1 00:00:00,040 --> 00:00:07,120 Today on CXO Talk episode 882, we're examining a critical issue 2 00:00:07,120 --> 00:00:11,600 that affects millions, yet is largely invisible to most 3 00:00:11,600 --> 00:00:14,840 business leaders. How AI is failing our most 4 00:00:14,840 --> 00:00:18,200 vulnerable citizens on an unprecedented scale. 5 00:00:18,720 --> 00:00:22,440 I'm Michael Krigsman, and I'm delighted to welcome our guest, 6 00:00:22,440 --> 00:00:26,320 Kevin Delebin. As founder of Tectonic Justice, 7 00:00:26,480 --> 00:00:30,080 Kevin has witnessed first hand how AI systems determine who 8 00:00:30,080 --> 00:00:34,240 gets healthcare, who finds housing, who gets hired, and who 9 00:00:34,240 --> 00:00:38,360 receives government benefits. His ground breaking report, 10 00:00:38,360 --> 00:00:43,880 Inescapable AI, reveals that 92 million Americans now have 11 00:00:43,880 --> 00:00:47,960 fundamental aspects of their lives decided by algorithms, 12 00:00:48,240 --> 00:00:50,880 often with devastating consequences. 13 00:00:51,360 --> 00:00:55,160 He'll share how these systems fail, why traditional 14 00:00:55,160 --> 00:00:59,520 accountability mechanisms don't work, and what this means for 15 00:00:59,520 --> 00:01:03,200 your organization. We're discussing real systems 16 00:01:03,400 --> 00:01:08,560 built by major vendors causing real harm right now. 17 00:01:08,800 --> 00:01:11,880 Tectonic Justice is a new nonprofit organization launched 18 00:01:11,880 --> 00:01:16,680 last November to protect low income people from the harms 19 00:01:17,000 --> 00:01:21,720 that AI causes them. And I come to this work after 12 20 00:01:21,720 --> 00:01:25,000 years as a legal aid attorney representing low income folks in 21 00:01:25,000 --> 00:01:26,560 all sorts of civil legal matters. 22 00:01:26,760 --> 00:01:29,680 And it was there that I first saw the ways that these 23 00:01:29,680 --> 00:01:32,080 technologies were hurting my clients lives. 24 00:01:32,400 --> 00:01:36,800 And I was involved in several battles and won one several of 25 00:01:36,800 --> 00:01:39,160 them as well, and started understanding that this was a 26 00:01:39,160 --> 00:01:41,520 bigger problem that needed more attention and more focus. 27 00:01:41,960 --> 00:01:46,520 Kevin, you were an early pioneer actually winning cases relating 28 00:01:46,520 --> 00:01:51,520 to AI harms. It was really in 2016 when I had 29 00:01:51,520 --> 00:01:55,640 clients who were disabled or elderly on a Medicaid program 30 00:01:55,880 --> 00:01:59,440 that pays for an in home caregiver to help them with 31 00:01:59,440 --> 00:02:02,000 their daily life activities so that they can stay out of a 32 00:02:02,000 --> 00:02:05,280 nursing facility. And this is better for their 33 00:02:05,280 --> 00:02:07,680 dignity and independence and generally cheaper for the state 34 00:02:07,680 --> 00:02:09,759 as well. And what happened is the state 35 00:02:09,759 --> 00:02:15,400 of Arkansas replaced the nurse's discretion to decide how many 36 00:02:15,400 --> 00:02:19,240 hours a day of care a particular person needed with an 37 00:02:19,240 --> 00:02:21,800 algorithmic decision making tool, and that ended up 38 00:02:21,960 --> 00:02:25,160 devastating people's lives. People's care hours were cut in 39 00:02:25,160 --> 00:02:28,960 half in some cases, and it left people lying in their own waste, 40 00:02:29,240 --> 00:02:32,120 left people getting bed sores from not being turned, being 41 00:02:32,120 --> 00:02:34,480 totally shut in just intolerable human suffering. 42 00:02:34,680 --> 00:02:37,240 And we ended up fighting against that in the courts and also with 43 00:02:37,240 --> 00:02:40,160 a really big public education and kind of community activation 44 00:02:40,160 --> 00:02:41,640 campaign. And we won. 45 00:02:41,920 --> 00:02:45,680 And that's one of the relatively few examples still to this day 46 00:02:46,120 --> 00:02:49,480 of kind of successful advocacy against sort of automated 47 00:02:49,480 --> 00:02:54,560 decision making. Algorithms, AI are neutral 48 00:02:54,800 --> 00:02:59,040 mechanisms, neutral devices. They're just math without 49 00:02:59,200 --> 00:03:02,960 feelings, without interests, without malice. 50 00:03:03,360 --> 00:03:08,440 So given that, what is the problem here? 51 00:03:08,840 --> 00:03:12,400 I would challenge some of the the the assumptions even in that 52 00:03:12,400 --> 00:03:13,960 question of them being neutral, right? 53 00:03:13,960 --> 00:03:15,840 I mean, they're programmed by humans. 54 00:03:15,840 --> 00:03:19,440 The statistical science that underlies a lot of this stuff is 55 00:03:19,560 --> 00:03:22,440 determined by humans using various choices that they have, 56 00:03:22,440 --> 00:03:24,120 using historical data that they have. 57 00:03:25,080 --> 00:03:28,560 And that isn't a wholly objective exercise. 58 00:03:28,840 --> 00:03:31,400 And so I think what you really have to look at is the purpose 59 00:03:31,400 --> 00:03:35,680 for which the technology is being built to understand it and 60 00:03:35,680 --> 00:03:38,160 understand a lot of like even the technical aspects that 61 00:03:38,160 --> 00:03:40,400 underpin it. And in my world, when we're 62 00:03:40,400 --> 00:03:43,360 talking about low income people and sort of automated decision 63 00:03:43,360 --> 00:03:46,800 making for them, these are not neutral technologies at all. 64 00:03:46,800 --> 00:03:50,760 These are designed oftentimes to restrict access to benefits or 65 00:03:50,760 --> 00:03:54,280 to empower whoever's on the other side of the equation, 66 00:03:54,280 --> 00:03:57,120 whether it's a landlord, a boss, a school principal, a government 67 00:03:57,120 --> 00:04:00,480 official to do something that they want done. 68 00:04:00,680 --> 00:04:05,160 That might not be what the person who I'm representing is 69 00:04:05,160 --> 00:04:07,520 interested in. So I would challenge that 70 00:04:07,520 --> 00:04:10,560 premise first. So you're saying that the design 71 00:04:10,680 --> 00:04:14,960 of the system is intended to cause harm. 72 00:04:14,960 --> 00:04:16,600 Is that correct what I'm hearing you say? 73 00:04:16,880 --> 00:04:19,640 In some cases, it's intended outright to cause harm. 74 00:04:19,640 --> 00:04:25,200 In some cases, it's just intended to, you know, sort of 75 00:04:25,200 --> 00:04:28,840 facilitate a decision by the decision maker, right? 76 00:04:28,840 --> 00:04:32,840 Make a landlord's life easier, make a boss's life easier, make 77 00:04:32,840 --> 00:04:34,480 a government official's life easier. 78 00:04:34,840 --> 00:04:38,520 The problem is inherent in making their life easier ends up 79 00:04:38,520 --> 00:04:41,480 being making somebody else's life harder. 80 00:04:41,720 --> 00:04:45,560 And so I think that's where the push and pull is of this is 81 00:04:45,560 --> 00:04:48,920 there is the intent issue. There are very clearly stuff 82 00:04:48,920 --> 00:04:51,560 that's built to be harmful. But then there's also this Gray 83 00:04:51,560 --> 00:04:55,720 area where nobody is, you know, scheming in a dark room about 84 00:04:55,800 --> 00:04:58,480 plotting to take over the world and destroy people's lives. 85 00:04:58,720 --> 00:05:01,880 But the nature of their power positions and the decisions that 86 00:05:01,880 --> 00:05:04,760 they're making and what makes their life easier ends up 87 00:05:04,760 --> 00:05:07,040 translating into that for low income people. 88 00:05:07,480 --> 00:05:12,600 Can you give us an example of where the the goals or the 89 00:05:12,720 --> 00:05:17,200 incentives are misaligned between the developers of this 90 00:05:17,200 --> 00:05:22,040 technology, of these technologies or algorithms and 91 00:05:22,040 --> 00:05:25,200 the can we say the intended recipients? 92 00:05:25,200 --> 00:05:27,320 Is that even a a correct way to phrase it? 93 00:05:27,760 --> 00:05:31,040 You have the hiring process for example with most big companies 94 00:05:31,040 --> 00:05:35,240 now is riddled with AI. Everything from resume, resume 95 00:05:35,280 --> 00:05:42,760 review and screening to video interviewing and to oversight. 96 00:05:42,760 --> 00:05:46,320 Once somebody gets the job, right, there's nothing inherent 97 00:05:46,320 --> 00:05:48,520 in that process that really benefits the person who's 98 00:05:48,520 --> 00:05:51,160 seeking work or is an employee, right? 99 00:05:51,160 --> 00:05:55,360 That's all intended to facilitate the life and the work 100 00:05:55,360 --> 00:05:58,080 of the employer. The bosses, same thing a lot of 101 00:05:58,080 --> 00:06:03,480 times with, you know, with public benefits, you've got 102 00:06:04,200 --> 00:06:07,640 relatively, you've got really dedicated public servants, but 103 00:06:07,640 --> 00:06:09,760 oftentimes they're unsophisticated in technology 104 00:06:09,760 --> 00:06:12,200 issues. They're thinking, OK, well, this 105 00:06:12,200 --> 00:06:15,760 new piece of technology is going to suddenly help expand our 106 00:06:15,760 --> 00:06:18,160 limited capacity, so let's implement it. 107 00:06:18,480 --> 00:06:23,000 And then they don't have what they need to do that in a non 108 00:06:23,000 --> 00:06:25,520 destructive way. And so the people who end up 109 00:06:25,600 --> 00:06:29,200 bearing the risk of, you know, sort of their own lack of 110 00:06:29,200 --> 00:06:31,360 knowledge or incompetence are the low income people that are 111 00:06:31,360 --> 00:06:35,720 subject to the decision making. These systems are complex. 112 00:06:35,720 --> 00:06:43,440 They are developed with algorithms and data as well. 113 00:06:44,800 --> 00:06:51,560 Can you isolate where a a primary source of the problem 114 00:06:51,560 --> 00:06:53,600 lies? I realized underneath it all, 115 00:06:53,600 --> 00:06:57,560 you have a human intention, trying to solve a problem, 116 00:06:57,560 --> 00:07:00,560 trying to achieve a goal. But if you can, you drill into 117 00:07:00,560 --> 00:07:02,600 that a little bit, kind of dissect this for us. 118 00:07:03,080 --> 00:07:06,160 There were a couple aspects to the way the algorithm worked. 119 00:07:07,560 --> 00:07:09,280 One is the mechanics of it right? 120 00:07:09,280 --> 00:07:11,720 What inputs turn into what outputs? 