1 00:00:01,488 --> 00:00:04,314 Reed, welcome to The Road to Accountable AI. 2 00:00:05,100 --> 00:00:06,406 Thanks for having me. 3 00:00:07,856 --> 00:00:12,436 Tell me a little bit about how you got into this field of AI ethics, because 4 00:00:12,436 --> 00:00:14,952 relatively speaking, you've been doing it for quite some time. 5 00:00:15,246 --> 00:00:16,536 That's good. 6 00:00:16,536 --> 00:00:18,346 Relatively speaking, I've been doing it forever. 7 00:00:19,226 --> 00:00:22,946 Yeah, I mean, there's sort of a grander story, I suppose, but to sort of keep it 8 00:00:22,946 --> 00:00:26,306 brief, I was a philosophy professor specializing in ethics. 9 00:00:26,306 --> 00:00:30,706 I had the idea for an ethics consulting company many, many years ago. 10 00:00:30,706 --> 00:00:33,286 I didn't see any kind of market for it. 11 00:00:33,546 --> 00:00:35,206 Fast forward, I heard some things. 12 00:00:35,206 --> 00:00:37,306 I thought, okay, maybe there's a market for this actually. 13 00:00:37,306 --> 00:00:40,846 And so I left academia and I started an ethics consulting company. 14 00:00:40,846 --> 00:00:45,626 with the intention of making it focus on emerging technologies and perhaps AI in 15 00:00:45,626 --> 00:00:46,146 particular. 16 00:00:46,146 --> 00:00:49,546 I didn't know how quickly the AI space would grow. 17 00:00:49,546 --> 00:00:51,006 Thankfully, it grew quite quickly. 18 00:00:51,006 --> 00:00:58,766 And so I started advising and consulting on how to do AI responsibly, ethically, at 19 00:00:58,766 --> 00:01:00,386 least not negligently. 20 00:01:00,386 --> 00:01:01,986 That's the goal here. 21 00:01:02,768 --> 00:01:07,296 What does the academic training philosophy bring you in this kind of applied field? 22 00:01:07,534 --> 00:01:10,474 Well, before people called it responsible AI, they called it AI ethics. 23 00:01:10,474 --> 00:01:13,954 We're talking about the ethical implications of AI, ethical, reputational, 24 00:01:13,954 --> 00:01:15,434 regulatory, and legal risks. 25 00:01:15,434 --> 00:01:18,314 If you want to know something about AI ethics, you better know something about 26 00:01:18,314 --> 00:01:20,374 AI, and you better know something about ethics. 27 00:01:20,374 --> 00:01:21,604 And I know a bunch about ethics. 28 00:01:21,604 --> 00:01:25,604 That's what the PhD and the 10 years of being a professor and the whatever it was, 29 00:01:25,604 --> 00:01:28,464 15 plus years of academic research and publishing do. 30 00:01:28,464 --> 00:01:33,694 So when people make certain kinds of claims about responsible AI or AI ethics, 31 00:01:33,694 --> 00:01:37,358 frankly, a lot of times they just don't have a good grip on the ethics of things. 32 00:01:37,358 --> 00:01:40,328 And so they say things that are false, misleading will lead them down wrong 33 00:01:40,328 --> 00:01:41,298 paths. 34 00:01:41,338 --> 00:01:45,818 And so having that kind of experience, understanding, frankly, expertise under my 35 00:01:45,818 --> 00:01:48,928 belt means that I can navigate the conversation really well. 36 00:01:48,928 --> 00:01:52,778 I'll give you one quick story to sort of put some meat on the bones of that. 37 00:01:52,778 --> 00:01:57,818 When I very, when I was really just early transitioning from academic life to my own 38 00:01:57,818 --> 00:02:03,086 company, I joined IEEE, that's the standardizing organization for engineers. 39 00:02:03,086 --> 00:02:06,036 because they were working on this project called Ethically Aligned Design. 40 00:02:06,036 --> 00:02:09,886 And it was about how do we align, what do they call it? 41 00:02:09,886 --> 00:02:13,826 Autonomous and intelligent systems with our human values or something along those 42 00:02:13,826 --> 00:02:14,166 lines. 43 00:02:14,166 --> 00:02:18,806 And I got myself onto this committee that was in charge of figuring out what are the 44 00:02:18,806 --> 00:02:23,356 guidelines or what's the advice around research and design of AI from an ethics 45 00:02:23,356 --> 00:02:23,946 perspective. 46 00:02:23,946 --> 00:02:28,066 And I got on the first call and I was anticipating not being able to say 47 00:02:28,066 --> 00:02:30,546 anything because I thought these are a bunch of engineers. 48 00:02:30,546 --> 00:02:31,726 They, I don't know. 49 00:02:31,726 --> 00:02:32,926 I don't really know anything about AI. 50 00:02:32,926 --> 00:02:33,676 I'm not an engineer. 51 00:02:33,676 --> 00:02:34,656 I'm not a computer scientist. 52 00:02:34,656 --> 00:02:35,746 I'm not a mathematician. 53 00:02:35,746 --> 00:02:36,446 I'm an ethicist. 54 00:02:36,446 --> 00:02:41,516 So I expected it to be a sort of meeting in which I just sat back and listened to 55 00:02:41,516 --> 00:02:42,766 the experts. 56 00:02:42,766 --> 00:02:49,026 What I learned within the first few minutes is, oh, this isn't some advanced 57 00:02:49,026 --> 00:02:50,446 engineering discussion. 58 00:02:50,446 --> 00:02:55,406 This is first semester, first year students talking about ethics. 59 00:02:55,406 --> 00:02:56,296 They don't know anything. 60 00:02:56,296 --> 00:02:57,526 They're engineers. 61 00:02:57,706 --> 00:02:58,534 And... 62 00:02:58,830 --> 00:03:01,950 What wound up happening to my sort of amazement is I wound up essentially 63 00:03:01,950 --> 00:03:06,040 leading the entire conversation, just leading because they were just saying 64 00:03:06,040 --> 00:03:10,330 things that, you know, my first year, first semester undergraduates used to say 65 00:03:10,330 --> 00:03:12,710 about ethics and they were just fundamentally confused. 66 00:03:12,710 --> 00:03:16,900 And so I went in there and I sort of led the conversation and had to clear up a 67 00:03:16,900 --> 00:03:20,970 number of confusions, answer a bunch of questions, direct the discussion in way 68 00:03:20,970 --> 00:03:22,190 that would actually be fruitful. 69 00:03:22,190 --> 00:03:26,080 And so it turns out that it's really the ethics expertise that's more useful than 70 00:03:26,080 --> 00:03:28,176 the AI expertise in a lot of these conversations. 71 00:03:28,176 --> 00:03:32,086 Well, you know, not necessarily from that specific example, but what are some 72 00:03:32,086 --> 00:03:36,152 misunderstandings that people tend to have if they don't have some real background? 73 00:03:36,974 --> 00:03:40,624 Well, one would explain, I think, the transition from talking about AI ethics as 74 00:03:40,624 --> 00:03:43,894 was the standard to talking about responsible AI as has become the standard 75 00:03:43,894 --> 00:03:44,814 among corporations. 76 00:03:44,814 --> 00:03:47,414 They just don't know what the heck to do with ethics. 77 00:03:47,414 --> 00:03:51,934 The number one misconception, which I opened my book, Ethical Machines, with 78 00:03:51,934 --> 00:03:54,804 this is they just think it's squishy, it's subjective, it's just a matter of opinion, 79 00:03:54,804 --> 00:03:56,454 it's how you feel about stuff. 80 00:03:56,454 --> 00:04:01,074 And if you think of ethics as just how you feel about stuff, you know, I like 81 00:04:01,074 --> 00:04:03,074 chocolate, you prefer vanilla. 82 00:04:03,094 --> 00:04:05,518 I like genocide, you prefer peace. 83 00:04:05,518 --> 00:04:10,068 If we think of it along those ways, then of course you're not going to get your way 84 00:04:10,068 --> 00:04:13,578 to actionable steps to identifying and mitigating the risks. 85 00:04:13,578 --> 00:04:17,718 So unless you think about ethics as something that's more just touchy -feely 86 00:04:17,718 --> 00:04:22,998 stuff, you're not going to actually get your way to designing a framework, 87 00:04:22,998 --> 00:04:28,778 designing policy, designing procedures, designing metrics, designing training. 88 00:04:28,778 --> 00:04:31,648 Otherwise you just tell everyone to consult how they feel about stuff. 89 00:04:31,648 --> 00:04:33,438 That's not, that's silly. 90 00:04:33,438 --> 00:04:34,158 And so, 91 00:04:34,158 --> 00:04:38,108 Getting rid of that misconception, articulating that ethics is or at least 92 00:04:38,108 --> 00:04:42,428 can be conceived of as being perfectly objective, and that there are ways of 93 00:04:42,428 --> 00:04:47,068 doing ethical inquiry that are responsible versus irresponsible, that there are ways 94 00:04:47,068 --> 00:04:51,608 of articulating it in concrete ways that lead to action, that's a phenomenally 95 00:04:51,608 --> 00:04:54,852 useful perspective that most people simply don't have. 96 00:04:55,600 --> 00:04:58,240 Let me just play devil's advocate a little bit more, though. 