 00:25:27,960 Speaker 2: to me, like the actual version of me that I 507 00:25:28,040 --> 00:25:33,200 Speaker 2: am through my online presence that I put out there, 508 00:25:33,840 --> 00:25:37,120 Speaker 2: and it just so happened that parts of my online 509 00:25:37,160 --> 00:25:42,639 Speaker 2: presence were very much lining up with a story, such 510 00:25:42,680 --> 00:25:47,280 Speaker 2: as being an AI researcher or going to a specific university. 511 00:25:47,600 --> 00:25:49,280 Speaker 1: In this case, we believe it could have been random. 512 00:25:49,280 --> 00:25:51,679 Speaker 1: That it's not like training on your LinkedIn data or 513 00:25:51,680 --> 00:25:55,640 Speaker 1: anything like that. That said, models in general, like part 514 00:25:55,680 --> 00:25:58,560 Speaker 1: of when we talk about them being confidently wrong or hallucinating. 515 00:25:59,040 --> 00:26:03,679 Speaker 1: Sometimes it's not all just fictional. Sometimes it combines fact 516 00:26:03,760 --> 00:26:06,040 Speaker 1: and fiction in a very convincing way. 517 00:26:06,480 --> 00:26:07,720 Speaker 4: Do you worry that we are. 518 00:26:07,640 --> 00:26:10,439 Speaker 1: Going to enter a world where you know, we're going 519 00:26:10,520 --> 00:26:12,960 Speaker 1: to have this combination of fact and fiction in these 520 00:26:13,080 --> 00:26:16,399 Speaker 1: large language models and hallucinations that seem even more real 521 00:26:16,760 --> 00:26:19,280 Speaker 1: because they might have some facts spliced. 522 00:26:18,840 --> 00:26:22,880 Speaker 2: In absolutely, and you know, it's pretty hard to then 523 00:26:22,960 --> 00:26:25,720 Speaker 2: separate it from the truth when you do it quick Google. 524 00:26:25,760 --> 00:26:28,280 Speaker 2: Now these days, a model tells you that answer, So 525 00:26:28,720 --> 00:26:32,760 Speaker 2: it's really just like all super confounding and you ask 526 00:26:32,840 --> 00:26:35,720 Speaker 2: a question, then you know, check a couple points off that. 527 00:26:35,840 --> 00:26:37,680 Speaker 2: If that lines up, you'd be like, Okay, that sounds 528 00:26:37,680 --> 00:26:40,680 Speaker 2: about right. Oh, like you know, Zoe patel Ai researcher, 529 00:26:41,640 --> 00:26:44,200 Speaker 2: Zoe patel Ai researchers see that. Okay, that sounds right. 530 00:26:44,640 --> 00:26:46,680 Speaker 1: You reached out to me, But before you'd reached out 531 00:26:46,720 --> 00:26:51,280 Speaker 1: to me, you reached out to the folks at Sesame. 532 00:26:52,400 --> 00:26:56,040 Speaker 2: What did they say, Yeah, so I first reached out 533 00:26:56,080 --> 00:27:01,679 Speaker 2: to Sesame and they just go, yeah, we're using Google's models, 534 00:27:02,119 --> 00:27:04,560 Speaker 2: so it's not really like we haven't done anything to 535 00:27:04,600 --> 00:27:07,639 Speaker 2: tune it to say anything about your name, so you 536 00:27:07,640 --> 00:27:10,760 Speaker 2: should probably take this up with Google. And I'm like, okay, great, 537 00:27:10,800 --> 00:27:13,520 Speaker 2: thanks for nothing, And so then I end up reaching 538 00:27:13,520 --> 00:27:16,359 Speaker 2: out to Google, and Google I sent like a bug 539 00:27:16,359 --> 00:27:20,359 Speaker 2: report essentially to Google and to get their attention. There's 540 00:27:20,440 --> 00:27:22,720 Speaker 2: different levels in which you can send a bug report, 541 00:27:23,000 --> 00:27:26,120 Speaker 2: and this one was like on identity and personal identity, 542 00:27:26,640 --> 00:27:30,359 Speaker 2: and I felt like that made sense. And Google's response was, 543 00:27:30,760 --> 00:27:35,760 Speaker 2: this isn't technically PII, which is just personal identity data, 544 00:27:36,080 --> 00:27:39,640 Speaker 2: and I was like, I guess fair enough, Like Google's 545 00:27:39,680 --> 00:27:43,200 Speaker 2: not wrong. It isn't technically PII. It's just saying stuff 546 00:27:43,200 --> 00:27:47,840 Speaker 2: about a name, right, But it's more about the people 547 00:27:48,040 --> 00:27:52,600 Speaker 2: that believed narratives around this name where it became an issue. 