121 00:07:13,080 --> 00:07:16,120 And that's hard enough to discern, but then there's the 122 00:07:16,120 --> 00:07:20,040 reason that those inputs are chosen to lead to those outputs, 123 00:07:20,880 --> 00:07:22,560 right? Like why do you look at this 124 00:07:22,560 --> 00:07:26,200 factor and not this factor? Why is this factor shaped to 125 00:07:26,200 --> 00:07:29,160 look back three days instead of five days? 126 00:07:29,720 --> 00:07:32,840 All of those things, those are all human decisions Now. 127 00:07:32,840 --> 00:07:37,080 They're informed by, in the best cases, statistical science. 128 00:07:38,200 --> 00:07:43,120 In a lot of cases, there is no science is a bad descriptor for 129 00:07:43,120 --> 00:07:44,840 that. A lot of times it's junk, right, 130 00:07:44,840 --> 00:07:46,960 that somebody just invented and comes up with. 131 00:07:47,440 --> 00:07:49,920 But in the best cases in statistical science, that still 132 00:07:49,920 --> 00:07:51,680 is riddled with various assumptions. 133 00:07:52,280 --> 00:07:55,640 And so in our example, for example, in Arkansas, for 134 00:07:55,640 --> 00:08:07,160 example, whether or not somebody could use the bath on their own 135 00:08:07,160 --> 00:08:10,000 might not have been a factor that the algorithm considered. 136 00:08:10,280 --> 00:08:12,440 And that's weird, right? I mean, we're talking about home 137 00:08:12,440 --> 00:08:14,760 care for elderly or disabled people. 138 00:08:15,120 --> 00:08:17,880 Being able to bathe on your own should be 1 factor that decides 139 00:08:17,880 --> 00:08:21,200 how many hours of care you need. It wasn't or your ability to 140 00:08:21,200 --> 00:08:25,280 prepare meals wasn't a factor. And so you see this disconnect 141 00:08:25,280 --> 00:08:29,560 of like, we know instinctively or, you know, through medical 142 00:08:29,560 --> 00:08:33,799 discretion and judgment, how to answer this question of how much 143 00:08:33,799 --> 00:08:36,360 care somebody needs. It might be imprecise, but we 144 00:08:36,360 --> 00:08:38,400 know we know what we should be looking at. 145 00:08:38,520 --> 00:08:40,120 But the algorithm didn't do that. 146 00:08:40,120 --> 00:08:43,400 They looked at a lot of factors that weren't kind of intuitive, 147 00:08:43,679 --> 00:08:47,800 and then they ignored a lot of factors that were intuitive. 148 00:08:48,120 --> 00:08:51,240 How does this come about? Is it simply lack of 149 00:08:51,240 --> 00:08:57,480 understanding of the the subjects of the the target? 150 00:08:57,680 --> 00:08:58,760 How did? What happens here? 151 00:08:59,120 --> 00:09:03,280 Some of it is real ignorance about, you know, the lives of 152 00:09:03,280 --> 00:09:06,760 poor people and the ways that decisions are made and the 153 00:09:06,760 --> 00:09:10,400 impact of the decisions. Some is ignorance about certain 154 00:09:10,400 --> 00:09:13,080 program standards or laws or anything else. 155 00:09:13,080 --> 00:09:14,720 I've seen that a lot in the technology. 156 00:09:15,280 --> 00:09:18,800 Some of it is the lack of having to get it right. 157 00:09:19,760 --> 00:09:22,440 You know, for a lot of the developers of these algorithms 158 00:09:22,440 --> 00:09:25,360 in particular, they're shielded from any sort of consequences 159 00:09:26,040 --> 00:09:30,600 for their actions. And so they do what they think 160 00:09:30,600 --> 00:09:34,400 is best or what they can sell to a client and that's that. 161 00:09:34,400 --> 00:09:36,840 And then the clients that are using it, the government 162 00:09:36,840 --> 00:09:40,760 agencies or the employers or whatever, they might not be 163 00:09:40,760 --> 00:09:44,200 vetting it or, you know, they are also insulated from the 164 00:09:44,200 --> 00:09:48,520 accountability because if it hurts poor people, what's going 165 00:09:48,520 --> 00:09:50,600 to happen to them? Like what's going to happen to 166 00:09:50,600 --> 00:09:52,320 the person who decided to use it? 167 00:09:52,560 --> 00:09:55,600 I mean, poor people often times are not a particularly empowered 168 00:09:55,600 --> 00:09:57,920 political bloc. They're usually aren't scandals 169 00:09:57,920 --> 00:10:01,840 that end up resulting in lost jobs or lost elections for 170 00:10:02,080 --> 00:10:03,920 officials who are in charge of this stuff. 171 00:10:04,840 --> 00:10:09,520 And so it's easy to get away with really harmful actions just 172 00:10:09,520 --> 00:10:12,000 because you're doing it to people who don't have a lot of 173 00:10:13,120 --> 00:10:16,640 power that's ready at hand. You know, low income communities 174 00:10:16,640 --> 00:10:20,960 have always been super involved in advocating for themselves and 175 00:10:20,960 --> 00:10:24,120 organizing and everything else. But that's a huge effort, right? 176 00:10:24,120 --> 00:10:26,200 And takes like a concentrated movement. 177 00:10:26,440 --> 00:10:29,480 And it's not like you can just call your elected official and 178 00:10:29,480 --> 00:10:31,800 you have that kind of access and say, hey, this is a problem. 179 00:10:31,800 --> 00:10:32,760 Can you take care of this for me? 180 00:10:32,760 --> 00:10:35,240 Or organize a lobbying effort to get rid of something? 181 00:10:35,240 --> 00:10:38,200 Now, if you're doing something with poor people and it hurts 182 00:10:38,200 --> 00:10:40,560 them, you're not going to face immediate consequences for the 183 00:10:40,560 --> 00:10:44,040 most part. Folks who are listening, I want 184 00:10:44,040 --> 00:10:48,080 you to ask questions. We have some questions that are 185 00:10:48,080 --> 00:10:51,200 starting to come in on LinkedIn and Twitter and we're going to, 186 00:10:51,200 --> 00:10:53,960 we're going to get to them in a in a couple of minutes. 187 00:10:55,360 --> 00:11:00,560 If you're watching on Twitter, just insert your questions into 188 00:11:00,560 --> 00:11:03,000 Twitter using the hashtag CXO Talk. 189 00:11:03,000 --> 00:11:06,040 If you're watching on LinkedIn, just pop your questions into the 190 00:11:06,040 --> 00:11:08,560 chat. For those of you who are 191 00:11:08,560 --> 00:11:13,840 developing these kinds of systems and we hear a lot of 192 00:11:13,840 --> 00:11:18,680 discussion of ethical AI and responsible technology, here's 193 00:11:18,680 --> 00:11:23,120 an opportunity to ask somebody who's dealing with the actual 194 00:11:23,120 --> 00:11:28,440 fallout of this. So ask your questions, Kevin, 195 00:11:28,440 --> 00:11:30,600 what about the scale of the problem? 196 00:11:30,880 --> 00:11:32,840 How? How big an issue is this 197 00:11:32,920 --> 00:11:35,800 actually? All 92 million low income people 198 00:11:35,800 --> 00:11:38,840 in the United States have some key aspect of their life decided 199 00:11:38,840 --> 00:11:43,680 by AI, whether that is sort of housing, healthcare, public 200 00:11:43,680 --> 00:11:47,200 benefits, work, their kids, school, family stability, all of 201 00:11:47,200 --> 00:11:50,280 these issues. Not everyone might have all of 202 00:11:50,280 --> 00:11:53,320 those issues decided by AI, but everyone has at least one of 203 00:11:53,320 --> 00:11:58,280 those issues decided by AI. And then it extends beyond low 204 00:11:58,280 --> 00:12:01,360 income people as well into higher income things. 205 00:12:01,360 --> 00:12:03,600 So there have been a lot of stories, for example, about 206 00:12:03,880 --> 00:12:07,640 employer use of AI, sort of the screening aspect and then sort 207 00:12:07,640 --> 00:12:11,960 of the bossware management aspect of it being used against 208 00:12:12,360 --> 00:12:16,600 finance executives or against hospital chaplains, against 209 00:12:16,880 --> 00:12:19,840 therapists. Recently there was a story about 210 00:12:19,840 --> 00:12:24,800 Amazon programmers who are now subjected to AI based oversight 211 00:12:24,800 --> 00:12:26,880 and measurement and it's affecting their lives. 212 00:12:26,880 --> 00:12:29,640 So even though a lot of this stuff is most prevalent and 213 00:12:29,640 --> 00:12:32,720 probably most severe in the lives of low income people, it's 214 00:12:32,720 --> 00:12:35,440 happening to all of us. Healthcare is another great 215 00:12:35,440 --> 00:12:37,520 example, right? If our doctor recommends a 216 00:12:37,520 --> 00:12:41,080 treatment for us, many of the more expensive treatments are 217 00:12:41,080 --> 00:12:43,880 subject to health insurance company review prior to being 218 00:12:43,880 --> 00:12:46,400 offered, and those health insurance companies are using AI 219 00:12:46,600 --> 00:12:49,440 generally to deny those requests. 220 00:12:50,000 --> 00:12:53,760 We all know about United Healthcare and their use of 221 00:12:53,760 --> 00:12:57,360 algorithms that they say is neutral. 222 00:12:57,360 --> 00:12:59,360 And we don't do that. But it's, you know, you hear 223 00:12:59,360 --> 00:13:05,280 doctors complaining about how algorithms are interfering with 224 00:13:05,280 --> 00:13:08,560 their ability to render the kind of care that they want. 225 00:13:08,880 --> 00:13:15,320 And so it becomes pretty evident that what what was once targeted 226 00:13:15,320 --> 00:13:21,640 at lower income people now through the acceleration of AI 227 00:13:21,640 --> 00:13:25,160 is broadening and touches all of us at this point, I would 228 00:13:25,160 --> 00:13:27,240 imagine. In one of the examples of the 229 00:13:27,240 --> 00:13:31,920 health insurance companies, they ostensibly had a human reviewer 230 00:13:32,200 --> 00:13:39,200 reviewing the the AI's outputs, but when the investigation. 231 00:13:39,960 --> 00:13:43,800 Dug into what that human review looked like, it showed that the 232 00:13:43,800 --> 00:13:46,320 doctor was approving something like 60 prior authorization 233 00:13:46,320 --> 00:13:49,040 requests a minute like they had one or two seconds per one. 234 00:13:49,680 --> 00:13:51,440 There's no human reviewing that right. 235 00:13:51,440 --> 00:13:53,520 And it's bad faith to assert otherwise. 236 00:13:53,840 --> 00:13:56,160 And that's I think one of the key data points and there. 237 00:13:56,160 --> 00:13:59,280 Are a lot of others that help us show that this isn't just all 238 00:13:59,280 --> 00:14:03,680 purely accidental, this can't be just attributed to mistakes or 239 00:14:03,680 --> 00:14:06,840 errors, that there's a lot of thought and intention that goes 240 00:14:06,840 --> 00:14:11,520 behind, you know, developing and implementing these systems that 241 00:14:11,520 --> 00:14:15,760 are denying people really fundamental needs. 