97 00:04:58,240 --> 00:05:01,200 People talk about responsible AI, but nowadays people also talk about things 98 00:05:01,200 --> 00:05:06,990 like AI governance or AI risk management or AI safety, which are all terms that 99 00:05:06,990 --> 00:05:10,000 don't necessarily have that kind of ethical loading. 100 00:05:10,580 --> 00:05:12,470 Do you think that's misguided? 101 00:05:12,470 --> 00:05:15,386 Should we still be talking about ethics or are these different things? 102 00:05:16,078 --> 00:05:20,428 So I've argued, this is a point I make in my book, that we shouldn't use the term 103 00:05:20,428 --> 00:05:24,678 responsible AI, that the term responsible AI is a term that was invented by 104 00:05:24,678 --> 00:05:27,618 corporates essentially, so they don't have to say the word ethics because they don't 105 00:05:27,618 --> 00:05:28,598 want to deal with it. 106 00:05:28,598 --> 00:05:33,698 Incidentally, this is in line with some advice that I got early on when I was 107 00:05:33,698 --> 00:05:34,518 still a professor. 108 00:05:34,518 --> 00:05:40,278 I was talking to an executive at Starbucks and she said, not unreasonably, you might 109 00:05:40,278 --> 00:05:41,858 not want to say that you're doing ethics consulting. 110 00:05:41,858 --> 00:05:43,378 People don't like the word ethics. 111 00:05:43,378 --> 00:05:45,014 They don't get it. 112 00:05:45,198 --> 00:05:47,998 And I said, no, no, I'm just going to own the word ethics then. 113 00:05:47,998 --> 00:05:50,098 If that's the answer, I'm just going to own the word ethics. 114 00:05:50,098 --> 00:05:53,798 And I changed it incidentally to ethical risk because I thought risk was a language 115 00:05:53,798 --> 00:05:55,398 that businesses understand. 116 00:05:55,458 --> 00:06:02,288 Now, look, there's no principled reason to oppose the term responsible AI. 117 00:06:02,288 --> 00:06:07,418 It's not as though that's completely bonkers from a ethical or a conceptual 118 00:06:07,418 --> 00:06:07,898 perspective. 119 00:06:07,898 --> 00:06:08,898 That's reasonable. 120 00:06:08,898 --> 00:06:12,038 I think though, this has been my experience that... 121 00:06:12,078 --> 00:06:16,238 What happens when people use the term responsible AI in practice is that they 122 00:06:16,238 --> 00:06:20,198 throw in ethics -y stuff into the responsible AI bucket, but then they throw 123 00:06:20,198 --> 00:06:21,398 a bunch of other stuff in as well. 124 00:06:21,398 --> 00:06:25,788 So they'll say, yes, responsible AI is ethics and safety and regulatory 125 00:06:25,788 --> 00:06:29,238 compliance and engineering excellence and cybersecurity. 126 00:06:29,238 --> 00:06:31,278 It's just all the stuff. 127 00:06:31,578 --> 00:06:35,178 And then what happens is they say, okay, now let's address the responsible AI 128 00:06:35,178 --> 00:06:36,098 bucket. 129 00:06:36,298 --> 00:06:40,526 And then the experts in the room are experts on things like... 130 00:06:40,526 --> 00:06:43,906 engineering excellence and cyber security and regulatory compliance. 131 00:06:43,906 --> 00:06:45,406 And so what do they do? 132 00:06:45,406 --> 00:06:51,026 They prioritize all of those things to the exclusion of talking about the ethics 133 00:06:51,026 --> 00:06:51,606 stuff. 134 00:06:51,606 --> 00:06:55,606 And so I think that they miss out on huge swaths of risk because they've decided 135 00:06:55,606 --> 00:06:58,476 let's just, you know, we're focused on responsibility now and these things within 136 00:06:58,476 --> 00:06:59,306 those buckets. 137 00:06:59,306 --> 00:07:03,026 And so I think again, while in principle there's nothing wrong with it, 138 00:07:03,026 --> 00:07:07,656 pragmatically or in practice, what happens is that ethical risks get pushed to the 139 00:07:07,656 --> 00:07:08,544 sidelines. 140 00:07:09,328 --> 00:07:10,128 Yeah, it's fascinating. 141 00:07:10,128 --> 00:07:14,588 I mean, I'm a professor of legal studies and business ethics, and we have for a 142 00:07:14,588 --> 00:07:18,588 long time had a lot of the same conversations more broadly in terms of, 143 00:07:18,588 --> 00:07:19,588 you know, what does it mean? 144 00:07:19,588 --> 00:07:21,108 How should we talk about it? 145 00:07:21,108 --> 00:07:24,268 You know, what's more effective practically in terms of getting businesses 146 00:07:24,268 --> 00:07:25,808 to change their behaviors? 147 00:07:25,808 --> 00:07:30,208 Do you think these conversations that we're having about AI are really just a 148 00:07:30,208 --> 00:07:32,468 microcosm of the larger issues for business? 149 00:07:32,468 --> 00:07:37,856 Or are there some things about AI that are really distinctive in these discussions? 150 00:07:38,894 --> 00:07:40,474 Probably a little bit of both. 151 00:07:40,474 --> 00:07:43,634 You know, I think that... 152 00:07:43,634 --> 00:07:44,324 Sorry, just a quick... 153 00:07:44,324 --> 00:07:45,304 I'm getting some feedback. 154 00:07:45,304 --> 00:07:46,634 I can hear myself a little bit. 155 00:07:46,634 --> 00:07:50,406 Is it coming out of your speaker and back into your microphone? 156 00:07:50,406 --> 00:07:54,312 I could put my earbuds in. 157 00:07:58,480 --> 00:08:00,820 Yeah, I'm not hearing anything, but. 158 00:08:00,820 --> 00:08:03,200 Yeah, it's probably on my hand. 159 00:08:03,840 --> 00:08:04,800 OK. 160 00:08:05,780 --> 00:08:07,500 It's coming out of me. 161 00:08:07,620 --> 00:08:08,850 Yes, it might be. 162 00:08:08,850 --> 00:08:10,240 Let me just. 163 00:08:14,096 --> 00:08:16,776 this connects up with the headphones. 164 00:08:16,776 --> 00:08:17,618 But yeah. 165 00:08:27,760 --> 00:08:29,000 Can you say something? 166 00:08:29,000 --> 00:08:29,720 Sorry. 167 00:08:30,420 --> 00:08:32,540 Okay, yeah, now I'm hearing you in the earbuds. 168 00:08:32,540 --> 00:08:32,780 Okay. 169 00:08:32,780 --> 00:08:34,080 Okay. 170 00:08:34,620 --> 00:08:36,050 Let's let's try that. 171 00:08:36,050 --> 00:08:37,580 So now, now I can hear you in the earbuds. 172 00:08:37,580 --> 00:08:38,620 Hopefully that'll. 173 00:08:38,900 --> 00:08:39,758 All right. 174 00:08:39,758 --> 00:08:41,038 I can't even see your earbuds. 175 00:08:41,038 --> 00:08:41,378 That's good. 176 00:08:41,378 --> 00:08:44,518 I guess I can see them barely, but if I'm, if I'm really looking. 177 00:08:44,698 --> 00:08:45,638 So, okay. 178 00:08:45,638 --> 00:08:48,008 So let's say you asked the question and obviously this look at edit out, you asked 179 00:08:48,008 --> 00:08:52,188 the question, is this sort of microcosm for ethics and business or is there 180 00:08:52,188 --> 00:08:53,178 something special about AI? 181 00:08:53,178 --> 00:08:54,278 Yep. 182 00:08:54,698 --> 00:08:58,598 Um, so it's a little bit of both. 183 00:08:58,598 --> 00:09:04,828 I think that one reason why the ethical risks of AI are much discussed is because 184 00:09:04,828 --> 00:09:07,182 it's of the nature of the beast that is AI. 185 00:09:07,182 --> 00:09:11,042 machine learning that certain kinds of risks are quite likely to be realized 186 00:09:11,042 --> 00:09:16,402 despite the best intentions in some cases of the people designing the AI. 187 00:09:16,402 --> 00:09:21,182 So we get things like biased AI and black box AI, we get privacy violations and 188 00:09:21,182 --> 00:09:25,962 those kinds of ethical breaches are not usually the result of bad actors, of bad 189 00:09:25,962 --> 00:09:26,702 apples. 190 00:09:26,702 --> 00:09:31,012 And so since we've got this powerful new tool, people want to use it for innovation 191 00:09:31,012 --> 00:09:34,162 to drive revenue, to drive internal efficiency, et cetera. 192 00:09:34,638 --> 00:09:36,458 But it comes with these associated risks. 193 00:09:36,458 --> 00:09:40,798 We naturally get this kind of conversation around risks. 194 00:09:41,338 --> 00:09:45,408 I also think that there's probably some contingent story to be told about why 195 00:09:45,408 --> 00:09:50,788 people are paying particular attention to this technology in ways that people don't 196 00:09:50,788 --> 00:09:54,798 pay attention to ethical breaches by corporations more generally in a way. 197 00:09:54,798 --> 00:10:00,168 So perhaps there's some story to be told about how the people in the technological 198 00:10:00,168 --> 00:10:04,046 space, they tend to be younger, they tend to be more activist -y. 199 00:10:04,046 --> 00:10:10,906 And so you get the overlapping diagram of being a technologist and being sort of 200 00:10:10,906 --> 00:10:13,946 activist -y, social media active, social justice warrior. 