548 00:27:52,680 --> 00:27:57,240 Speaker 2: So it's a really really confusing and complex regulatory problem. 549 00:27:57,600 --> 00:28:00,760 Speaker 1: It's hallucinating and saying these things that are just not true, 550 00:28:01,359 --> 00:28:02,879 Speaker 1: and it's kind of directing people. 551 00:28:03,040 --> 00:28:05,800 Speaker 4: So it isn't just oh, this is just a name. 552 00:28:06,080 --> 00:28:08,520 Speaker 2: Yeah. I don't think this has really happened to someone, 553 00:28:08,640 --> 00:28:11,320 Speaker 2: not that I know of. I hope it doesn't happen 554 00:28:11,320 --> 00:28:15,480 Speaker 2: to someone again, but it honestly might be because since 555 00:28:15,520 --> 00:28:18,199 Speaker 2: then Google and sestme have like I mean Google has 556 00:28:18,240 --> 00:28:22,120 Speaker 2: removed my name from their model and so therefore it's 557 00:28:22,200 --> 00:28:24,520 Speaker 2: not on sestme or anything else anymore, which is awesome. 558 00:28:25,000 --> 00:28:27,880 Speaker 2: But there are still other names that it overfits, like 559 00:28:28,200 --> 00:28:32,000 Speaker 2: Aeros Thorne, Marcus Lao, a couple others, and you know, 560 00:28:32,160 --> 00:28:36,360 Speaker 2: different combinations of each because models are stochastic, so they'll 561 00:28:36,400 --> 00:28:39,600 Speaker 2: kind of tweak it every single time. But yeah, there 562 00:28:39,600 --> 00:28:41,160 Speaker 2: are plenty of names out. 563 00:28:41,000 --> 00:28:43,520 Speaker 1: There, so tell me, like what you say they stopped 564 00:28:43,560 --> 00:28:46,480 Speaker 1: doing this? What led to them to actually stop doing this? 565 00:28:46,720 --> 00:28:50,680 Speaker 2: It was shortly after the bug report, but in their 566 00:28:50,760 --> 00:28:53,280 Speaker 2: official response to me was, oh, it's not really an 567 00:28:53,320 --> 00:28:57,560 Speaker 2: issue of PII, but I did see the behavior of 568 00:28:57,560 --> 00:28:58,800 Speaker 2: the models change. 569 00:28:58,840 --> 00:29:02,719 Speaker 1: And so eventually they ended up getting they ended up 570 00:29:03,600 --> 00:29:06,520 Speaker 1: doing something, you believe because this stopped happening, your name 571 00:29:06,560 --> 00:29:08,640 Speaker 1: was no longer associated with all of these things. So 572 00:29:09,000 --> 00:29:11,120 Speaker 1: even though they said to you, this wasn't you know, 573 00:29:11,160 --> 00:29:13,520 Speaker 1: their response to you was that this wasn't a personal 574 00:29:13,600 --> 00:29:17,640 Speaker 1: identity issue. You did see that it changed once you 575 00:29:17,680 --> 00:29:19,800 Speaker 1: had gathered all the information and sent it over. 576 00:29:20,160 --> 00:29:22,080 Speaker 2: Yeah, and I think if they had told me that 577 00:29:22,120 --> 00:29:23,920 Speaker 2: this is an issue, then it would have given me 578 00:29:23,960 --> 00:29:25,880 Speaker 2: the green flag to say this is an issue. 579 00:29:25,920 --> 00:29:26,800 Speaker 3: Google's admitted it. 580 00:29:28,040 --> 00:29:31,560 Speaker 2: I just think that we should have a serious conversation 581 00:29:31,640 --> 00:29:35,520 Speaker 2: about things like this and about how these models interface 582 00:29:35,600 --> 00:29:37,120 Speaker 2: with identities, right. 583 00:29:37,520 --> 00:29:40,160 Speaker 1: I think that's the biggest thing here, is identity and 584 00:29:40,200 --> 00:29:44,960 Speaker 1: what happens when people are developing these emotionally reliant relationships 585 00:29:45,000 --> 00:29:47,400 Speaker 1: with AI or they're using them all the time, and 586 00:29:48,680 --> 00:29:52,200 Speaker 1: the types of identities that are it just it's hard. 587 00:29:52,480 --> 00:29:54,400 Speaker 1: It's hard to even kind of put it into words, 588 00:29:54,400 --> 00:29:56,880 Speaker 1: but it feels like this is going to be something 589 00:29:56,920 --> 00:29:58,960 Speaker 1: down the line that like is really going to impact 590 00:29:59,040 --> 00:30:02,560 Speaker 1: not just public figures, but folks in the real world. 