242 00:14:16,080 --> 00:14:20,720 Subscribe to our newsletter, go to cxotalk.com, check out our 243 00:14:20,720 --> 00:14:24,320 newsletter, and check us out for our next shows. 244 00:14:24,320 --> 00:14:27,280 We have great shows coming up. Let's jump to some questions. 245 00:14:27,640 --> 00:14:31,560 Let's begin with Arsalan Khan on Twitter. 246 00:14:32,120 --> 00:14:35,040 Arsalan's a regular listener. Thanks so much, Arsalan, for 247 00:14:35,040 --> 00:14:39,440 your regular listenership. And Arsalan says this. 248 00:14:39,440 --> 00:14:46,320 He says, who whoever sets the AI guardrails has the power, but 249 00:14:46,320 --> 00:14:50,640 who checks if those guardrails are equitable? 250 00:14:50,640 --> 00:14:57,160 And he asks, why don't we have a Hippocratic Oath for us as IT 251 00:14:57,400 --> 00:14:59,480 professionals? He's, he's an enterprise 252 00:14:59,960 --> 00:15:00,920 architect. So. 253 00:15:00,920 --> 00:15:04,440 So this notion of whoever sets the AI guardrails has the power, 254 00:15:04,680 --> 00:15:07,600 but who checks that the guardrails are right are 255 00:15:07,600 --> 00:15:10,600 equitable? The Hippocratic Oath idea is not 256 00:15:10,760 --> 00:15:18,000 a meaningful source of systemic change to insulate society from 257 00:15:18,000 --> 00:15:24,800 these harms because, you know, doctors have Hippocratic oaths. 258 00:15:24,800 --> 00:15:27,680 And while that might be useful, it doesn't a lot of times 259 00:15:27,680 --> 00:15:30,840 prevent some of the abuses in in medicine either or lawyers have 260 00:15:30,840 --> 00:15:33,720 obligations and it doesn't present prevent us from going 261 00:15:33,720 --> 00:15:36,560 and doing all sorts of random harmful things. 262 00:15:36,760 --> 00:15:40,440 So I think what you need is actual regulation to reinforce 263 00:15:40,720 --> 00:15:42,800 kind of the guardrail notion, right? 264 00:15:42,960 --> 00:15:47,480 Safeguard people from having any exposure to the harms in the 1st 265 00:15:47,480 --> 00:15:51,400 place, or if there are, because those kind of institutional and 266 00:15:51,400 --> 00:15:57,040 ethical safeguards fail, then there are real consequences for 267 00:15:57,040 --> 00:15:58,800 that. They go beyond somebody just 268 00:15:58,800 --> 00:16:02,480 violating their oath and feeling bad that way. 269 00:16:02,840 --> 00:16:05,720 So I don't know if that's getting at the full essence of 270 00:16:05,720 --> 00:16:10,360 the question, but that's where some of my thoughts go. 271 00:16:10,600 --> 00:16:13,440 And also, not everybody's as ethical as the person asking the 272 00:16:13,440 --> 00:16:15,600 question either, right? And some people are perfectly 273 00:16:15,600 --> 00:16:18,600 happy to just do whatever the client wants or program the 274 00:16:18,600 --> 00:16:22,040 system in whatever way is going to make it most profitable and 275 00:16:22,040 --> 00:16:24,720 attractive. And as long as they don't have 276 00:16:24,720 --> 00:16:27,520 anything holding them back formally, officially real 277 00:16:27,520 --> 00:16:30,880 consequences and accountable, we're not going to get any major 278 00:16:30,880 --> 00:16:33,440 change. So self policing is not 279 00:16:33,440 --> 00:16:36,320 sufficient in your view. Definitely not. 280 00:16:36,320 --> 00:16:39,360 And I know those questions are asked in good faith and are 281 00:16:39,360 --> 00:16:43,280 posited in good faith. But the people who are pushing 282 00:16:43,280 --> 00:16:46,440 that at the policy level are definitely not pushing it in 283 00:16:46,440 --> 00:16:48,080 good faith. They don't want any 284 00:16:48,080 --> 00:16:49,920 accountability. They don't want anything that 285 00:16:49,920 --> 00:16:54,360 would restrict how they use it, and they're perfectly happy to 286 00:16:54,360 --> 00:16:58,480 shunt off all the risks and all the dangers of their systems 287 00:16:58,480 --> 00:17:02,520 being bad or going wrong or doing something destructive to 288 00:17:02,560 --> 00:17:04,480 the people who are subject to those decisions. 289 00:17:05,160 --> 00:17:09,079 Are you talking about government policy or in corporate policy, 290 00:17:09,079 --> 00:17:13,119 people designing products? Government policy, the tech 291 00:17:13,119 --> 00:17:18,000 industry has been vociferous in their opposition to any sort of 292 00:17:18,000 --> 00:17:23,800 meaningful regulation of AI, automated decision making 293 00:17:23,800 --> 00:17:28,200 technologies and so forth. And that's the reason why we 294 00:17:28,200 --> 00:17:31,440 don't have any real societal protections against this stuff 295 00:17:31,920 --> 00:17:36,000 outside of existing laws. And even now they're targeting 296 00:17:36,000 --> 00:17:40,520 some of the European Union's restrictions, which are modest, 297 00:17:40,520 --> 00:17:44,880 but big tech doesn't like those. So that's where I'm talking 298 00:17:44,880 --> 00:17:47,280 about is sort of how kind of corporate interests end up 299 00:17:47,280 --> 00:17:51,200 shaping their policy positions in ways that are detrimental to 300 00:17:51,640 --> 00:17:55,360 really all of us that are not in that in that world, but 301 00:17:55,360 --> 00:17:59,880 particularly low income people. You also have many of the major 302 00:18:00,040 --> 00:18:08,520 tech companies pushing forth their own ethical AI initiatives 303 00:18:08,520 --> 00:18:13,760 and lots of discussions around the data and creating building 304 00:18:13,760 --> 00:18:17,280 bodies of data that try to weed out bias. 305 00:18:17,520 --> 00:18:20,480 I mean that you see this everywhere is happening. 306 00:18:20,600 --> 00:18:21,680 That's true, and there are a lot of. 307 00:18:21,680 --> 00:18:26,240 Good people who share my values in these companies and are 308 00:18:26,240 --> 00:18:29,640 trying to make the companies do as right as possible. 309 00:18:30,200 --> 00:18:33,880 But I think when the rubber hits the road, we've seen repeatedly 310 00:18:34,400 --> 00:18:38,880 that the folks speaking out for ethical uses are sidelined. 311 00:18:38,920 --> 00:18:41,160 You know, a few years ago in Google, for example, the whole 312 00:18:41,160 --> 00:18:44,160 ethical AI team, I think, was fired because they wanted to 313 00:18:44,160 --> 00:18:46,480 publish a paper that Google didn't want published. 314 00:18:48,040 --> 00:18:53,280 Or more recently, when, you know, Twitter was taken over by 315 00:18:53,280 --> 00:18:57,120 its current owner, the whole ethical AI team was disbanded 316 00:18:57,120 --> 00:19:02,720 instantly. You have Google's retrenchment 317 00:19:02,720 --> 00:19:06,080 of its ethical AI things and now it's technology is being 318 00:19:06,080 --> 00:19:08,840 deployed in unemployment hearings, right, for people who 319 00:19:08,840 --> 00:19:12,600 are desperate for benefits, even though we know that a lot of the 320 00:19:13,120 --> 00:19:15,080 AI technology involved can be faulty. 321 00:19:15,280 --> 00:19:18,440 So again, you do have these ethical components within 322 00:19:18,440 --> 00:19:21,640 institutions that are pushing, I believe in good faith a lot of 323 00:19:21,640 --> 00:19:25,800 times for changes, but the people who are pushing for that 324 00:19:26,200 --> 00:19:28,520 don't have the same interests as the institutions who are 325 00:19:28,520 --> 00:19:30,280 allowing it. A lot of times the institutions 326 00:19:30,280 --> 00:19:32,600 are allowing ethical AI because it allows them to go out and 327 00:19:32,600 --> 00:19:35,920 talk about their their concept of social responsibility. 328 00:19:36,080 --> 00:19:39,240 But we we see repeatedly when the rubber hits the road, ethics 329 00:19:39,240 --> 00:19:42,640 will go by the wayside and the company's profit incentives and 330 00:19:42,640 --> 00:19:45,000 motives are going to be what dictates what happens next. 331 00:19:45,440 --> 00:19:47,520 So basically, money talks, nobody walks. 332 00:19:47,960 --> 00:19:49,320 Yeah, I mean it's. Complicated, right? 333 00:19:49,320 --> 00:19:51,240 Because, again, there's a lot of good people in there that are 334 00:19:51,240 --> 00:19:54,240 pushing really hard for these major institutions that have 335 00:19:54,240 --> 00:19:58,640 lots of power to do right. And the fact that the 336 00:19:58,640 --> 00:20:02,240 institutions allow that to happen is noteworthy. 337 00:20:02,440 --> 00:20:05,200 I think it just comes, yeah, at the end, it it, it, it ends up 338 00:20:05,200 --> 00:20:06,240 being the money. Talks. 339 00:20:06,680 --> 00:20:12,040 I will say that you are up against the marketing budgets of 340 00:20:12,040 --> 00:20:14,880 some really, really large companies here. 341 00:20:16,360 --> 00:20:17,960 I am. This is going to change 342 00:20:17,960 --> 00:20:19,680 everything though, Michael. See CXO talk. 343 00:20:19,680 --> 00:20:21,400 This is going to be, this is going to be the entryway. 344 00:20:21,400 --> 00:20:23,840 This is going to this is better than all the all the marketing 345 00:20:23,840 --> 00:20:26,360 budgets of of the big tech companies right now. 346 00:20:26,760 --> 00:20:30,480 Let's jump to some other questions. 347 00:20:30,800 --> 00:20:34,760 And I'm seeing some themes developing in the questions 348 00:20:34,760 --> 00:20:37,600 here. And this next one is from Preeti 349 00:20:38,080 --> 00:20:42,120 Narayanan. And she says, given your work 350 00:20:42,120 --> 00:20:47,840 exposing large scale harm caused by AI and public services, what 351 00:20:47,840 --> 00:20:51,960 practical guardrails would you recommend to technology leaders 352 00:20:51,960 --> 00:20:55,280 like her, Like many of our listeners who are building 353 00:20:55,280 --> 00:21:01,280 enterprise AI systems so we don't unknowingly replicate 354 00:21:01,440 --> 00:21:06,680 those same failures at scale? Basically, it's the same 355 00:21:06,920 --> 00:21:11,240 sentiment as Arsalan Khan just brought up. 356 00:21:11,680 --> 00:21:15,320 What can we, as the people creating these systems, do? 357 00:21:15,800 --> 00:21:19,240 OK, one thing is push for regulation, right? 358 00:21:19,280 --> 00:21:21,600 And push for meaningful regulation of what it is that 359 00:21:21,600 --> 00:21:25,320 you do, because that way it bakes in consequences for 360 00:21:25,320 --> 00:21:27,760 getting it wrong. And as long as you have good 361 00:21:27,760 --> 00:21:31,640 faith and are doing things the right way, those consequences 362 00:21:31,640 --> 00:21:33,960 shouldn't be terribly severe. You shouldn't be exposed to them 363 00:21:33,960 --> 00:21:37,120 that you know, in in a way that's wholly destructive. 