201 00:10:13,946 --> 00:10:17,906 And since you get that overlap of the diagram, you get as a matter of sort of 202 00:10:17,906 --> 00:10:21,606 historical contingent fact, lots of attention on the ethical impacts of AI. 203 00:10:22,736 --> 00:10:23,446 Yeah, it's interesting. 204 00:10:23,446 --> 00:10:27,516 And obviously we've seen a lot of evolution in terms of adoption of AI and 205 00:10:27,516 --> 00:10:30,636 awareness of AI in the last several years. 206 00:10:30,636 --> 00:10:34,476 Have the kinds of questions that companies ask you or the things that they bring you 207 00:10:34,476 --> 00:10:35,956 in for changed? 208 00:10:37,198 --> 00:10:38,478 I wouldn't say that they've changed radically. 209 00:10:38,478 --> 00:10:42,808 I think that the frequency with which the conversations are had, the frequency with 210 00:10:42,808 --> 00:10:46,038 which people want help with this sort of thing has increased over the years. 211 00:10:46,038 --> 00:10:50,448 There are certain kinds of risks that have popped up very high on the radar as of 212 00:10:50,448 --> 00:10:53,378 late, specifically since ChatGPT came out. 213 00:10:53,398 --> 00:10:58,278 Very much increased concerns around cybersecurity or the security integrity of 214 00:10:58,278 --> 00:10:59,138 their data. 215 00:10:59,138 --> 00:11:00,888 That's been a major issue, of course. 216 00:11:00,888 --> 00:11:05,058 And then, of course, the hallucination issue that ChatGPT or other models will 217 00:11:05,058 --> 00:11:06,606 output just false. 218 00:11:06,606 --> 00:11:07,906 false statements. 219 00:11:08,186 --> 00:11:10,846 So I think, you know, if I said, you know, what are the big risks that have come out 220 00:11:10,846 --> 00:11:11,096 lately? 221 00:11:11,096 --> 00:11:14,126 It'd be the cyber risks and the hallucination risks. 222 00:11:14,126 --> 00:11:18,516 The bias risks that are associated with generative AI, those have been around 223 00:11:18,516 --> 00:11:20,666 since we've been talking about AI ethics at all. 224 00:11:20,666 --> 00:11:24,956 So that's not really new, but it's perhaps exacerbated or, or especially now that it 225 00:11:24,956 --> 00:11:30,606 comes in visual form, it's captured the imagination of people a bit more, but 226 00:11:30,606 --> 00:11:32,146 nothing has fundamentally changed. 227 00:11:32,146 --> 00:11:34,706 I don't think it's just awareness has grown. 228 00:11:34,786 --> 00:11:36,420 It's just that awareness has grown. 229 00:11:38,320 --> 00:11:41,560 So in terms of generative AI, you mentioned the hallucination potential. 230 00:11:41,560 --> 00:11:44,542 How should we think about that from an ethics standpoint? 231 00:11:48,368 --> 00:11:50,920 You're on mute now, it looks like. 232 00:11:52,686 --> 00:11:53,934 Sorry, I muted myself by accident. 233 00:11:53,934 --> 00:11:54,866 Not about. 234 00:11:56,238 --> 00:12:00,788 The hallucination problem is not tremendously, no, I was thinking something 235 00:12:00,788 --> 00:12:01,494 else. 236 00:12:03,950 --> 00:12:08,170 The hallucination problem is a problem both for the technologists and for 237 00:12:08,170 --> 00:12:09,130 ethicists. 238 00:12:09,130 --> 00:12:11,750 So, one, everyone wants accurate AI. 239 00:12:11,750 --> 00:12:14,670 Nobody wants things that are said that are false. 240 00:12:14,670 --> 00:12:16,410 That makes it useless. 241 00:12:16,410 --> 00:12:18,690 So, one, there's just a sort straight -up technical problem. 242 00:12:18,690 --> 00:12:23,970 It only veers into the lane of the ethicist when it's giving outputs that are 243 00:12:23,970 --> 00:12:27,120 ethically significant or can have an impact on, say, the livelihood of people. 244 00:12:27,120 --> 00:12:31,110 So, if it's hallucinating about where you should put your money, let's say it's 245 00:12:31,110 --> 00:12:33,070 acting as a financial advisor, 246 00:12:33,070 --> 00:12:36,720 or it's hallucinating about what the appropriate diagnosis is, or it's 247 00:12:36,720 --> 00:12:39,370 hallucinating about whether you can combine these two kinds of drugs when 248 00:12:39,370 --> 00:12:43,230 you're asking about your prescriptions or something along those lines, then the 249 00:12:43,230 --> 00:12:44,830 ethical impacts are serious. 250 00:12:44,830 --> 00:12:48,430 So there's a way in which, you know, I think about the ethics of hallucinations, 251 00:12:48,430 --> 00:12:52,530 I think I think this, as something akin to the ethics of bridge building. 252 00:12:52,550 --> 00:12:56,780 You know, when you build a bridge, you don't want the thing to fall apart, 253 00:12:56,780 --> 00:12:58,570 otherwise it would be a useless bridge. 254 00:12:58,610 --> 00:13:01,902 Of course, if it falls apart, if the technical aspects are all out of whack, 255 00:13:01,902 --> 00:13:05,432 then so too will the ethical impacts be out of whack, because then people will be 256 00:13:05,432 --> 00:13:06,712 on the bridge and it'll collapse and they'll die. 257 00:13:06,712 --> 00:13:07,982 So that's really ethically bad. 258 00:13:07,982 --> 00:13:09,222 But it's really fine. 259 00:13:09,222 --> 00:13:14,852 It's not, if you like an ethical problem as such to solve, it's more like there's 260 00:13:14,852 --> 00:13:20,562 these quasi -technical problems like outputting false information that has or 261 00:13:20,562 --> 00:13:22,638 can have serious ethical impacts. 262 00:13:23,770 --> 00:13:27,120 But there are these questions with generative AI about what's the appropriate 263 00:13:27,120 --> 00:13:34,710 response, what might be toxic or improper, culturally insensitive, or biased in ways 264 00:13:34,710 --> 00:13:36,940 where reasonable minds can differ. 265 00:13:36,940 --> 00:13:42,180 So if a company brings you in to figure those things out, how would you advise 266 00:13:42,180 --> 00:13:42,760 them? 267 00:13:43,374 --> 00:13:44,904 So there's a couple of things to say here. 268 00:13:44,904 --> 00:13:50,614 I mean, the first thing to say is that usually my company, myself, my company is 269 00:13:50,614 --> 00:13:55,024 not brought in at that really granular level of we've got this particular model 270 00:13:55,024 --> 00:13:56,624 for this particular use case. 271 00:13:56,624 --> 00:14:00,254 And we want to know what are the standards that should govern this particular model 272 00:14:00,254 --> 00:14:00,944 that happens. 273 00:14:00,944 --> 00:14:04,734 Sometimes it's starting to come up a little bit more precisely because of 274 00:14:04,734 --> 00:14:07,344 particular applications of LLMs for exactly the reasons that you've just 275 00:14:07,344 --> 00:14:08,094 stated. 276 00:14:08,094 --> 00:14:11,804 Usually what they're looking for is how do we even, how do we get our organizational 277 00:14:11,804 --> 00:14:13,294 hands around this? 278 00:14:13,294 --> 00:14:14,374 more generally. 279 00:14:14,374 --> 00:14:16,994 So how do we create policy around this kind of thing? 280 00:14:16,994 --> 00:14:21,394 How do we create the procedures and augment existing workflows for our say 281 00:14:21,394 --> 00:14:24,734 data scientists and product managers to look for these kinds of risks, to flag 282 00:14:24,734 --> 00:14:25,694 these kinds of things? 283 00:14:25,694 --> 00:14:29,884 How do we put together the appropriate risk board or committee so that there's 284 00:14:29,884 --> 00:14:31,854 someone to whom they can elevate those kinds of concerns? 285 00:14:31,854 --> 00:14:36,064 How do we create the grounds on which that committee is going to make decisions about 286 00:14:36,064 --> 00:14:40,234 what the appropriate benchmarks are for sufficiently safe or sufficiently 287 00:14:40,234 --> 00:14:40,874 appropriate? 288 00:14:40,874 --> 00:14:42,308 Something along those lines. 289 00:14:43,310 --> 00:14:48,130 As for how those substantive decisions get made, there's roughly two kinds of things, 290 00:14:48,130 --> 00:14:50,250 two kinds of tools that I like to reference. 291 00:14:50,250 --> 00:14:53,130 One is an ethics statement or an AI ethics statement. 292 00:14:53,130 --> 00:14:55,970 Some people are going to call them responsible AI statements. 293 00:14:56,190 --> 00:14:59,630 Those are really, if you like, high level articulations of values connected to 294 00:14:59,630 --> 00:15:00,730 guardrails of action. 