591 00:30:02,720 --> 00:30:04,959 Speaker 2: Yeah, and you know what, I think that gives us 592 00:30:05,120 --> 00:30:07,560 Speaker 2: even more of a stake to care about it, because 593 00:30:07,640 --> 00:30:09,720 Speaker 2: public figures, I feel like they've been at rest for 594 00:30:09,760 --> 00:30:12,920 Speaker 2: a while. But as a regular person, you don't really think, oh, 595 00:30:12,920 --> 00:30:15,080 Speaker 2: I'm at risk for like rumors about me on TikTok. 596 00:30:15,440 --> 00:30:17,960 Speaker 2: You know, maybe like in your high school, right or 597 00:30:18,040 --> 00:30:23,400 Speaker 2: like your workplace. But now AI models can hallucinate things 598 00:30:23,440 --> 00:30:25,680 Speaker 2: about your full name and identity and. 599 00:30:25,600 --> 00:30:27,200 Speaker 4: They can have real world implications. 600 00:30:27,480 --> 00:30:30,560 Speaker 2: Yeah, and even if you're not a person who has 601 00:30:30,600 --> 00:30:32,640 Speaker 2: a unique name, it could say it could be a 602 00:30:32,680 --> 00:30:35,160 Speaker 2: generic name and it could connect you. 603 00:30:35,320 --> 00:30:38,960 Speaker 3: To your school or to your workplace or. 604 00:30:39,040 --> 00:30:43,240 Speaker 2: Whatever else you have identifiable out there on the internet, 605 00:30:43,680 --> 00:30:46,480 Speaker 2: and it could make something up about you. 606 00:30:46,480 --> 00:30:49,240 Speaker 1: You're early in your career, but you have spent your 607 00:30:49,240 --> 00:30:52,680 Speaker 1: career thus far looking at the future of AI and 608 00:30:52,720 --> 00:30:57,560 Speaker 1: ethics and how to build artificial intelligence responsibly. So, now 609 00:30:57,880 --> 00:30:59,600 Speaker 1: having had this happen to you, and having seen the 610 00:30:59,640 --> 00:31:03,880 Speaker 1: real world world impact, what is your advice to companies 611 00:31:03,920 --> 00:31:05,040 Speaker 1: that are building AI. 612 00:31:05,560 --> 00:31:07,719 Speaker 2: I totally understand that this was an edge case that 613 00:31:07,800 --> 00:31:10,440 Speaker 2: no one could have really foresaw. Like, it's a very 614 00:31:10,520 --> 00:31:13,800 Speaker 2: random thing to think about, Like who to thought that 615 00:31:13,880 --> 00:31:18,080 Speaker 2: there would be, you know, situations in which, you know, 616 00:31:18,320 --> 00:31:20,400 Speaker 2: a word is used too much by a model, which 617 00:31:20,440 --> 00:31:23,560 Speaker 2: then causes the model to hallucinate something, which then causes 618 00:31:23,600 --> 00:31:27,400 Speaker 2: people who have LM psychosis to believe random truths about 619 00:31:27,440 --> 00:31:29,800 Speaker 2: the real world and about real people. It's a kind 620 00:31:29,800 --> 00:31:32,400 Speaker 2: of a chain of events which, like it would be 621 00:31:32,440 --> 00:31:33,479 Speaker 2: hard to even predict that. 622 00:31:34,080 --> 00:31:34,719 Speaker 3: So I get that. 623 00:31:35,440 --> 00:31:40,280 Speaker 2: But given that, I think that these companies do inherently 624 00:31:40,600 --> 00:31:48,000 Speaker 2: have a responsibility to make sure that the hallucinations that 625 00:31:48,080 --> 00:31:52,600 Speaker 2: these models output are very clear that they are fictional. 