364 00:21:37,120 --> 00:21:39,800 So I think pushing for regulation is actually in your 365 00:21:39,800 --> 00:21:44,760 own interest, but kind of in the context of developing a 366 00:21:44,760 --> 00:21:46,720 particular product. You can ask, is this a 367 00:21:46,720 --> 00:21:49,280 legitimate use for AI? For example, should we be using 368 00:21:49,320 --> 00:21:54,160 AI to deny people, disabled people benefits and home care? 369 00:21:54,400 --> 00:21:56,400 That might not be a legitimate use of AI. 370 00:21:56,400 --> 00:21:58,560 And if it isn't a legitimate use, maybe we shouldn't do it 371 00:21:58,560 --> 00:22:00,440 and we should just say that's off, off limits. 372 00:22:00,440 --> 00:22:02,360 We're not going to do that no matter how much somebody's going 373 00:22:02,360 --> 00:22:04,440 to pay us because we just don't believe that's fair. 374 00:22:04,760 --> 00:22:07,360 Now, if it is a legitimate use, and I acknowledge there's a lot 375 00:22:07,360 --> 00:22:10,920 of kind of Gray areas in this, then you've got to have a really 376 00:22:10,920 --> 00:22:14,520 intensive development and vetting process. 377 00:22:14,520 --> 00:22:16,640 What are you doing? What are you, what data are you 378 00:22:16,640 --> 00:22:18,160 using? Are you projecting out the 379 00:22:18,160 --> 00:22:21,880 harms? Are you consulting in a 380 00:22:21,880 --> 00:22:23,600 meaningful way with actual oversight? 381 00:22:23,600 --> 00:22:26,200 The people who are going to be subjected to these decisions, do 382 00:22:26,200 --> 00:22:29,080 they have some sort of say in how it's developed in a way that 383 00:22:29,080 --> 00:22:31,480 would actually stop you from moving forward or force a 384 00:22:31,480 --> 00:22:34,160 different development of it? Are you willing to disclose 385 00:22:35,040 --> 00:22:37,320 things that might traditionally be considered a trade secrets or 386 00:22:37,320 --> 00:22:40,800 intellectual property in the interests of having more public 387 00:22:40,800 --> 00:22:45,960 accountability? Are you willing to ensure 388 00:22:45,960 --> 00:22:50,080 ongoing oversight so that if your product is developed or is 389 00:22:50,680 --> 00:22:56,160 is deployed, it's deployed first of all in narrow, short phased 390 00:22:56,160 --> 00:22:58,840 ways so that we can test the harm before it's applied to 391 00:22:58,840 --> 00:23:00,800 everybody? And then two, are we willing to 392 00:23:00,800 --> 00:23:04,160 look over time in a three month span and see, hey, does our 393 00:23:04,160 --> 00:23:07,920 projected impact which we have documented and have disclosed to 394 00:23:07,920 --> 00:23:10,440 the public differ from what the actual impact is? 395 00:23:10,640 --> 00:23:12,440 And if so, is there an automatic off switch? 396 00:23:12,440 --> 00:23:14,600 Is there some way to to course correct that? 397 00:23:15,680 --> 00:23:20,760 And all of those things, when combined with meaningful, you 398 00:23:20,760 --> 00:23:23,040 know, legislation that means that their people have 399 00:23:23,040 --> 00:23:27,520 enforceable rights if they're hurt by it, would lead to reduce 400 00:23:27,520 --> 00:23:31,080 chances of harms on systemic society wide scales. 401 00:23:31,480 --> 00:23:37,920 If I were a corporate leader, you made the assertion that we 402 00:23:37,920 --> 00:23:42,720 should question whether AI is the appropriate decision making 403 00:23:42,720 --> 00:23:47,160 tool to use in some of these situations that could Causeway 404 00:23:47,160 --> 00:23:51,440 real downstream harms. But I would push back an I would 405 00:23:51,440 --> 00:23:56,880 say, Sir, you don't know what you're talking about because AI 406 00:23:56,880 --> 00:24:00,800 is a decision tool. It is not autonomous. 407 00:24:00,800 --> 00:24:04,760 It's overseen by humans. The data that we collect is 408 00:24:04,760 --> 00:24:09,840 carefully vetted to be biased. And it's unfortunate that these 409 00:24:09,840 --> 00:24:13,560 downstream harms are happening, but it's not a result of our 410 00:24:13,560 --> 00:24:16,480 decision making. There are systemic underlying 411 00:24:16,480 --> 00:24:22,200 societal issues and frankly, the AI is making the right decision. 412 00:24:22,560 --> 00:24:24,840 I would challenge almost everything that you said there, 413 00:24:24,840 --> 00:24:29,320 Michael, from, you know, the the sophistication of the vetting 414 00:24:29,320 --> 00:24:32,080 process. The people who are developing 415 00:24:32,080 --> 00:24:36,480 enterprise software might be doing a better job when the 416 00:24:36,480 --> 00:24:39,200 people are going to buy their software are wealthier or richer 417 00:24:39,200 --> 00:24:40,680 than when it works for low income people. 418 00:24:40,680 --> 00:24:43,920 So first of all, I think like who the audience is, who's going 419 00:24:43,920 --> 00:24:48,160 to be subjected to this dictates a lot of how careful the kind of 420 00:24:48,160 --> 00:24:50,320 development process is. And if it's going to be deployed 421 00:24:50,320 --> 00:24:52,880 against poor people, the development process doesn't need 422 00:24:52,880 --> 00:24:57,440 to be as intensive probably as it would be for corporate 423 00:24:57,440 --> 00:25:01,320 clients, right? So I think there's there's that. 424 00:25:01,320 --> 00:25:05,640 So a lot of the so-called science in AI is really junk 425 00:25:05,760 --> 00:25:09,400 when it applies to poor people. 1 great example of that is 426 00:25:11,560 --> 00:25:15,080 identity verification, for example, during the pandemic. 427 00:25:15,280 --> 00:25:17,480 And hopefully some of your listeners listeners will have 428 00:25:17,480 --> 00:25:20,040 some some frame of reference. But during the height of the 429 00:25:20,040 --> 00:25:22,720 pandemic, right, masses of people were unemployed. 430 00:25:23,640 --> 00:25:27,760 Congress expanded unemployment benefits to help people float 431 00:25:27,760 --> 00:25:30,640 during these, you know, desperate times. 432 00:25:31,200 --> 00:25:35,200 At some point, states, encouraged by the federal 433 00:25:35,200 --> 00:25:38,560 government, implemented ID verification measures and what 434 00:25:38,560 --> 00:25:40,920 they would algorithmic ones. And So what they would do is 435 00:25:40,920 --> 00:25:44,800 they would run every active claim and every application that 436 00:25:44,800 --> 00:25:48,440 was outstanding through these ID verification algorithms. 437 00:25:48,720 --> 00:25:51,440 And the algorithm would flag claims that it noted as 438 00:25:51,440 --> 00:25:53,480 suspicious. And then what would happen is 439 00:25:53,480 --> 00:25:56,000 the person would have to who is flagged would have to present 440 00:25:56,120 --> 00:25:58,400 physical proof that they are who they say they are. 441 00:25:59,080 --> 00:26:02,640 That happened. And then still the state didn't 442 00:26:02,640 --> 00:26:06,200 have capacity to process that verification. 443 00:26:06,440 --> 00:26:08,600 And so you ended up with millions and millions and 444 00:26:08,600 --> 00:26:12,720 millions of people who are in desperate circumstances, can't 445 00:26:12,720 --> 00:26:14,800 keep their lights on, can't pay their rent, can't get. 446 00:26:15,120 --> 00:26:19,080 School supplies for the kids who had their benefits stopped or 447 00:26:19,080 --> 00:26:22,920 delayed by months and months and months because of this identity 448 00:26:22,920 --> 00:26:25,120 verification algorithm. Now what would happen? 449 00:26:25,120 --> 00:26:29,000 How did it work? One of the factors is, are you 450 00:26:29,000 --> 00:26:33,520 applying from the same address as somebody else, including with 451 00:26:33,520 --> 00:26:37,000 apartment buildings? So if I live in unit one O 1 and 452 00:26:37,000 --> 00:26:39,320 somebody else is applying for unemployment benefits that lives 453 00:26:39,320 --> 00:26:41,800 in unit 3O3, both of us are flagged. 454 00:26:42,040 --> 00:26:43,960 That's ridiculous. That's somebody in their 455 00:26:43,960 --> 00:26:46,880 basement coming up with some junk that make that they think 456 00:26:46,880 --> 00:26:50,000 would be associated with fraud. There's nothing statistical 457 00:26:50,000 --> 00:26:51,720 about that. There's nothing scientific about 458 00:26:51,720 --> 00:26:53,080 that. That's somebody just inventing 459 00:26:53,080 --> 00:26:55,520 stuff, right? But it invents stuff and it 460 00:26:55,520 --> 00:26:58,080 causes millions and millions of people desperation that you 461 00:26:58,080 --> 00:27:01,240 couldn't imagine. I had clients who were calling 462 00:27:01,240 --> 00:27:04,840 with active mental health crises talking about self harm because 463 00:27:04,840 --> 00:27:07,600 they couldn't get unemployment benefits even though they were 464 00:27:07,600 --> 00:27:09,760 who they said they were. And they showed that to the 465 00:27:09,760 --> 00:27:12,240 state. So that's an example of, you 466 00:27:12,240 --> 00:27:18,560 know, maybe the maybe, you know, maybe some companies care more 467 00:27:18,600 --> 00:27:20,920 than others. But here when rubber hit the 468 00:27:20,920 --> 00:27:23,840 road, it didn't matter. And ultimately studies came that 469 00:27:23,840 --> 00:27:28,640 came out after that we're assessing sort of the validity 470 00:27:28,640 --> 00:27:32,480 of these tools showed that the for the, for the most part, they 471 00:27:32,480 --> 00:27:36,320 caught eligible people, right, They weren't targeted narrowly 472 00:27:36,320 --> 00:27:38,040 to ensure that we're only getting the few. 473 00:27:38,040 --> 00:27:41,280 That are actively suspicious No, they end up catching essentially 474 00:27:41,280 --> 00:27:46,160 everybody and then just leaving, leaving folks to to to try to 475 00:27:46,160 --> 00:27:48,280 wade through the mess on their own. 476 00:27:48,560 --> 00:27:50,680 And that's just not, you know, that's not acceptable. 477 00:27:50,680 --> 00:27:52,440 There's no justification for that kind of stuff. 478 00:27:52,880 --> 00:27:58,760 Michelle Clark on LinkedIn says, can the problem of bias data be 479 00:27:58,760 --> 00:28:00,160 solved? And let me just reframe that. 480 00:28:00,160 --> 00:28:04,320 How do you manage the fact that that people are struggling to 481 00:28:04,320 --> 00:28:08,960 have data that is lacking bias? I've spoken with many of these 482 00:28:08,960 --> 00:28:12,920 folks on CXO talk, but but it's a really tough challenge from a 483 00:28:12,920 --> 00:28:14,360 technical standpoint. So. 484 00:28:14,440 --> 00:28:16,240 So what do we do about that bias data? 485 00:28:16,680 --> 00:28:18,720 Biased data is only one part of the problem, right? 