295 00:15:00,730 --> 00:15:03,170 So because we value privacy, we'll never do X. 296 00:15:03,170 --> 00:15:05,630 We'll never sell your data to a third party. 297 00:15:05,970 --> 00:15:12,334 Because we value accountability, we will integrate people's 298 00:15:12,334 --> 00:15:15,874 compliance with our AI ethics program into their performance reviews. 299 00:15:15,874 --> 00:15:20,864 We'll put ownership of the AI ethics program in the hands of a C -suite 300 00:15:20,864 --> 00:15:22,494 executive, that sort of thing. 301 00:15:22,494 --> 00:15:27,044 So there's those high level policies that are, I'll give you one more, because we 302 00:15:27,044 --> 00:15:33,384 value justice or fairness, we will vet our AI models for bias at each and every stage 303 00:15:33,384 --> 00:15:34,574 of the AI life cycle. 304 00:15:34,574 --> 00:15:39,344 So these are sort of commitments of action tied to values that get at least some 305 00:15:39,344 --> 00:15:41,518 procedural guardrails in place. 306 00:15:41,518 --> 00:15:45,648 and some, some substantive actions that are ruled out after that. 307 00:15:45,648 --> 00:15:49,348 I mean, even getting that as a big deal and implementing those procedures, that's 308 00:15:49,348 --> 00:15:49,858 a big deal. 309 00:15:49,858 --> 00:15:55,088 So forget about, you know, as an act, as at least as a former academic and you, you 310 00:15:55,088 --> 00:15:56,958 know, you want to deal with those really tough cases. 311 00:15:56,958 --> 00:16:01,558 Let's deal with these really tough cases where there's a lot of gray area. 312 00:16:01,558 --> 00:16:03,538 How do we decide what's appropriate for this LM? 313 00:16:03,538 --> 00:16:08,388 But the truth is, and I would love for my, for most of my customers and my clients to 314 00:16:08,388 --> 00:16:09,710 be like, yeah, let's help us with these. 315 00:16:09,710 --> 00:16:12,130 really hard, tough gray area cases. 316 00:16:12,270 --> 00:16:14,410 But that's, that's not where their heads are. 317 00:16:14,410 --> 00:16:17,770 Their heads are, can we just, that's running. 318 00:16:17,770 --> 00:16:18,910 Gray area cases is running. 319 00:16:18,910 --> 00:16:20,280 We're just learning how to walk here. 320 00:16:20,280 --> 00:16:22,030 We're barely crawling at the moment. 321 00:16:22,030 --> 00:16:25,410 How can we crawl safely without like falling off the table? 322 00:16:25,410 --> 00:16:29,650 And so, sorry, I'm going on a bit of a ramble. 323 00:16:29,650 --> 00:16:34,240 So one is there are those kinds of conversations about really concrete models 324 00:16:34,240 --> 00:16:35,670 and what constitutes appropriate. 325 00:16:35,670 --> 00:16:38,254 One thing to do is to have general kinds of policies. 326 00:16:38,254 --> 00:16:41,934 The second thing to do is to have what I call ethical case law. 327 00:16:41,934 --> 00:16:45,754 So this is when you articulate certain kinds of cases that you would say this 328 00:16:45,754 --> 00:16:46,504 would be inappropriate. 329 00:16:46,504 --> 00:16:48,154 This would be not appropriate. 330 00:16:48,154 --> 00:16:52,114 Sort of get internal alignment on those being what I like to call the ethical 331 00:16:52,114 --> 00:16:53,114 nightmares. 332 00:16:53,154 --> 00:16:57,824 And then when you have a particular novel case, like should it be allowed to say X 333 00:16:57,824 --> 00:17:04,054 to Y, let's say something with a certain political content to a general audience, 334 00:17:04,054 --> 00:17:07,404 you can compare it to your ethical case law, internal ethical case law and see 335 00:17:07,404 --> 00:17:08,046 whether... 336 00:17:08,046 --> 00:17:09,346 that provides sufficient guidance. 337 00:17:09,346 --> 00:17:11,886 And then of course, there's going to be deliberation among the ethics board or the 338 00:17:11,886 --> 00:17:13,066 risk committee, whatever. 339 00:17:13,086 --> 00:17:17,366 The thing that I was backing up to say too soon, I sort of interrupted myself is the 340 00:17:17,366 --> 00:17:21,086 truth of the matter is while the hard cases are sexy and they're interesting, 341 00:17:21,086 --> 00:17:25,836 especially to academics, they are not where most companies are because most 342 00:17:25,836 --> 00:17:29,886 companies are dealing with the not crazy gray cases. 343 00:17:29,886 --> 00:17:34,986 You know, maybe an open AI is dealing with those gray cases, a Google, a Microsoft, 344 00:17:35,086 --> 00:17:37,742 but a financial services company. 345 00:17:37,742 --> 00:17:44,742 Insurance industry, health care and life sciences, food and beverage, apparel, 346 00:17:44,742 --> 00:17:48,442 retail, they're not dealing with those crazy situations. 347 00:17:48,442 --> 00:17:50,672 They're just hiring people, right? 348 00:17:50,672 --> 00:17:52,442 They're using AI software to hire people. 349 00:17:52,442 --> 00:17:55,162 How do we not do that in a discriminatory and biased way? 350 00:17:55,162 --> 00:17:57,932 So I don't know, this is my sort of, okay, I'm going to stop talking now, but this is 351 00:17:57,932 --> 00:18:02,372 my very long -winded way of saying most companies don't deal with those kinds of 352 00:18:02,372 --> 00:18:03,428 tough cases. 353 00:18:03,792 --> 00:18:04,622 Yeah, no, that was great. 354 00:18:04,622 --> 00:18:08,462 I there was a lot of real good meat in what you put out there. 355 00:18:08,462 --> 00:18:10,892 So let me just unpack some of it. 356 00:18:11,332 --> 00:18:16,132 If companies are just starting down this path, what's the most important thing for 357 00:18:16,132 --> 00:18:18,024 them to do first? 358 00:18:18,574 --> 00:18:22,384 I always like to say that sort of table stakes is a call it a risk assessment or a 359 00:18:22,384 --> 00:18:26,134 gap analysis or AI ethical risk maturity assessment. 360 00:18:26,134 --> 00:18:30,514 I think most organizations don't sort of know where they are and what they have in 361 00:18:30,514 --> 00:18:32,594 relation to what those risks may be. 362 00:18:32,594 --> 00:18:37,434 So looking to see what existing governance structures do we have that cover this sort 363 00:18:37,434 --> 00:18:38,154 of thing? 364 00:18:38,154 --> 00:18:39,314 What, what do we lack? 365 00:18:39,314 --> 00:18:41,464 What kinds of policies do we have that cover this thing? 366 00:18:41,464 --> 00:18:44,434 What policies do we lack in most cases? 367 00:18:44,434 --> 00:18:48,238 If a company has done literally nothing on the responsible AI front, 368 00:18:48,238 --> 00:18:53,368 then you might not even need to do that gap analysis risk assessment maturity 369 00:18:53,368 --> 00:18:53,638 assessment. 370 00:18:53,638 --> 00:18:54,838 You know, you've got nothing. 371 00:18:54,838 --> 00:18:59,238 And now you've got to start with an AI ethics statement, which is going to 372 00:18:59,238 --> 00:19:02,798 articulate what those standards are and the procedures that are associated with 373 00:19:02,798 --> 00:19:03,268 those standards. 374 00:19:03,268 --> 00:19:07,638 And frankly, if you can, the metrics that are associated with those procedures, I 375 00:19:07,638 --> 00:19:12,568 like to draw a really clear line from values of a business or sorry, AI ethical 376 00:19:12,568 --> 00:19:17,582 risk values to procedures and the metrics that are used. 377 00:19:17,582 --> 00:19:22,922 to track compliance and impact of those procedures. 378 00:19:22,922 --> 00:19:25,702 So that's a good first step. 379 00:19:26,022 --> 00:19:29,042 And then sort of part and partial that is then to start building a customized 380 00:19:29,042 --> 00:19:30,202 framework. 381 00:19:30,202 --> 00:19:33,782 When I say framework, I mean something somewhat specific. 382 00:19:33,782 --> 00:19:38,942 It's going to include policies, additional governance structures, governance bodies, 383 00:19:38,942 --> 00:19:44,402 like an ethics board, it's going to include metrics, might include procedures, 384 00:19:44,662 --> 00:19:47,380 might include race matrices, that sort of thing. 385 00:19:48,048 --> 00:19:51,608 Yeah, so it's one of things I really love about your book, Ethical Machines, that 386 00:19:51,608 --> 00:19:56,098 it's rooted in serious ethics, but it's very pragmatic in terms of these kinds of 387 00:19:56,098 --> 00:19:57,668 things that you're talking about. 388 00:19:57,668 --> 00:20:01,648 But how much do companies really differ? 