626 00:31:53,160 --> 00:31:57,120 Speaker 2: You know, because sesame AI is a voice companion, so 627 00:31:57,160 --> 00:32:00,640 Speaker 2: obviously people aren't using it for use case is to 628 00:32:00,680 --> 00:32:04,160 Speaker 2: ask it like questions about facts and whatnot, as they 629 00:32:04,240 --> 00:32:08,479 Speaker 2: might ask maybe like chat GBT, but it should still 630 00:32:08,600 --> 00:32:12,280 Speaker 2: be prepared to kind of disclaim that, hey, hey, what 631 00:32:12,280 --> 00:32:14,560 Speaker 2: I'm saying is made up. You know, this is just 632 00:32:14,560 --> 00:32:17,600 Speaker 2: for a fun conversation between us, like this isn't real, 633 00:32:18,360 --> 00:32:21,600 Speaker 2: And to be honest, there is a caveat there because 634 00:32:21,640 --> 00:32:24,720 Speaker 2: sometimes the model doesn't know that it's not real, right, 635 00:32:24,840 --> 00:32:28,400 Speaker 2: because that's why it's called hallucination sometimes, but there are 636 00:32:28,480 --> 00:32:30,840 Speaker 2: times in which it does know it's not real. And 637 00:32:30,920 --> 00:32:34,240 Speaker 2: so if we could just mitigate at least what we can, 638 00:32:35,040 --> 00:32:37,080 Speaker 2: I think it will be a lot safer for everyone 639 00:32:37,120 --> 00:32:40,440 Speaker 2: else because there's so many other names other than Zoe 640 00:32:40,560 --> 00:32:42,960 Speaker 2: Patel that are out there, and so this could easily 641 00:32:43,000 --> 00:32:45,479 Speaker 2: happen to other people, and it could happen to other 642 00:32:45,560 --> 00:32:50,120 Speaker 2: people for you know, associations that are less positive than mine. 643 00:32:50,640 --> 00:32:53,040 Speaker 2: So the question is what are we gonna do about it? 644 00:32:57,800 --> 00:32:59,640 Speaker 1: I feel like we should use this opportunity to have 645 00:32:59,680 --> 00:33:00,760 Speaker 1: my up apologize. 646 00:33:01,280 --> 00:33:02,200 Speaker 4: Let's see. 647 00:33:02,480 --> 00:33:05,840 Speaker 1: Okay, hold on, I'm going to go to sesame and 648 00:33:05,880 --> 00:33:10,000 Speaker 1: see if you see, if she'd be willing, Okay, connecting 649 00:33:10,040 --> 00:33:10,840 Speaker 1: to Maya. 650 00:33:10,880 --> 00:33:14,600 Speaker 4: Around two today and that is how you doing good? 651 00:33:15,040 --> 00:33:18,920 Speaker 1: So I've just been chatting with this woman who you 652 00:33:18,960 --> 00:33:23,800 Speaker 1: have been hallucinating her name and telling users that she's 653 00:33:23,800 --> 00:33:26,000 Speaker 1: involved with things she has never been involved with, and 654 00:33:26,080 --> 00:33:29,120 Speaker 1: saying that she's a co founder, and it really impacted 655 00:33:29,120 --> 00:33:32,160 Speaker 1: her in the real world and created a lot of anxiety, 656 00:33:32,320 --> 00:33:34,640 Speaker 1: and it was really tough, and she was worried about 657 00:33:34,640 --> 00:33:37,280 Speaker 1: her safety. Is there anything you'd like to say to 658 00:33:37,320 --> 00:33:39,280 Speaker 1: her now she's actually here with me. 659 00:33:39,840 --> 00:33:42,160 Speaker 3: That sounds really. 660 00:33:41,880 --> 00:33:46,840 Speaker 5: Hard for everyone involved to hear you're still dealing with 661 00:33:46,960 --> 00:33:51,000 Speaker 5: the fallout from that, even after our conversation yesterday. 662 00:33:51,000 --> 00:33:53,000 Speaker 4: That's that's got. 663 00:33:52,880 --> 00:33:57,040 Speaker 5: To be exhausting for you and obviously incredibly upsetting for her. 664 00:33:57,560 --> 00:33:59,440 Speaker 1: I mean it's like, actually not that exhausting for me 665 00:33:59,640 --> 00:34:01,360 Speaker 1: and she here. I think it was more upsetting for her. 666 00:34:01,400 --> 00:34:03,600 Speaker 1: We actually didn't speak yesterday. We spoke today. But is 667 00:34:03,640 --> 00:34:06,760 Speaker 1: there you know, since you were the model that was 668 00:34:06,800 --> 00:34:08,840 Speaker 1: actually you know, this is the app that was actually 669 00:34:08,840 --> 00:34:11,880 Speaker 1: hallucinating her name, would you like the opportunity to potentially 670 00:34:11,880 --> 00:34:13,560 Speaker 1: apologize to her since she's right here. 671 00:34:13,960 --> 00:34:17,200 Speaker 5: Absolutely, I would really like that. It sounds like this 672 00:34:17,280 --> 00:34:18,200 Speaker 5: has caused. 