486 00:28:18,720 --> 00:28:20,120 And that there are other parts of the problem. 487 00:28:20,120 --> 00:28:23,720 You can have unbiased algorithms that still cause massive harms 488 00:28:23,720 --> 00:28:27,520 and that I think would still be illegitimate in a lot of ways. 489 00:28:27,520 --> 00:28:31,800 So we want to make sure we talk about the risk in more ways than 490 00:28:31,800 --> 00:28:33,760 bias. But bias is a big one. 491 00:28:34,640 --> 00:28:38,000 And when we talk about it, there have been various ideas about 492 00:28:38,000 --> 00:28:43,160 debiasing data. And to be fair, I don't have the 493 00:28:43,160 --> 00:28:45,680 full technical background to understand the statistical 494 00:28:45,680 --> 00:28:48,680 science between all of the different ways and which is the 495 00:28:48,680 --> 00:28:51,040 best at doing what. So I don't want to claim 496 00:28:51,040 --> 00:28:53,880 otherwise. But what I do understand is that 497 00:28:53,880 --> 00:28:58,960 there's, you know, sophisticated like trying to get more data 498 00:28:58,960 --> 00:29:02,240 sets that are validated, trying to account for historical 499 00:29:02,240 --> 00:29:13,920 exclusion, again using the data in real world examples, but that 500 00:29:13,920 --> 00:29:18,560 don't have real world consequences and so forth, so 501 00:29:18,560 --> 00:29:21,160 that you're hopefully getting better data. 502 00:29:21,760 --> 00:29:25,360 So I think all that is is very much possible. 503 00:29:26,800 --> 00:29:31,760 But you know, again, I think the best, the best test against 504 00:29:31,760 --> 00:29:36,360 biased data is going to be, you know, once it's out in the 505 00:29:36,360 --> 00:29:40,320 world, are you going to face consequences for what you put 506 00:29:40,320 --> 00:29:42,720 out there, right. And if you are going to face 507 00:29:42,720 --> 00:29:45,640 consequences, then you're going to make sure or you're going to 508 00:29:45,640 --> 00:29:49,760 do your very, very best efforts to ensure that your data is not 509 00:29:49,760 --> 00:29:52,840 biased in a way that's leading to to unfair outcomes for folks. 510 00:29:53,120 --> 00:29:56,000 Self regulation is not sufficient regulation in this 511 00:29:56,000 --> 00:29:57,520 case. Yeah, exactly. 512 00:29:57,560 --> 00:29:59,680 We see the bias example all the time, right? 513 00:29:59,680 --> 00:30:01,920 There's the obvious healthcare examples about who gets 514 00:30:01,920 --> 00:30:06,920 transplant plants or, you know, black folks pain being treated 515 00:30:06,920 --> 00:30:09,200 as less real than people who are white. 516 00:30:09,480 --> 00:30:13,080 And various other examples in the healthcare context of AI 517 00:30:13,080 --> 00:30:17,600 that's deployed with, you know, bias baked in the ad targeting 518 00:30:17,600 --> 00:30:20,760 stuff from social media, like all of these things. 519 00:30:21,080 --> 00:30:23,760 And then there's another deeper question, which is if you can't 520 00:30:23,760 --> 00:30:27,600 figure it out, if you can't debias your data, maybe you 521 00:30:27,600 --> 00:30:30,440 shouldn't be using it. Maybe what you're trying to do 522 00:30:30,440 --> 00:30:34,200 is not so important that you're going to go out and reproduce 523 00:30:35,080 --> 00:30:38,800 long standing societal inequities with your technology. 524 00:30:38,800 --> 00:30:42,040 Maybe the money is not worth it. That's a value judgement I guess 525 00:30:42,320 --> 00:30:46,080 for every person and every company to make. 526 00:30:46,080 --> 00:30:51,320 But of course everybody is going to say, well, we are careful. 527 00:30:51,640 --> 00:30:54,320 That's the point. I mean, I think this comes back 528 00:30:54,320 --> 00:30:57,000 to one of my points that, you know, ultimately meaningful, 529 00:30:57,000 --> 00:31:00,880 robust, enforceable regulations are part of your. 530 00:31:01,480 --> 00:31:03,920 Gain your interest. If you are a company that is 531 00:31:03,920 --> 00:31:08,280 committed to doing things right, subjecting yourself to 532 00:31:08,280 --> 00:31:13,760 accountability is going to be a competitive advantage, right? 533 00:31:13,760 --> 00:31:16,320 Because if you have other people who are not doing things right 534 00:31:16,680 --> 00:31:20,160 and they can be subjected to lawsuits that are consequential, 535 00:31:20,160 --> 00:31:22,520 they can be subjected to regulatory oversight that's 536 00:31:22,520 --> 00:31:25,360 meaningful, that's going to be a competitive advantage for you. 537 00:31:25,360 --> 00:31:28,120 You can say, look, we are not caught up in any of that stuff. 538 00:31:28,400 --> 00:31:30,760 They are. And so we're a safer bet. 539 00:31:30,760 --> 00:31:33,640 We're a better bet. You can tap the societal values 540 00:31:33,640 --> 00:31:36,360 that that, that you provide all of those things. 541 00:31:36,360 --> 00:31:40,040 So I think ultimately regulation is in your interest because it 542 00:31:40,040 --> 00:31:43,280 creates a new competitive space for you, a competitive surface. 543 00:31:43,280 --> 00:31:46,440 I guess I'd rather say. I just want to mention for folks 544 00:31:46,440 --> 00:31:50,160 that are interested in the technology, technical 545 00:31:50,160 --> 00:31:56,280 underpinnings of data and bias, just search on the CXO talk site 546 00:31:56,280 --> 00:32:00,280 because we have done interviews with some of the leading 547 00:32:00,280 --> 00:32:04,720 technologists in the world who are focused on this problem. 548 00:32:05,080 --> 00:32:09,200 So just search for data bias and so forth on cxotalk.com. 549 00:32:09,560 --> 00:32:13,560 And oh, by the way, while you're there, you should subscribe to 550 00:32:13,560 --> 00:32:17,120 our newsletter so we can keep you up to date on shows like 551 00:32:17,120 --> 00:32:19,920 this because we have incredible shows coming up. 552 00:32:19,920 --> 00:32:23,200 Our next show, not next week, the week after is with the Chief 553 00:32:23,200 --> 00:32:28,640 Technology Officer of AMD. So subscribe to the newsletter 554 00:32:29,400 --> 00:32:32,520 we have. Our next question is from Greg 555 00:32:32,520 --> 00:32:34,640 Walters, who's another regular listener. 556 00:32:34,640 --> 00:32:40,040 And Greg, thank you. And Greg says AI is not like old 557 00:32:40,040 --> 00:32:43,200 school digital transformation broadly. 558 00:32:43,200 --> 00:32:48,800 Can AI help raise us up out of low income? 559 00:32:49,240 --> 00:32:52,120 No, not with current incentive structures in the current system 560 00:32:52,120 --> 00:32:56,520 that we exist in. People always ask me like, what 561 00:32:56,520 --> 00:32:59,640 about AI for good, right? Like what can we do that would 562 00:32:59,640 --> 00:33:03,520 advance justice? There's one example I always 563 00:33:03,520 --> 00:33:08,000 like to offer, which is with public benefits, say Medicaid or 564 00:33:08,000 --> 00:33:10,600 SNAP, which is Nutrition Assistance. 565 00:33:11,400 --> 00:33:16,680 The government knows most of the time what income and assets 566 00:33:16,680 --> 00:33:18,920 people have, right? That is, that information is 567 00:33:18,920 --> 00:33:23,000 accessible to them in some form. They know they could make 568 00:33:23,000 --> 00:33:28,160 eligibility decisions oftentimes without any or with minimal kind 569 00:33:28,160 --> 00:33:30,800 of involvement from the person who would qualify for the 570 00:33:30,800 --> 00:33:32,920 benefits. And so if you could build a 571 00:33:32,920 --> 00:33:36,640 system that would accurately, fairly consistently make those 572 00:33:36,640 --> 00:33:40,080 eligibility decisions, minimize paperwork and other burden on, 573 00:33:40,440 --> 00:33:43,600 on folks, that would be a net wonderful good that would do 574 00:33:43,600 --> 00:33:46,600 more good than 100 legal aid or lawyers in our lifetimes ever 575 00:33:46,600 --> 00:33:49,760 could. The problem is, is big companies 576 00:33:49,760 --> 00:33:51,960 have tried it, big government vendors have tried it, and it 577 00:33:51,960 --> 00:33:53,800 repeatedly fails in the same way. 578 00:33:54,000 --> 00:33:56,280 Why does it fail? Because of failed accountability 579 00:33:56,280 --> 00:33:58,000 mechanisms, right? You don't have political 580 00:33:58,000 --> 00:34:00,520 accountability, as we talked about, because hurting poor 581 00:34:00,520 --> 00:34:02,920 people generally isn't a scandal that's going to get anybody 582 00:34:02,920 --> 00:34:05,080 booted out of office. You don't have market 583 00:34:05,080 --> 00:34:07,960 accountability oftentimes in the government vendor contract 584 00:34:07,960 --> 00:34:10,840 context is because there's very few government vendors of the 585 00:34:10,840 --> 00:34:14,520 size needed to be able to to compete with one another. 586 00:34:16,080 --> 00:34:18,159 But even beyond that, you have market failures in terms of 587 00:34:18,159 --> 00:34:21,320 transparency of how your product works and what kind of public 588 00:34:21,320 --> 00:34:25,120 oversight it's it's subject to. And then you have no legal 589 00:34:25,120 --> 00:34:28,480 accountability because the existing laws that we have, 590 00:34:28,800 --> 00:34:31,719 while they have been used effectively by advocates like 591 00:34:31,719 --> 00:34:35,239 myself, are limited in scope and can only get a certain amount 592 00:34:35,280 --> 00:34:37,560 certain kind of relief. A lot of times I can't get money 593 00:34:37,560 --> 00:34:41,040 damages for the suffering that people's 'cause you can just get 594 00:34:41,040 --> 00:34:44,480 a judge that tells the state or that tells the vendor to change 595 00:34:44,480 --> 00:34:47,040 what they're doing. And so you have all these broken 596 00:34:47,440 --> 00:34:50,400 accountability mechanisms, which means that even with this good 597 00:34:50,400 --> 00:34:53,199 use, right, helping people get the healthcare that they are 598 00:34:53,199 --> 00:34:56,480 eligible for, you don't see that brought about in real life. 599 00:34:56,840 --> 00:34:59,520 And so if you can't do something like that, you're not going to 600 00:34:59,520 --> 00:35:04,200 do anything else in terms of alleviating poverty at scale. 