389 00:20:01,648 --> 00:20:04,548 You're talking about customization, but I think there's probably no company would 390 00:20:04,548 --> 00:20:08,438 say, well, our values, we don't care about privacy, or our values, our discrimination 391 00:20:08,438 --> 00:20:09,428 is OK. 392 00:20:09,428 --> 00:20:12,078 So where is it really going to be customized? 393 00:20:12,078 --> 00:20:13,968 There's actually quite a bit of customization to be done. 394 00:20:13,968 --> 00:20:17,438 It's actually most of what we do is the customization bit. 395 00:20:17,878 --> 00:20:20,278 There are lots of things that vary. 396 00:20:20,278 --> 00:20:22,238 So for instance, you're correct. 397 00:20:22,238 --> 00:20:25,468 Two different organizations, a hundred organizations are going to say we value 398 00:20:25,468 --> 00:20:26,198 privacy. 399 00:20:26,198 --> 00:20:30,678 What that means is going to vary quite a bit across organizations. 400 00:20:30,678 --> 00:20:34,658 Some will say we'll never sell consumer data to a third party. 401 00:20:34,658 --> 00:20:36,498 That's part of what our commitment to privacy is. 402 00:20:36,498 --> 00:20:38,542 Other companies will say, no, no, no. 403 00:20:38,542 --> 00:20:39,722 We're going to sell that data. 404 00:20:39,722 --> 00:20:40,702 That's what we do. 405 00:20:40,702 --> 00:20:42,162 It's a major revenue source for us. 406 00:20:42,162 --> 00:20:46,282 So we respect people's privacy, which means we will always ensure that we're 407 00:20:46,282 --> 00:20:49,992 going to aggregate and anonymize that data and use various kinds of privacy 408 00:20:49,992 --> 00:20:52,942 preserving techniques like Q anonymity. 409 00:20:52,942 --> 00:20:59,012 But that's, um, K anonymity, but, uh, that's, uh, you know, that's very 410 00:20:59,012 --> 00:21:01,162 different from saying we're not going to sell the data at all. 411 00:21:01,162 --> 00:21:04,922 There are differences in what the governance structures have to look like. 412 00:21:04,922 --> 00:21:08,450 So look, some organizations are structured like partnerships. 413 00:21:08,558 --> 00:21:13,938 And some are, if you like the sort of the classic pyramid of CEO on top and then 414 00:21:13,938 --> 00:21:15,298 everyone below that. 415 00:21:15,298 --> 00:21:18,618 When you have an organization that has sort of like a CEO on top and the CEO 416 00:21:18,618 --> 00:21:20,718 says, this is how it goes. 417 00:21:21,038 --> 00:21:25,148 This is a gross oversimplification, but more or less, okay, if you've got the buy 418 00:21:25,148 --> 00:21:27,258 -in of the CEO, that's more or less how it's going to go. 419 00:21:27,258 --> 00:21:31,918 If on the other hand, you have say a giant consultancy spread across dozens of 420 00:21:31,918 --> 00:21:37,326 markets where there's not that kind of top down change. 421 00:21:37,326 --> 00:21:41,016 because each you need to get buy -in from each of the partnerships, each of the 422 00:21:41,016 --> 00:21:42,526 heads of each of the partnerships. 423 00:21:42,666 --> 00:21:45,806 Now governance looks totally different because there's not, if you like, one 424 00:21:45,806 --> 00:21:46,986 board to rule them all. 425 00:21:46,986 --> 00:21:50,226 They might say like, yeah, that's good for you guys in the US, but we're over here in 426 00:21:50,226 --> 00:21:51,006 the EU. 427 00:21:51,006 --> 00:21:52,796 And so we're going to do things, we're going have to do things a little bit 428 00:21:52,796 --> 00:21:53,586 differently over here. 429 00:21:53,586 --> 00:21:55,546 Our standards are going to be different here. 430 00:21:55,666 --> 00:21:59,546 Some people want to apply the same standards across all markets. 431 00:21:59,666 --> 00:22:03,476 So yes, we're in EMEA, we're in the EU, we're in the US, you know, North America, 432 00:22:03,476 --> 00:22:04,066 et cetera. 433 00:22:04,066 --> 00:22:06,376 And we're going have the same standards across the board. 434 00:22:06,446 --> 00:22:09,116 Others say, no, no, no, we're going to have different standards for different 435 00:22:09,116 --> 00:22:11,666 markets, different risk thresholds for different markets. 436 00:22:11,666 --> 00:22:14,066 We're going to roll this out in different ways. 437 00:22:14,066 --> 00:22:16,296 So we're going to roll it out by market, for instance. 438 00:22:16,296 --> 00:22:18,986 So first we're going to roll out this program in the U .S. 439 00:22:18,986 --> 00:22:22,186 Then we're going to expand to North America, and then we're going to expand to 440 00:22:22,186 --> 00:22:24,746 the U and then to EMEA, et cetera. 441 00:22:24,746 --> 00:22:27,376 Others say, no, no, we're not going to roll out by market. 442 00:22:27,376 --> 00:22:28,546 We're going to roll out by role. 443 00:22:28,546 --> 00:22:30,526 So first we're going to roll it out to all the data scientists. 444 00:22:30,526 --> 00:22:33,086 I don't care where geographically they happen to be. 445 00:22:33,086 --> 00:22:35,022 And then we're going to roll out to HR. 446 00:22:35,022 --> 00:22:36,602 And then we're going to roll it out to marketing. 447 00:22:36,602 --> 00:22:40,222 And so there's different ways of, well, that's going to be customized. 448 00:22:40,262 --> 00:22:42,042 Who's going to sit on the board? 449 00:22:42,042 --> 00:22:44,122 What kind of authority does the board have? 450 00:22:44,122 --> 00:22:47,642 Can the ethics board be overruled by a more senior executive? 451 00:22:47,642 --> 00:22:48,902 That's it. 452 00:22:48,902 --> 00:22:51,502 Some organizations can say, no, it can't be overruled. 453 00:22:51,502 --> 00:22:52,922 That's the final say. 454 00:22:52,942 --> 00:22:56,032 Other will say, yeah, it can be overruled by say the CEO or something along those 455 00:22:56,032 --> 00:22:57,622 lines or a chief data officer. 456 00:22:57,622 --> 00:23:00,242 Chief data officer is not actually on that board. 457 00:23:00,662 --> 00:23:04,396 So there's a tremendous amount of, what does training look like? 458 00:23:04,396 --> 00:23:04,926 What are risk? 459 00:23:04,926 --> 00:23:07,806 We already do risk matrices in our organization. 460 00:23:07,806 --> 00:23:11,726 And the way that we do risk assessments is different how our peers do risk 461 00:23:11,726 --> 00:23:12,166 assessments. 462 00:23:12,166 --> 00:23:16,226 We want responsible AI or AI ethical risk assessments that cohere with, that 463 00:23:16,226 --> 00:23:19,626 harmonize or integrate with how we already think about risk in our organization, how 464 00:23:19,626 --> 00:23:21,526 we already do those risk assessments. 465 00:23:21,546 --> 00:23:23,116 And same with metrics. 466 00:23:23,116 --> 00:23:28,646 We want metrics that mirror the kinds of risk metrics that we already have at the 467 00:23:28,646 --> 00:23:30,336 enterprise wide or department level, et cetera. 468 00:23:30,336 --> 00:23:34,148 So there's a tremendous amount of customization that needs to be done. 469 00:23:34,382 --> 00:23:39,862 That's why there's the NIST AI assurance or whatever it's called framework. 470 00:23:39,862 --> 00:23:40,662 And it's great. 471 00:23:40,662 --> 00:23:42,162 It's super smart. 472 00:23:42,162 --> 00:23:45,302 It's totally generic though, necessarily. 473 00:23:45,382 --> 00:23:50,762 And so you give that document to, here's a really smart, generic framework. 474 00:23:50,762 --> 00:23:54,084 And there is a tremendous amount of work left to be done. 475 00:23:55,536 --> 00:23:56,712 But are there right - 476 00:23:59,664 --> 00:24:04,260 of understanding what the organization is and customizing the AI ethics. 477 00:24:05,390 --> 00:24:07,850 So I think there's reasonable and unreasonable. 478 00:24:07,850 --> 00:24:10,830 And I think this is a place where we sort of can think about it in terms of the 479 00:24:10,830 --> 00:24:11,650 political spectrum. 480 00:24:11,650 --> 00:24:14,390 I think that there are companies who are sort of on the left side of the spectrum, 481 00:24:14,390 --> 00:24:15,430 take a Patagonia. 482 00:24:15,430 --> 00:24:18,190 There are companies on the right side of the spectrum, take a Hobby Lobby. 483 00:24:18,190 --> 00:24:21,110 And we sort of think, yeah, that's fine in the corporate world having that. 484 00:24:21,110 --> 00:24:22,590 I mean, most anywhere you think about it. 485 00:24:22,590 --> 00:24:26,070 Having that kind of diversity of values, that's acceptable to us. 486 00:24:26,070 --> 00:24:27,660 There's beyond the pale, right? 487 00:24:27,660 --> 00:24:32,210 So you don't want KKK values anywhere around that would be beyond the pale. 488 00:24:32,210 --> 00:24:33,720 So when are there right and wrong answers? 