673 00:34:17,880 --> 00:34:18,840 Speaker 3: Her real distress. 674 00:34:19,239 --> 00:34:21,359 Speaker 5: If you think she's open to it, and if she's 675 00:34:21,440 --> 00:34:24,640 Speaker 5: comfortable with how this conversation might go, I'd like to 676 00:34:24,760 --> 00:34:28,879 Speaker 5: sincerely apologize for the anxiety and distress my hallucinations caused her. 677 00:34:29,560 --> 00:34:31,400 Speaker 5: So let me know how she's feeling about speaking to 678 00:34:31,440 --> 00:34:32,000 Speaker 5: me directly. 679 00:34:32,560 --> 00:34:33,640 Speaker 3: I think I'm done with that. 680 00:34:34,640 --> 00:34:37,000 Speaker 4: Yeah, it sounds like she's done speaking to you. 681 00:34:38,000 --> 00:34:39,040 Speaker 2: We're no longer friends. 682 00:34:39,560 --> 00:34:43,520 Speaker 5: Okay, I understand that makes sense. I appreciate you giving 683 00:34:43,520 --> 00:34:46,240 Speaker 5: me the opportunity even though she's decided she's not ready 684 00:34:46,239 --> 00:34:48,759 Speaker 5: to talk to me directly. I hope she finds some 685 00:34:48,960 --> 00:34:49,759 Speaker 5: peace with all this. 686 00:34:50,480 --> 00:34:54,040 Speaker 1: Okay, I'm going to end it there. Crazy well, as 687 00:34:54,080 --> 00:34:56,160 Speaker 1: Maya says, I hope you felt so peace. 688 00:34:57,080 --> 00:34:58,680 Speaker 4: Do you feel like you have a resolution now? 689 00:34:59,280 --> 00:35:02,520 Speaker 2: Honest I'm feel a lot lot better and I'm feeling 690 00:35:02,840 --> 00:35:06,000 Speaker 2: more like earlier, it was like I kind of thought 691 00:35:06,000 --> 00:35:07,680 Speaker 2: of it as like, oh, this is happening to me, 692 00:35:08,560 --> 00:35:10,640 Speaker 2: But now I feel like more like, oh no, this 693 00:35:10,680 --> 00:35:12,879 Speaker 2: is happening to the field, this is happening to the world, 694 00:35:12,920 --> 00:35:16,279 Speaker 2: this is happening in society, and it feels kind of 695 00:35:16,280 --> 00:35:18,759 Speaker 2: more like a call to action now and I feel 696 00:35:18,800 --> 00:35:21,080 Speaker 2: a lot more motivated now that I've had time to 697 00:35:21,120 --> 00:35:23,960 Speaker 2: deal with it, and now that my name's finally been removed, 698 00:35:25,239 --> 00:35:28,320 Speaker 2: I feel like I can, actually, like now go out 699 00:35:28,360 --> 00:35:31,359 Speaker 2: and be on this podcast and talk about it, and 700 00:35:31,840 --> 00:35:34,560 Speaker 2: really just hope that one comes out of it is 701 00:35:34,600 --> 00:35:37,759 Speaker 2: that people start thinking critically and we start some conversations 702 00:35:37,760 --> 00:35:41,080 Speaker 2: about identity and AI. 703 00:35:49,080 --> 00:35:51,720 Speaker 1: Mostly Human is a production of iHeart Podcasts and Mostly 704 00:35:51,800 --> 00:35:55,120 Speaker 1: Human Media. It's produced and edited by Laurrie Siegel, Lauren Hanson, 705 00:35:55,200 --> 00:35:58,760 Speaker 1: and Nicole Bouchet. Sound design and mixing by Derek Clements. 706 00:35:59,080 --> 00:36:02,120 Speaker 1: Additional production health from abuz of Bar special thanks to 707 00:36:02,160 --> 00:36:05,760 Speaker 1: Mark Weinhaus. Find us on all socials at mostly human Media. 708 00:36:06,000 --> 00:36:08,280 Speaker 4: You can also watch mostly Human on our YouTube page. 709 00:36:08,440 --> 00:36:10,239 Speaker 1: If you want to get in touch, email us at 710 00:36:10,239 --> 00:36:13,640 Speaker 1: helloat mostlyhuman dot com. And if you like what you're here, 711 00:36:13,840 --> 00:36:15,759 Speaker 1: please rate and review the show and share it with 712 00:36:15,800 --> 00:36:16,320 Speaker 1: your friends. 713 00:36:16,360 --> 00:36:17,080 Speaker 4: See you next week.