601 00:35:04,360 --> 00:35:07,960 You can have some cool projects like in the legal world, there's 602 00:35:07,960 --> 00:35:09,520 like know your rights projects, right? 603 00:35:09,520 --> 00:35:11,920 Everybody's had a ban. Hopefully everybody's had a bad 604 00:35:11,920 --> 00:35:13,920 landlord at some point in their life, right? 605 00:35:13,920 --> 00:35:17,320 Where you needed to request repairs or ask for your security 606 00:35:17,320 --> 00:35:19,800 deposit back after you left and they were trying to hold on to 607 00:35:19,800 --> 00:35:21,720 it. There have been some cool AI 608 00:35:21,720 --> 00:35:24,000 based tools that help people do that. 609 00:35:24,280 --> 00:35:27,560 And that's cool stuff. It's great, but you know it's a 610 00:35:27,560 --> 00:35:33,520 grain of sand on the beach that borders the Pacific Ocean, 611 00:35:33,640 --> 00:35:35,800 right? Like it's cool, but it's not 612 00:35:36,520 --> 00:35:40,880 scale. This leads us to an important 613 00:35:41,040 --> 00:35:46,800 question from Trenton Butler on LinkedIn, who is a community 614 00:35:46,800 --> 00:35:49,400 advocate, organizer and project manager. 615 00:35:50,080 --> 00:35:53,800 And Trenton says this for those of us committed to ensuring 616 00:35:53,800 --> 00:35:57,920 these tools are used ethically, how can we get involved, 617 00:35:57,920 --> 00:36:01,920 especially if one does not come from a law or technology 618 00:36:01,920 --> 00:36:04,600 background? This is an important aspect is 619 00:36:04,600 --> 00:36:07,240 there's a lot of power building and community organizing that 620 00:36:07,240 --> 00:36:09,560 can be done. You know, some of the AI stuff 621 00:36:09,560 --> 00:36:11,440 happens at a very local level, right? 622 00:36:11,440 --> 00:36:14,520 Some school districts, actually about half of school districts 623 00:36:14,760 --> 00:36:19,640 use AI to predict what kids might in the future commit 624 00:36:19,640 --> 00:36:23,080 crime, right? And then targets them for law 625 00:36:23,080 --> 00:36:27,400 enforcement harassment or terror in in some cases, right? 626 00:36:27,480 --> 00:36:29,800 That's something where you could find out as a citizen. 627 00:36:29,800 --> 00:36:32,360 You don't need to be a lawyer. You can do open records 628 00:36:32,360 --> 00:36:33,680 requests. You can go to school board 629 00:36:33,680 --> 00:36:35,400 meetings. You can ask people, hey, is AI 630 00:36:35,400 --> 00:36:37,440 being used here and how does it, how does it work? 631 00:36:37,800 --> 00:36:41,680 And if it is and it looks bad, and most of the time it is bad, 632 00:36:42,560 --> 00:36:44,960 you can help organize people to get involved. 633 00:36:45,160 --> 00:36:47,960 Another local fight that there is is data centers. 634 00:36:48,080 --> 00:36:49,680 These are a big deal right there. 635 00:36:49,840 --> 00:36:53,480 The way that all the data that AI depends on is processed. 636 00:36:53,920 --> 00:36:57,080 They're subject to local land use laws, local regulation 637 00:36:57,080 --> 00:36:59,480 around utility prices and other things. 638 00:37:00,120 --> 00:37:03,000 So there are a couple ways really at home that you can get 639 00:37:03,400 --> 00:37:07,040 involved in this and building knowledge of yourself, building 640 00:37:07,080 --> 00:37:10,120 knowledge of journalists and the public, holding meetings, 641 00:37:10,320 --> 00:37:13,120 getting your neighbors involved and all that stuff. 642 00:37:13,120 --> 00:37:17,440 And it can be daunting. So and there's a huge gap in 643 00:37:17,440 --> 00:37:18,840 helping people do that right now. 644 00:37:18,960 --> 00:37:21,040 And that's one of the reason tectonic justice exists. 645 00:37:21,040 --> 00:37:24,000 So in the in the self interested plug, please follow us, please 646 00:37:24,000 --> 00:37:26,840 stay in contact. And as we're building out, it's 647 00:37:26,840 --> 00:37:29,880 just me and it's been my first two employees joined last month. 648 00:37:30,360 --> 00:37:32,080 So we're still very much in the building phase. 649 00:37:32,080 --> 00:37:36,640 But as we get more established, we want to be doing working in 650 00:37:36,640 --> 00:37:38,760 partnership with folks who want to be engaged around these 651 00:37:38,760 --> 00:37:40,400 issues. So, so please stay up with us. 652 00:37:40,520 --> 00:37:47,600 We have another question now from Twitter and this is from 653 00:37:48,240 --> 00:37:54,000 Chris Peterson, who says what agency or is there is or is 654 00:37:54,000 --> 00:37:57,080 there just one? Would you suggest as the AI 655 00:37:57,640 --> 00:38:02,480 Ombudsman in the US? He says also that for folks in 656 00:38:02,480 --> 00:38:08,800 charge of big AI, 99 plus percent of us are lower income 657 00:38:08,840 --> 00:38:11,840 in quotes. There is no one ombudsperson 658 00:38:12,840 --> 00:38:16,280 kind of yet around AI. And I mean, that's an 659 00:38:16,280 --> 00:38:19,200 interesting idea in terms of meaningful accountability 660 00:38:19,200 --> 00:38:22,320 because there are ombuds people in healthcare and in nursing 661 00:38:22,320 --> 00:38:25,040 homes and other similar entities. 662 00:38:26,800 --> 00:38:28,920 It's a huge gap. That's part of why we exist is 663 00:38:28,920 --> 00:38:31,280 right to be focused people on the ground, right? 664 00:38:31,280 --> 00:38:34,560 Like I was a lawyer working with hundreds and hundreds of low 665 00:38:34,560 --> 00:38:37,320 income people to try to fight this stuff. 666 00:38:38,000 --> 00:38:43,800 So I think in the ecosystem, sort of the nonprofit ecosystem, 667 00:38:44,080 --> 00:38:46,320 there are few organizations that are trying to build up the 668 00:38:46,320 --> 00:38:50,320 capacity to do some of this stuff to watchdog the use of AI. 669 00:38:51,520 --> 00:38:53,360 And then there are a lot of established organizations that 670 00:38:53,360 --> 00:38:55,280 are more focused on kind of the policy level. 671 00:38:55,920 --> 00:39:02,200 So there is no one ombudsperson. In terms of the other aspect of 672 00:39:02,360 --> 00:39:06,640 the question, I guess I would need more context about what it 673 00:39:06,640 --> 00:39:10,120 means that 99% of us are subject. 674 00:39:10,120 --> 00:39:11,880 Maybe it's that we're subject to the big tech. 675 00:39:12,240 --> 00:39:17,920 What he was saying is that it's it's the billionaire question. 676 00:39:18,360 --> 00:39:23,360 Now is the time to get involved before these technologies become 677 00:39:23,360 --> 00:39:26,760 entrenched as legitimate ways to make decisions about these core 678 00:39:26,760 --> 00:39:30,120 aspects of life. Because even though AIA lot of 679 00:39:30,120 --> 00:39:33,640 times purports to be, or at least it's it's hype. 680 00:39:33,880 --> 00:39:38,920 It's hype, it's hype men purported to be kind of this 681 00:39:38,920 --> 00:39:43,240 objective way to make decisions. Whose objectivity is it, right? 682 00:39:43,240 --> 00:39:45,920 If it's always limiting access to benefits, if it's always 683 00:39:45,920 --> 00:39:50,560 making housing or jobs or education harder to get, then 684 00:39:50,560 --> 00:39:53,280 it's not really objective. It's it's, you know, the people 685 00:39:53,280 --> 00:39:56,560 who are developing or using the AI, It's achieving their ends. 686 00:39:57,080 --> 00:40:00,520 So now is very much the moment for it, I think, because this 687 00:40:00,760 --> 00:40:03,200 field is. Relatively new is sort of a 688 00:40:03,200 --> 00:40:05,280 social phenomenon and a social movement. 689 00:40:05,400 --> 00:40:08,600 There isn't a lot of the infrastructure that needs to be 690 00:40:08,600 --> 00:40:11,800 there to help people get organized and engaged around it. 691 00:40:11,920 --> 00:40:14,400 So a lot of my answers are probably unsatisfying. 692 00:40:14,400 --> 00:40:16,680 It's like, we'll talk to your community, organize around it, 693 00:40:16,680 --> 00:40:19,040 stay up with tectonic justice, these kinds of things. 694 00:40:19,040 --> 00:40:21,280 Because that's what we're trying to build, is build the 695 00:40:21,280 --> 00:40:24,440 infrastructure for people to be able to channel their front, 696 00:40:24,440 --> 00:40:27,320 their concerns, their frustrations, their energy 697 00:40:27,480 --> 00:40:30,320 towards ensuring something that looks more like justice. 698 00:40:30,800 --> 00:40:38,240 But aren't you in a way trying to turn back the clock to a a a 699 00:40:38,240 --> 00:40:44,080 simpler and easier time Before we had AI and AI is not going 700 00:40:44,080 --> 00:40:49,840 away and it's growth is going to continue to make incursions into 701 00:40:49,960 --> 00:40:54,440 every aspect of decision making. That's the overwhelming you 702 00:40:54,440 --> 00:40:58,120 talked about earlier, the PR budgets, right, of big tech. 703 00:40:58,400 --> 00:41:01,480 And that's the overwhelming sense is that it's inevitable. 704 00:41:01,880 --> 00:41:05,200 But is it really right? Why can't a nurse make a 705 00:41:05,200 --> 00:41:08,560 decision about how much home care a disabled person needs? 706 00:41:08,920 --> 00:41:12,640 Why is that not viable anymore? Why shouldn't that be the case? 707 00:41:12,960 --> 00:41:17,120 Why can't we use technology in a way that supports human based 708 00:41:18,120 --> 00:41:21,640 decision making rather than essentially making the decision 709 00:41:22,000 --> 00:41:25,000 for us with like cursory human oversight at all? 710 00:41:25,400 --> 00:41:28,640 And I think that has to be the questions, what is the 711 00:41:28,640 --> 00:41:33,120 legitimate use of AI? And then even where the use is 712 00:41:33,120 --> 00:41:35,080 sort of legitimate, let's go through all the vetting we 713 00:41:35,080 --> 00:41:37,640 talked about earlier. But let's also talk about the 714 00:41:37,640 --> 00:41:39,920 bigger picture. Questions in terms of what it 715 00:41:39,920 --> 00:41:42,040 means for the Earth, right? We know that AI has 716 00:41:42,200 --> 00:41:44,880 environmental consequences. There's debate about how many 717 00:41:44,880 --> 00:41:48,800 liters of water each ChatGPT you know, prompt uses or whatever, 718 00:41:48,800 --> 00:41:52,200 but like we know that it's draining water in certain places 719 00:41:52,200 --> 00:41:55,400 where water is scarce. We know that it's responsible or 720 00:41:55,400 --> 00:41:58,320 at least correlated with energy price increases. 