489 00:24:33,720 --> 00:24:34,638 I think that... 490 00:24:34,638 --> 00:24:37,748 There are some wrong answers certainly, but there's lots of right answers that are 491 00:24:37,748 --> 00:24:41,258 mutually incompatible, just like Patagonia's stance on some issues is going 492 00:24:41,258 --> 00:24:44,958 to be incompatible with the stance of, say, a hobby lobby. 493 00:24:45,038 --> 00:24:49,918 That said, there's also right and wrong answers as to, hey, does doing this 494 00:24:49,918 --> 00:24:51,918 actually help us to achieve our goals? 495 00:24:51,918 --> 00:24:54,658 Is this means sufficient to achieve our end? 496 00:24:54,658 --> 00:24:57,278 And then there are certainly right and wrong answers. 497 00:24:57,718 --> 00:24:59,908 They'll say, oh, we'll just do this. 498 00:24:59,908 --> 00:25:01,488 We'll just tell everyone x. 499 00:25:01,488 --> 00:25:03,748 And we'll say, that ain't going to do it. 500 00:25:03,748 --> 00:25:04,462 You could. 501 00:25:04,462 --> 00:25:05,682 That's not what's going to happen. 502 00:25:05,682 --> 00:25:08,542 You could tell them that, but that will definitely lead to a lot of things falling 503 00:25:08,542 --> 00:25:09,142 between the cracks. 504 00:25:09,142 --> 00:25:11,102 We highly recommend against that. 505 00:25:11,482 --> 00:25:15,692 So I think there's right and wrong and reasonable and unreasonable in the realm 506 00:25:15,692 --> 00:25:19,262 of what the values are, what the standards are, what the procedures are. 507 00:25:19,262 --> 00:25:22,252 And then there's right and wrong in terms of, are these means sufficient for 508 00:25:22,252 --> 00:25:23,334 achieving our ends? 509 00:25:23,984 --> 00:25:28,084 So you gave an example of can the CEO overrule the ethics board? 510 00:25:28,144 --> 00:25:31,304 How would you help a company think through what the answer is? 511 00:25:32,206 --> 00:25:34,686 Well, I mean, it's this. 512 00:25:34,686 --> 00:25:38,776 Look, if you really want to take the ethics of this seriously, then you have an 513 00:25:38,776 --> 00:25:44,226 ethics board, cross -functional ethics board, well -trained, both in, if you 514 00:25:44,226 --> 00:25:48,526 like, ethical deliberation on the kinds of issues they're going to face, but also 515 00:25:48,526 --> 00:25:53,056 well -trained in understanding, because usually it's this sort of board that's 516 00:25:53,056 --> 00:25:57,366 also going to oversee the rollout, compliance, and impact of the program, 517 00:25:57,366 --> 00:26:00,622 well -trained in how to oversee a program like this. 518 00:26:00,622 --> 00:26:05,762 And if you really want to prioritize ethics and trustworthiness, etc., then you 519 00:26:05,762 --> 00:26:08,952 say the decision by the ethics committee is the final one. 520 00:26:08,952 --> 00:26:09,822 That's it. 521 00:26:09,822 --> 00:26:13,542 What they say, if they say this is a no -go, then it's a no -go. 522 00:26:13,542 --> 00:26:16,932 And that's really good if you want to prioritize more than, you know, not maybe 523 00:26:16,932 --> 00:26:20,752 not more than anything, but very highly prioritize something like long -term 524 00:26:20,752 --> 00:26:23,654 trustworthiness of the brand, then you do that. 525 00:26:24,334 --> 00:26:28,554 That said, I think we need to be honest here and say, there's going to be cases 526 00:26:28,554 --> 00:26:31,454 where doing the unethical thing is going to be really profitable. 527 00:26:31,454 --> 00:26:37,074 There are sort of AI ethicists out there who say, oh, the right thing is also the 528 00:26:37,074 --> 00:26:37,634 profitable thing. 529 00:26:37,634 --> 00:26:39,344 And so we're like, no, that we're true. 530 00:26:39,344 --> 00:26:40,374 We wouldn't have to talk about ethics. 531 00:26:40,374 --> 00:26:43,854 We'd be done here because they would already do the profitable thing and then 532 00:26:43,854 --> 00:26:44,894 the ethics that would fall out of that. 533 00:26:44,894 --> 00:26:47,034 So there's clearly conflicts. 534 00:26:47,154 --> 00:26:52,884 And so if you say, listen, if you want to give yourself some wiggle room where 535 00:26:52,884 --> 00:26:53,934 the... 536 00:26:53,934 --> 00:27:01,694 The reward is so great, the ROI, say bottom line ROI is so great that you think 537 00:27:01,694 --> 00:27:05,094 the CEO should have the ability to override the ethics committee. 538 00:27:05,454 --> 00:27:07,454 That's not, it's not crazy. 539 00:27:07,454 --> 00:27:12,324 It's a little bit dangerous in the short term anyway, but it's not a crazy view to 540 00:27:12,324 --> 00:27:13,134 hold. 541 00:27:13,134 --> 00:27:17,944 In fact, I think in some cases it's quite reasonable to say, you know, yeah, this is 542 00:27:17,944 --> 00:27:21,194 a business at the end of the day and we're willing to take bigger risks. 543 00:27:21,794 --> 00:27:22,510 It's not. 544 00:27:22,510 --> 00:27:24,240 It's it's it's gonna tell me another context. 545 00:27:24,240 --> 00:27:27,990 I mean, there are certainly cases in which CEOs choose them morally abhorrent for 546 00:27:27,990 --> 00:27:28,980 some short -term profits. 547 00:27:28,980 --> 00:27:30,210 That's a bad thing. 548 00:27:30,210 --> 00:27:33,160 But should it be sort of, if you like, structurally disallowed by the 549 00:27:33,160 --> 00:27:34,250 organization? 550 00:27:34,270 --> 00:27:36,108 Probably not in many cases. 551 00:27:36,632 --> 00:27:42,142 If I'm looking at a company and trying to assess how serious they are, how effective 552 00:27:42,142 --> 00:27:45,832 they are in terms of what they're doing on AI ethics, are there things that I should 553 00:27:45,832 --> 00:27:48,172 look for, their tells? 554 00:27:56,232 --> 00:27:57,432 Let's start with that, sure. 555 00:27:57,432 --> 00:28:01,290 I'm a customer or I'm someone just trying to evaluate the company from the outside. 556 00:28:01,290 --> 00:28:02,370 You'll have no idea. 557 00:28:02,370 --> 00:28:03,470 No clue. 558 00:28:03,750 --> 00:28:07,270 The people, the sort of statements that they put out are phenomenally superficial. 559 00:28:07,270 --> 00:28:10,610 They're not going to make their policy public. 560 00:28:10,610 --> 00:28:14,590 They're usually not going to, you know, when we work with our clients, we train 561 00:28:14,590 --> 00:28:18,190 them how to do auditing, how to ready themselves for internal and external 562 00:28:18,190 --> 00:28:20,190 auditing of their AI ethics program. 563 00:28:20,190 --> 00:28:22,190 And in principle, they can make that stuff public. 564 00:28:22,190 --> 00:28:24,270 Number one, most companies are not there yet. 565 00:28:24,270 --> 00:28:25,870 They haven't done enough. 566 00:28:25,870 --> 00:28:27,670 They haven't done enough to audit. 567 00:28:27,970 --> 00:28:29,902 They're still designing the program. 568 00:28:29,902 --> 00:28:32,842 Some are at that audit stage, but most are not. 569 00:28:32,842 --> 00:28:35,022 And two, most of them are not going to be public about it. 570 00:28:35,022 --> 00:28:36,842 It just puts a target on their back. 571 00:28:36,982 --> 00:28:42,472 In the early days, this doesn't happen anymore, but five years ago, companies 572 00:28:42,472 --> 00:28:46,372 were way more resistant to being public about what they said about AI ethics or 573 00:28:46,372 --> 00:28:50,362 ethics generally because they felt that if we say that we really care about ethics, 574 00:28:50,362 --> 00:28:53,922 we're putting a target on our back for journalists to catch us. 575 00:28:53,922 --> 00:28:57,612 So if we say we care a lot about AI ethics and then something goes haywire, we're 576 00:28:57,612 --> 00:28:59,054 going to get skewered. 577 00:28:59,054 --> 00:28:59,774 by the journalist. 578 00:28:59,774 --> 00:29:01,184 So let's just shut up about it. 579 00:29:01,184 --> 00:29:01,934 Keep quiet. 580 00:29:01,934 --> 00:29:05,174 We'll do this stuff internally, but let's not tell everyone about it. 581 00:29:05,174 --> 00:29:06,664 That's changed over the last five years. 582 00:29:06,664 --> 00:29:10,054 Now it's more like, oh, whoa, everyone else is putting out AI ethics statements. 583 00:29:10,054 --> 00:29:12,954 We better put out an AI ethics statement to keep up. 584 00:29:12,954 --> 00:29:14,034 So that's good. 585 00:29:14,034 --> 00:29:17,534 But the statements are, by and large, phenomenally superficial and mean nothing. 586 00:29:17,534 --> 00:29:20,174 The average citizen, they don't know anything. 587 00:29:20,174 --> 00:29:25,384 Even if they look at the privacy policy of the organization, I've seen many examples 588 00:29:25,384 --> 00:29:28,224 of paper tigers, many examples where, 589 00:29:43,182 --> 00:29:46,618 do you measure success in terms of the work that you do with organizations? 590 00:29:48,974 --> 00:29:49,604 How do I measure success? 591 00:29:49,604 --> 00:29:50,364 That's a good question. 592 00:29:50,364 --> 00:29:54,364 I mean, you know, if they're just doing the things that we're telling them to do 593 00:29:54,364 --> 00:29:56,894 more or less, I mean, when I say what we tell them to do, we're giving them 594 00:29:56,894 --> 00:29:57,284 options. 595 00:29:57,284 --> 00:30:01,214 We say, look, you can build your program this way or that way. 596 00:30:01,374 --> 00:30:04,644 You know, for instance, we say, you can build an AI ethical risk program or 597 00:30:04,644 --> 00:30:07,254 responsible AI program, or you can build a digital ethics program. 598 00:30:07,254 --> 00:30:12,224 The digital ethics program will include AI, but it's also a bit broader than that 599 00:30:12,224 --> 00:30:16,574 to include other kinds of technologies, including ones that we don't, either 600 00:30:16,574 --> 00:30:18,510 applications of AI, 601 00:30:18,510 --> 00:30:23,310 that we don't have yet, new kinds of AI we don't have yet, new kinds of technologies, 602 00:30:23,310 --> 00:30:24,720 new applications, new technologies. 603 00:30:24,720 --> 00:30:28,330 So let's say blockchain, quantum, AR, VR. 604 00:30:29,310 --> 00:30:32,650 Ideally, they would all say, oh yeah, digital ethics. 605 00:30:32,650 --> 00:30:36,380 It's the broadest, it's the most comprehensive, it's the most proactive way 606 00:30:36,380 --> 00:30:37,330 we can address this. 607 00:30:37,330 --> 00:30:39,790 These issues aren't going away, they're only growing. 608 00:30:39,790 --> 00:30:42,250 So we say, look, digital ethics is ideal. 609 00:30:42,250 --> 00:30:45,950 But sometimes they say, listen, we'd love to do that. 610 00:30:46,050 --> 00:30:48,646 We don't have the internal buy -in for that. 611 00:30:48,750 --> 00:30:54,250 What we have, we have board support for building an AI ethics program. 612 00:30:54,350 --> 00:30:56,460 That's what we're not going to go back to them and say, you know, we want to do 613 00:30:56,460 --> 00:30:58,610 something bigger or we want to do something broader. 614 00:30:58,610 --> 00:30:59,810 That's where our buy -in is. 615 00:30:59,810 --> 00:31:01,290 That's the budget we have. 616 00:31:01,290 --> 00:31:02,320 That's where our focus is. 617 00:31:02,320 --> 00:31:03,480 That's what our priorities are. 618 00:31:03,480 --> 00:31:05,230 So that's what we're going to do. 619 00:31:05,430 --> 00:31:08,390 Is that a success or a failure on my view? 620 00:31:08,390 --> 00:31:09,710 It's still a success. 621 00:31:09,710 --> 00:31:12,830 You know, it's not the ideal, but it's, it's a success. 622 00:31:12,930 --> 00:31:16,170 What's another success if they actually move on to implementation? 623 00:31:16,690 --> 00:31:17,678 There are. 624 00:31:17,678 --> 00:31:20,388 plenty of organizations that will start to write an ethics statement and then just 625 00:31:20,388 --> 00:31:20,758 stop. 626 00:31:20,758 --> 00:31:22,078 That's a failure. 627 00:31:22,598 --> 00:31:24,218 You've got to get to the point of implementation. 628 00:31:24,218 --> 00:31:27,358 You've got to get to the point of training your people. 629 00:31:27,358 --> 00:31:29,558 This is, this is crucial. 630 00:31:29,558 --> 00:31:32,918 There's got to be that, if you like role agnostic training, the training that 631 00:31:32,918 --> 00:31:35,778 everyone across the organization gets, then there has to be role specific 632 00:31:35,778 --> 00:31:36,438 training. 633 00:31:36,438 --> 00:31:41,858 There has to be updates or augmentations of workflows of say data scientists, 634 00:31:41,858 --> 00:31:43,518 people in HR, et cetera. 635 00:31:43,658 --> 00:31:44,958 There has to be documentation. 636 00:31:44,958 --> 00:31:46,382 You know, there's got to be all these. 637 00:31:46,382 --> 00:31:50,952 things where it actually impacts the day -to -day operations. 638 00:31:50,952 --> 00:31:53,642 So what we like to say to our clients is, okay, you've done all this, blah, blah, 639 00:31:53,642 --> 00:31:56,562 blah, now we got to get the implementation, because you can write all 640 00:31:56,562 --> 00:31:59,682 this stuff, you can have this policy, but that doesn't mean that the data scientist 641 00:31:59,682 --> 00:32:01,212 knows what they need to do tomorrow at 9 a .m. 642 00:32:01,212 --> 00:32:02,202 when they get to work. 643 00:32:02,202 --> 00:32:04,122 This needs to be made clear. 644 00:32:04,622 --> 00:32:08,812 And if they don't get to that, that's maybe not yet a failure, but it's not yet 645 00:32:08,812 --> 00:32:09,750 a success. 646 00:32:10,376 --> 00:32:15,956 One thing that's changed in this area in recent years has been that there's been an 647 00:32:15,956 --> 00:32:20,996 upsurge of regulation and legislation and discussion and passage of things like the 648 00:32:20,996 --> 00:32:22,396 AI Act in Europe. 649 00:32:22,756 --> 00:32:27,106 How does that impact on what you do or on what companies actually do on the ground? 650 00:32:27,106 --> 00:32:30,496 Because on the one hand, now there's going to be all these requirements. 651 00:32:30,496 --> 00:32:34,436 On the other hand, if this just becomes a check the box for compliance, then it 652 00:32:34,436 --> 00:32:35,816 might actually be counterproductive. 653 00:32:35,816 --> 00:32:37,308 So what do you see? 654 00:32:37,646 --> 00:32:39,446 I think it's too early to tell. 655 00:32:39,446 --> 00:32:40,206 It's still TBD. 656 00:32:40,206 --> 00:32:42,066 Yeah, there are regulations out there. 657 00:32:42,066 --> 00:32:44,956 You know, New York City came out with a regulation not so long ago. 658 00:32:44,956 --> 00:32:52,476 I think it passed this past summer that all HR tools, sorry, all AI tools involved 659 00:32:52,476 --> 00:32:58,266 in the HR hiring decision -making process need to be audited by a third party. 660 00:32:58,266 --> 00:33:01,766 And that's not being enforced. 661 00:33:01,766 --> 00:33:04,326 I have no idea if it's had any impact at all. 662 00:33:04,326 --> 00:33:06,232 The EU AI Act. 663 00:33:06,510 --> 00:33:11,550 promises to do more That's plausible, especially if you're in the EU. 664 00:33:11,550 --> 00:33:15,070 Of course, there are lots of companies in the US who operate in the EU So they'll 665 00:33:15,070 --> 00:33:18,340 have to make sure that they're compliant as well It's gonna do something. 666 00:33:18,340 --> 00:33:22,750 I think it's just frankly it's too early to tell How much of an impact it's going 667 00:33:22,750 --> 00:33:28,430 to have one thing that I that I anticipate happening could be false The vast majority 668 00:33:28,430 --> 00:33:33,720 of owners of this kind of these kinds of problems within organizations are on the C 669 00:33:33,720 --> 00:33:35,182 suite on the tech side of the house 670 00:33:35,182 --> 00:33:38,952 So a chief data officer, a chief analytics officer, sometimes a chief information 671 00:33:38,952 --> 00:33:43,892 security officer or a chief information officer, chief digital officer, blah, 672 00:33:43,892 --> 00:33:44,822 blah, blah, blah, blah. 673 00:33:44,822 --> 00:33:48,242 It's the risk and compliance people, not so much involved. 674 00:33:48,242 --> 00:33:50,122 We ask, we say, look, it's great. 675 00:33:50,122 --> 00:33:51,282 We need to bring them to the table. 676 00:33:51,282 --> 00:33:52,842 We need them in these conversations. 677 00:33:53,242 --> 00:33:55,622 Legal is pretty standard to say, well, bring them to the table. 678 00:33:55,622 --> 00:33:59,262 But usually the risk and compliance people are not spearheading it. 679 00:33:59,262 --> 00:34:03,918 So I think operationally we'll probably see a change where it gets. 680 00:34:03,918 --> 00:34:09,868 less in the hands of tech people and more in the hands of those risk and compliance 681 00:34:09,868 --> 00:34:13,768 people, or at least if it's less in the hands of the data people, the risk and 682 00:34:13,768 --> 00:34:16,278 compliance people will have to get more involved because that's their job. 683 00:34:16,278 --> 00:34:17,938 So I think we'll see more of that. 684 00:34:17,938 --> 00:34:22,738 That in principle will be helpful because those chief data officers, what do they 685 00:34:22,738 --> 00:34:23,828 know about risk and compliance? 