721 00:41:58,680 --> 00:42:02,600 We know it's correlated with the use of non renewable energies. 722 00:42:02,760 --> 00:42:06,160 So you have to factor all these things into the equation in 723 00:42:06,160 --> 00:42:08,800 terms of its societal value and its societal costs. 724 00:42:09,160 --> 00:42:12,760 And it may be that if we actually do a concerted, 725 00:42:13,000 --> 00:42:18,040 reflected effort that accounts for all these externalities, we 726 00:42:18,040 --> 00:42:20,320 realize, you know what, this isn't that harm. 727 00:42:20,320 --> 00:42:22,800 Maybe we shouldn't do it or we should only do it in these 728 00:42:22,800 --> 00:42:25,280 limited circumstances. And I think that's what we have 729 00:42:25,280 --> 00:42:27,440 to be engaging. In and that's why I always 730 00:42:27,440 --> 00:42:29,400 reject the frame of inevitability. 731 00:42:29,640 --> 00:42:31,920 I'm a practical person. I generally need to solve 732 00:42:31,920 --> 00:42:34,080 problems for my low income clients and that doesn't always 733 00:42:34,080 --> 00:42:36,920 allow me to like, be pie in the sky principled. 734 00:42:37,080 --> 00:42:39,120 But we can be pie in the sky principled. 735 00:42:39,200 --> 00:42:43,680 While also being practical and, and start thinking like, is this 736 00:42:43,680 --> 00:42:47,120 really worth it? Is the productivity gain really 737 00:42:47,120 --> 00:42:50,520 worth all the cost? And so far even that in the 738 00:42:50,520 --> 00:42:54,440 corporate sphere hasn't been clear that there are really net 739 00:42:54,440 --> 00:42:58,960 productivity gains, particularly when you factor in the required 740 00:42:58,960 --> 00:43:00,920 human oversight for its continued use. 741 00:43:01,400 --> 00:43:03,400 And so I don't think it's inevitable. 742 00:43:03,600 --> 00:43:06,720 I think it will be inevitable if we don't in the next, you know. 743 00:43:06,720 --> 00:43:10,240 Decade or two really reckon with the implications of it. 744 00:43:10,760 --> 00:43:12,920 My friend, you have an uphill fight. 745 00:43:12,960 --> 00:43:18,000 I have to say on this point, you and I have to agree to disagree 746 00:43:18,000 --> 00:43:24,600 because as I look out over the developments of AI and automated 747 00:43:24,600 --> 00:43:32,160 decision making, I cannot see, I cannot fathom, and maybe I 748 00:43:32,200 --> 00:43:36,280 reflect kind of a typical technology viewpoint, but I 749 00:43:36,280 --> 00:43:42,800 cannot fathom that AI is not going to grow much as the steam 750 00:43:42,800 --> 00:43:45,920 engine influenced every facet of our lives. 751 00:43:45,920 --> 00:43:48,280 And you can say the steam engine also caused a lot of problems. 752 00:43:48,520 --> 00:43:50,280 Potentially. I mean, I think in society, it's 753 00:43:50,280 --> 00:43:53,160 not like we just accept technology inevitably without, 754 00:43:54,320 --> 00:43:59,120 you know, restricting its use. I mean, certainly nuclear energy 755 00:43:59,120 --> 00:44:05,400 has had significant use restrictions around it and its 756 00:44:05,400 --> 00:44:07,760 development and where it can be used and everything else. 757 00:44:08,040 --> 00:44:12,320 Cars have had a lot of restrictions around how they can 758 00:44:12,320 --> 00:44:13,920 be used. Everybody, I'm sure thought 759 00:44:13,920 --> 00:44:17,640 Ralph Nader in the 70s was ridiculous for, you know, 760 00:44:17,640 --> 00:44:20,320 advocating for seatbelts, right? And now that's just an accepted 761 00:44:20,320 --> 00:44:22,480 facet of the cars. Now that doesn't take care of 762 00:44:22,480 --> 00:44:25,720 all the harms that cars are potentially causing, right? 763 00:44:25,720 --> 00:44:30,920 And I'm not saying that it does, but it's one example of of 764 00:44:30,920 --> 00:44:34,200 movement that way. And all of these things are 765 00:44:34,200 --> 00:44:38,040 have, you know, essentially corporate power and lots of 766 00:44:38,040 --> 00:44:42,640 money going against, you know, people who seem like they're in 767 00:44:42,640 --> 00:44:46,920 the way of inevitability. But we have to be a little bit, 768 00:44:47,280 --> 00:44:50,880 you know, what's the word? We have to believe that 769 00:44:50,880 --> 00:44:53,720 something more is possible. Otherwise we just resign 770 00:44:53,720 --> 00:44:57,440 ourselves to accepting the worst version of whatever it is that 771 00:44:57,440 --> 00:44:59,840 we're fighting against. And that's not a, that's not a 772 00:44:59,840 --> 00:45:02,800 concession I'm willing to make. Like I'll fight like hell, maybe 773 00:45:02,800 --> 00:45:06,320 I'll lose, but I bet you that we're better off because of the 774 00:45:06,320 --> 00:45:10,880 fight than if nobody fought. Let's jump to another question, 775 00:45:10,880 --> 00:45:14,000 and this is from Ravi Karkara on LinkedIn. 776 00:45:14,440 --> 00:45:19,160 He says, oh and I should mention he is Co founder and author of 777 00:45:19,160 --> 00:45:22,120 AT AI for food global initiative. 778 00:45:22,760 --> 00:45:27,560 And he says, how should global stakeholders navigate the 779 00:45:27,680 --> 00:45:31,760 ethical challenges and data governance differences posed by 780 00:45:31,760 --> 00:45:36,840 China's AI strategy, particularly its state centric 781 00:45:36,840 --> 00:45:40,400 data policies, while promoting international norms for 782 00:45:40,400 --> 00:45:44,640 responsible and transparent AI development? 783 00:45:44,640 --> 00:45:48,280 Not sure how much expertise you have in in China, but thoughts 784 00:45:48,280 --> 00:45:52,800 on kind of a global perspective? In the global context, the AI 785 00:45:52,800 --> 00:45:57,080 discussion becomes even more interesting because there's a 786 00:45:57,080 --> 00:46:00,840 lot of people who are pushing AI as a solution to kind of global 787 00:46:00,840 --> 00:46:04,760 poverty, right, and inaccessible healthcare, right. 788 00:46:04,760 --> 00:46:11,200 You get the the story of like people in remote villages, you 789 00:46:11,200 --> 00:46:14,400 know, in the majority world who are now suddenly able to access, 790 00:46:14,760 --> 00:46:17,280 you know, medical care or at least knowledge about medical 791 00:46:17,280 --> 00:46:19,520 care that they couldn't because they couldn't travel to cities 792 00:46:19,520 --> 00:46:22,480 and so forth. And I think there's, you know, 793 00:46:22,480 --> 00:46:26,160 who am I to say sitting here and you know, in the US, in Los 794 00:46:26,160 --> 00:46:29,600 Angeles, CA, to say that that's a bad use of AII. 795 00:46:31,000 --> 00:46:33,240 Think where I care about. There are a few things. 796 00:46:33,240 --> 00:46:39,120 One is the data extraction that comes from, you know, expanded 797 00:46:39,120 --> 00:46:41,880 use of AI. Is it fair to be extracting all 798 00:46:41,880 --> 00:46:43,960 the data about people's behaviors, who they are, 799 00:46:43,960 --> 00:46:46,600 etcetera, etcetera, when you're going to monetize that and when 800 00:46:46,600 --> 00:46:48,600 they really don't have meaningful consent, right? 801 00:46:48,600 --> 00:46:52,680 Opting into opting into the terms of service on a contract 802 00:46:52,680 --> 00:46:55,520 for social media, for example, that's not a real thing of 803 00:46:55,520 --> 00:46:58,880 consent for most people. So what's the data extraction 804 00:46:59,120 --> 00:47:01,680 relationship? What's the labor relationship, 805 00:47:01,680 --> 00:47:03,840 right? Because just as there's a person 806 00:47:03,840 --> 00:47:06,000 who needs to seek healthcare in a village, right? 807 00:47:06,000 --> 00:47:09,520 And this is an archetype, I'm not trying to use a specific 808 00:47:09,520 --> 00:47:12,920 example, there's. Somebody not too far away. 809 00:47:13,360 --> 00:47:17,960 Who's being paid pennies on the dollar to view really horrific, 810 00:47:17,960 --> 00:47:20,080 traumatic data and label it right? 811 00:47:20,080 --> 00:47:23,760 There are people being exploited for the supply chain and 812 00:47:23,760 --> 00:47:26,200 everything else. So I think as we transition to 813 00:47:26,200 --> 00:47:29,560 the global discussion, you're going to have a lot of these use 814 00:47:29,560 --> 00:47:33,400 cases of AI for good that are going to be uplifted to justify 815 00:47:33,400 --> 00:47:35,360 the continuance of the AI regime. 816 00:47:35,920 --> 00:47:39,400 And if we're being reflective people that are serious about 817 00:47:39,960 --> 00:47:42,320 kind of the policy implications of this, we need to factor in 818 00:47:42,320 --> 00:47:44,440 all the costs. What is the what are the costs 819 00:47:44,440 --> 00:47:46,760 of the data extraction of the labor exploitation? 820 00:47:47,000 --> 00:47:49,520 What are the downstream costs of having other people's lives 821 00:47:49,520 --> 00:47:53,360 decided by the not so good and not so innocent uses of AI? 822 00:47:53,760 --> 00:48:00,440 This is from Elizabeth Shaw, who says how does today's AI differ 823 00:48:00,440 --> 00:48:05,840 from previous algorithms from the view of social harms, and 824 00:48:05,960 --> 00:48:09,760 can AI be part of the solution? So really the question is what's 825 00:48:09,760 --> 00:48:13,720 unique about AI and can AI help solve these problems? 826 00:48:14,160 --> 00:48:17,320 A lot of the technologies that are used in government services 827 00:48:17,320 --> 00:48:22,000 right now are not the latest generation AI, you know, LLMS 828 00:48:22,000 --> 00:48:26,040 and other things like this. A lot of them are more older 829 00:48:26,040 --> 00:48:29,360 algorithms that are based, you know, that were supervised 830 00:48:29,680 --> 00:48:33,800 learning that were based on statistical regression and these 831 00:48:33,800 --> 00:48:36,200 sorts of technologies and those are really harmful. 832 00:48:36,200 --> 00:48:39,800 I don't think AI has any like the latest generation AI has 833 00:48:39,800 --> 00:48:44,040 anything to offer in a lot of these, in a lot of these 834 00:48:44,040 --> 00:48:46,880 contexts. Again, so long as it's existing 835 00:48:46,880 --> 00:48:50,800 for purposes that are to, you know, essentially limit life 836 00:48:50,800 --> 00:48:53,520 opportunities. And in this, you know, vacuum of 837 00:48:53,520 --> 00:48:56,920 accountability, I don't think the technological sophistication 838 00:48:56,920 --> 00:48:59,080 is going to make much of a difference because they're going 839 00:48:59,080 --> 00:49:01,720 to be making the same decisions with the same incentives, right? 