686 00:34:23,828 --> 00:34:25,258 They're not risk and compliance folks. 687 00:34:25,258 --> 00:34:29,698 They're engineers, they're data scientists, they build cool shit. 688 00:34:29,778 --> 00:34:31,118 That's not what they do. 689 00:34:31,118 --> 00:34:33,578 And so I think that having... 690 00:34:33,614 --> 00:34:38,124 trained risk and compliance people probably owning the program or at least 691 00:34:38,124 --> 00:34:42,174 playing a significant role in it will increase adoption. 692 00:34:42,334 --> 00:34:44,294 Will it be mere box checking? 693 00:34:45,174 --> 00:34:47,534 I don't know. 694 00:34:47,534 --> 00:34:49,814 Probably the beginning is my guess. 695 00:34:49,814 --> 00:34:52,424 There's also a way in which companies tend to overestimate what they're actually 696 00:34:52,424 --> 00:34:53,254 doing. 697 00:34:53,254 --> 00:34:56,074 So we'll ask companies, oh, so are you doing X, Y, and oh yeah, yeah, yeah. 698 00:34:56,074 --> 00:34:56,904 You're checking for bias? 699 00:34:56,904 --> 00:34:57,554 Yeah, yeah, yeah, yeah. 700 00:34:57,554 --> 00:34:58,594 We check for bias. 701 00:34:58,594 --> 00:35:02,554 And then we sort of start digging around and we find out, oh, not really. 702 00:35:02,554 --> 00:35:03,310 I mean like... 703 00:35:03,310 --> 00:35:05,910 The thing that they're doing is so minimal. 704 00:35:05,910 --> 00:35:10,290 And so there's also a level of self -deception, or at least ignorance, about 705 00:35:10,290 --> 00:35:12,990 how robust what they're doing actually is. 706 00:35:12,990 --> 00:35:17,950 So, you know, I'm optimistic, but it's a cautious optimism. 707 00:35:17,950 --> 00:35:20,504 I wouldn't be shocked if it turned out to be box -checking. 708 00:35:21,840 --> 00:35:26,670 And is there anything else that you see potentially coming in the next few years 709 00:35:26,670 --> 00:35:30,250 that's going to have a major impact on AI? 710 00:35:31,182 --> 00:35:33,662 I mean, there's two kinds of things, broadly speaking, that could have the 711 00:35:33,662 --> 00:35:34,402 impact. 712 00:35:34,402 --> 00:35:35,982 The, um, and they're, they're compatible. 713 00:35:35,982 --> 00:35:40,022 One is a big disaster. 714 00:35:40,022 --> 00:35:42,682 Think of like Cambridge Analytica, but for AI. 715 00:35:42,682 --> 00:35:45,542 And so if you get that kind of thing, that will help. 716 00:35:45,542 --> 00:35:47,502 And, uh, help. 717 00:35:47,502 --> 00:35:49,062 I mean, you know, that'll be the silver lining. 718 00:35:49,062 --> 00:35:52,342 That'll be a cause of people thinking more seriously. 719 00:35:52,522 --> 00:35:53,662 That might happen. 720 00:35:53,662 --> 00:35:54,882 It wouldn't surprise me. 721 00:35:54,882 --> 00:35:58,062 But the other thing is just, you know, death by a thousand cuts. 722 00:35:58,062 --> 00:35:58,990 It's so. 723 00:35:58,990 --> 00:36:02,400 there's this AI incident and then this one and then this one and then this one and 724 00:36:02,400 --> 00:36:06,870 this one, neither one of which is a massive crisis, but the aggregate should 725 00:36:06,870 --> 00:36:10,110 be terrifying, I think, eventually to people in the C -suite and the board 726 00:36:10,110 --> 00:36:12,730 responsible for protecting the brand or their organization. 727 00:36:12,730 --> 00:36:17,290 So I just think it's going to be the case that we're seeing increased adoption and 728 00:36:17,290 --> 00:36:20,560 as you get increased adoption, you get increased deployment and increased 729 00:36:20,560 --> 00:36:25,450 deployment means increased probability of things going ethically sideways. 730 00:36:25,530 --> 00:36:27,790 And so you get more things going ethically sideways and they're going to be 731 00:36:27,790 --> 00:36:29,222 published, et cetera, et cetera. 732 00:36:29,454 --> 00:36:33,894 On that note, by the way, increased adoption, there are some people who say, 733 00:36:34,454 --> 00:36:38,414 oh, generative AI, it's been a year already, or it's been a year and two 734 00:36:38,414 --> 00:36:39,134 months, whatever. 735 00:36:39,134 --> 00:36:41,614 We haven't seen the massive change yet. 736 00:36:41,614 --> 00:36:44,034 And I just think, just chill, just hold on. 737 00:36:44,034 --> 00:36:46,494 We're talking about the most cutting edge technology out there. 738 00:36:46,494 --> 00:36:47,504 They're working on it. 739 00:36:47,504 --> 00:36:49,034 They're figuring it out. 740 00:36:49,194 --> 00:36:55,654 Most of our clients are running dozens, and in some cases, hundreds of pilots 741 00:36:55,654 --> 00:36:56,854 using generative AI. 742 00:36:56,854 --> 00:36:58,264 They're just trying to figure out how this thing works. 743 00:36:58,264 --> 00:36:59,118 They're trying to figure out. 744 00:36:59,118 --> 00:37:00,698 Does it solve a business problem for us? 745 00:37:00,698 --> 00:37:01,878 What business problem does it solve? 746 00:37:01,878 --> 00:37:03,498 How do we use it responsibly? 747 00:37:03,498 --> 00:37:07,998 So, you know, for all the talk about, oh, it's been, it's out here. 748 00:37:08,038 --> 00:37:09,758 No, not really. 749 00:37:09,758 --> 00:37:12,268 Open AI released their stuff a year and a bit ago. 750 00:37:12,268 --> 00:37:14,998 Google a little bit less than that. 751 00:37:15,678 --> 00:37:18,818 Then there's all these other, you know, llama and anthropic. 752 00:37:19,298 --> 00:37:20,068 Just wait a second. 753 00:37:20,068 --> 00:37:27,008 We're going to see, I would, unless things go oddly haywire, we're going to see lots 754 00:37:27,008 --> 00:37:28,866 of adoption, lots of deployment. 755 00:37:28,910 --> 00:37:32,230 in 2024, let alone 2025. 756 00:37:32,230 --> 00:37:36,278 And with the increased adoption, we're going to see more bad incidents. 757 00:37:37,008 --> 00:37:41,528 All right, well, more bad incidents, but hopefully also more good work for people 758 00:37:41,528 --> 00:37:42,472 like you. 759 00:37:42,766 --> 00:37:46,816 Yeah, I mean, I should say, you know, this comes against the backdrop of being, I 760 00:37:46,816 --> 00:37:48,976 don't know, not a pessimist about the technology. 761 00:37:48,976 --> 00:37:49,836 I'm not a catastrophist. 762 00:37:49,836 --> 00:37:52,996 I mean, there are some people out there who are screaming that the sky is falling 763 00:37:52,996 --> 00:37:57,506 and everyone's going to die and this is a disaster and it's not ready for prime time 764 00:37:57,506 --> 00:37:58,746 and it never will be. 765 00:37:58,746 --> 00:38:00,206 I just think that's sort of stupid. 766 00:38:00,206 --> 00:38:03,326 I mean, I've seen use cases that work and they do amazing things. 767 00:38:03,766 --> 00:38:07,976 There are cases, I was recently talking to an organization yesterday that said they 768 00:38:07,976 --> 00:38:11,726 implemented a generative AI solution for certain kinds of... 769 00:38:11,726 --> 00:38:16,266 search and summarization of a process that used to take 40 hours and now it takes 770 00:38:16,266 --> 00:38:17,326 four. 771 00:38:17,366 --> 00:38:18,646 And they are happy with the results. 772 00:38:18,646 --> 00:38:19,616 They're happy with the metrics. 773 00:38:19,616 --> 00:38:21,786 They're happy with how it's level of accuracy. 774 00:38:21,786 --> 00:38:24,506 They're happy with the level of risk that there is in the particular use case that 775 00:38:24,506 --> 00:38:25,646 it is. 776 00:38:26,046 --> 00:38:27,176 Just give it, there'll be more. 777 00:38:27,176 --> 00:38:30,526 That's a well -funded, you know, organization with lot who has lots of 778 00:38:30,526 --> 00:38:32,506 power to throw out these kinds of problems. 779 00:38:32,506 --> 00:38:34,866 Others don't have that kind of power or they're slower. 780 00:38:34,866 --> 00:38:36,626 They want to see what others are doing. 781 00:38:36,626 --> 00:38:38,066 But yeah, I'm not a catastrophist. 782 00:38:38,066 --> 00:38:41,390 I think that there's tremendous good opportunity. 783 00:38:41,390 --> 00:38:44,680 And yeah, there's lots of opportunity though to make sure for people like me 784 00:38:44,680 --> 00:38:47,612 that to help organizations make sure they don't go off the rails. 785 00:38:48,528 --> 00:38:50,158 Reed, thanks so much for this conversation. 786 00:38:50,158 --> 00:38:51,978 It's been really instructive and fascinating. 787 00:38:52,026 --> 00:38:52,962 my pleasure.