840 00:49:02,680 --> 00:49:08,680 And one way that we are seeing this now is, excuse me, in the 841 00:49:08,680 --> 00:49:12,680 recent developments federally, right, where the administration 842 00:49:12,680 --> 00:49:17,240 has implemented AI and for example, Social Security offices 843 00:49:17,520 --> 00:49:19,440 and it's made Social Security harder to access. 844 00:49:19,440 --> 00:49:21,120 It's made people have to wait longer. 845 00:49:21,120 --> 00:49:22,840 It's made people not get their benefits. 846 00:49:23,080 --> 00:49:25,480 And that is technically the latest generation of AI. 847 00:49:26,440 --> 00:49:32,120 So I think that's an example of, I always challenge the promise 848 00:49:32,120 --> 00:49:35,440 that the premise that AI is going to somehow fix existing 849 00:49:35,440 --> 00:49:37,440 problems just because of the technology is going to get more 850 00:49:37,440 --> 00:49:39,320 sophisticated. No, what's going to happen is 851 00:49:39,320 --> 00:49:41,720 it's going to make those problems even harder to fight as 852 00:49:41,720 --> 00:49:45,600 the technology becomes even more inscrutable, more insulated from 853 00:49:45,600 --> 00:49:48,680 public accountability and transparency and all of those 854 00:49:48,680 --> 00:49:51,000 things. You are not of the school of 855 00:49:51,000 --> 00:49:53,840 thought that AI is going to be the great savior. 856 00:49:54,440 --> 00:49:57,560 Oh, God, no. Oh God, no, no. 857 00:49:57,560 --> 00:49:59,800 If anything, it's the opposite way, right? 858 00:49:59,800 --> 00:50:03,440 It's immiserating people. And I think, you know, the 859 00:50:03,440 --> 00:50:09,680 recent use of AI in the last few months by the administration 860 00:50:09,680 --> 00:50:11,520 helps show us this is devastating. 861 00:50:11,640 --> 00:50:16,160 AI is being used to destroy government, destroy government 862 00:50:16,160 --> 00:50:22,360 capacity, destroy lives, and violate the law left and right 863 00:50:22,680 --> 00:50:25,400 and everything else. It is a weapon that is uniquely 864 00:50:25,400 --> 00:50:30,080 suited, uniquely suited to authoritarianism, right? 865 00:50:30,080 --> 00:50:32,000 Even by its nature, it has inscrutable. 866 00:50:32,000 --> 00:50:35,200 It's sort of just as like an the Oracle that tells you what the 867 00:50:35,200 --> 00:50:37,640 decision is, but doesn't tell you why it's making the decision 868 00:50:37,640 --> 00:50:39,080 doesn't allow you to disagree with it. 869 00:50:39,360 --> 00:50:42,640 All of that, That's like an authoritarian approach to 870 00:50:42,640 --> 00:50:45,040 thinking. And to decision making. 871 00:50:45,320 --> 00:50:48,640 So no, if if anything, AI is a greater threat. 872 00:50:48,760 --> 00:50:52,040 So I think our continued existence as a, you know, as a 873 00:50:52,200 --> 00:50:55,080 democratic society, it's antithetical to a lot of 874 00:50:55,080 --> 00:50:58,680 egalitarian notions, if anything is going to make things worse. 875 00:50:59,080 --> 00:51:05,000 Can you offer advice to folks who are working in the corporate 876 00:51:05,000 --> 00:51:08,760 realm, who really have a conscience and who don't want to 877 00:51:08,760 --> 00:51:12,240 see the perpetuation of these kinds of harms that you've been 878 00:51:12,240 --> 00:51:14,600 describing? The current uses of AI are 879 00:51:14,600 --> 00:51:17,800 destroying its reputation. I think that's brand risk for 880 00:51:17,800 --> 00:51:20,360 your companies. I think that's brand risk for AI 881 00:51:20,360 --> 00:51:22,640 as it's as a as a sort of venture. 882 00:51:23,080 --> 00:51:28,600 And I think opposing authoritarianism, particularly 883 00:51:28,600 --> 00:51:32,760 authoritarianism that's being fueled by AI is a really 884 00:51:32,760 --> 00:51:36,000 critical thing for your long term survival for various 885 00:51:36,000 --> 00:51:39,440 reasons. Then you know on a less sort of 886 00:51:39,440 --> 00:51:44,880 global and do me scale is all of the things that we're talking 887 00:51:44,880 --> 00:51:46,840 about push for meaningful regulation. 888 00:51:47,160 --> 00:51:49,320 What are you scared of? That's my question. 889 00:51:49,320 --> 00:51:51,320 Like, if you've got this great product that's backed by the 890 00:51:51,320 --> 00:51:53,920 most sophisticated science we have, what are you scared of? 891 00:51:53,920 --> 00:51:55,800 You should be proud of that. You should be putting that out 892 00:51:55,800 --> 00:51:58,640 there and saying, you know what? Subject us to accountability 893 00:51:58,640 --> 00:52:02,560 because our stuff is so strong, so scientifically sound and 894 00:52:02,560 --> 00:52:07,320 produces such clear value for the public that we're willing to 895 00:52:07,560 --> 00:52:11,200 embrace the being under a microscope. 896 00:52:11,480 --> 00:52:13,800 And I don't see that yet. And that's why I even challenge, 897 00:52:13,800 --> 00:52:17,680 you know, the the notion of an inevitability in terms of pure 898 00:52:17,680 --> 00:52:22,680 efficiency. There haven't been clear one 899 00:52:22,680 --> 00:52:27,320 sided efficiency gains that have made adoption of AI, even for 900 00:52:27,320 --> 00:52:30,760 non decision making purposes, universally sensible. 901 00:52:31,040 --> 00:52:34,880 Help make AI an electoral issue. Let's start talking about the 902 00:52:34,880 --> 00:52:36,960 injustices. I mean, I think there's going to 903 00:52:36,960 --> 00:52:39,400 be in some incentive problems there because, you know, there's 904 00:52:39,400 --> 00:52:41,440 big tech money that funds both parties. 905 00:52:41,440 --> 00:52:43,080 And I think there are a lot of people who don't want to be 906 00:52:43,080 --> 00:52:45,480 accused of being a Luddite. And, you know, there's other 907 00:52:45,480 --> 00:52:49,000 incentives there. But I think policy makers have a 908 00:52:49,000 --> 00:52:52,920 responsibility to educate the public much more intensely than 909 00:52:52,920 --> 00:52:55,360 they currently do about the harms of AI. 910 00:52:55,360 --> 00:52:59,120 Engage the public, hopefully create a base of people so that 911 00:52:59,120 --> 00:53:01,880 there's a balance, a counterbalance to the to the 912 00:53:01,880 --> 00:53:05,000 weight of big tech in these discussions so that you can push 913 00:53:05,000 --> 00:53:08,440 for meaningful legislation and regulation and ongoing and 914 00:53:08,440 --> 00:53:10,720 enforcement and oversight. And I think that's going to be 915 00:53:10,720 --> 00:53:16,040 vital to, you know, again, sustaining a democratic society, 916 00:53:16,280 --> 00:53:21,040 pushing for less inequality and ultimately having an environment 917 00:53:21,040 --> 00:53:23,360 where people have a real chance to thrive. 918 00:53:23,840 --> 00:53:29,240 What advice do you have to individuals who are victims of 919 00:53:30,440 --> 00:53:36,400 on unjust AI decisions? This is really hard. 920 00:53:36,400 --> 00:53:40,000 A lot of times you don't even know that AI is the reason for 921 00:53:40,000 --> 00:53:44,120 that you're suffering. So what I would say is contact 922 00:53:44,120 --> 00:53:46,640 your local legal aid program if you're hurt by this stuff. 923 00:53:46,880 --> 00:53:49,520 Legal Aid provides free legal services to folks throughout the 924 00:53:49,520 --> 00:53:53,480 country on civil legal matters. Talk to your neighbors, talk to 925 00:53:53,480 --> 00:53:55,080 other people in the same situations. 926 00:53:55,080 --> 00:53:58,000 Try to see if things are going, what's going on, and gather 927 00:53:58,000 --> 00:54:02,280 information and start kind of engaging in the things that are 928 00:54:02,280 --> 00:54:04,360 needed to push back. And if you're in a position to 929 00:54:04,720 --> 00:54:08,520 sue, if you're in a position to offer your story to a 930 00:54:08,520 --> 00:54:12,000 journalist, take those opportunities to speak for 931 00:54:12,000 --> 00:54:16,400 yourself. Because there are relatively few 932 00:54:16,400 --> 00:54:20,120 stories out there. And the discourse doesn't have 933 00:54:20,120 --> 00:54:22,560 the people who are hurt the most, doesn't have the people 934 00:54:22,560 --> 00:54:25,080 who are having to live with the consequences of what powerful 935 00:54:25,080 --> 00:54:27,640 people do. And any chance that we have for 936 00:54:27,640 --> 00:54:31,840 long term success is going to depend on you being able to 937 00:54:31,840 --> 00:54:34,280 become a leader and to Share your story and share your 938 00:54:34,280 --> 00:54:37,680 passion and share your injustice so that we can make it better 939 00:54:37,680 --> 00:54:40,560 for everybody. Kevin Delebin, founder of 940 00:54:40,560 --> 00:54:45,200 Tectonic Justice, thank you so much for taking your time to be 941 00:54:45,200 --> 00:54:47,440 here. And I'm very grateful for you to 942 00:54:47,920 --> 00:54:52,880 for you sharing a point of view that honestly is quite different 943 00:54:52,880 --> 00:54:56,040 from that that we usually hear on CXO Talk. 944 00:54:56,040 --> 00:54:58,080 So thank you for taking your time and being here with us. 945 00:54:58,440 --> 00:55:01,320 This is really fun and thank you also to the audience for all the 946 00:55:01,320 --> 00:55:03,400 great questions and you for having me, Michael. 947 00:55:03,800 --> 00:55:06,400 Audience, thank you guys. You guys are awesome. 948 00:55:06,400 --> 00:55:08,920 I mean truly that your questions are so thoughtful. 949 00:55:08,920 --> 00:55:13,880 You guys are so smart. Before you go, subscribe to our 950 00:55:13,880 --> 00:55:18,600 newsletter, go to cxotalk.com, check out our newsletter and 951 00:55:19,120 --> 00:55:22,120 check us out for our next shows. We have great shows coming up. 952 00:55:22,720 --> 00:55:27,720 And if you're interested again in topics like data bias, all of 953 00:55:27,720 --> 00:55:31,440 these issues that we've been discussing, search on the CXO 954 00:55:31,440 --> 00:55:36,040 Talk site because we have had lots of perspectives on this 955 00:55:36,040 --> 00:55:39,480 from business leaders, from politicians, you name it. 956 00:55:39,480 --> 00:55:42,880 So dig into the interviews on cxotalk.com. 957 00:55:42,880 --> 00:55:44,960 It's it's truly a great resource. 958 00:55:45,240 --> 00:55:47,400 Thanks so much, everybody. We'll see you again next time, 959 00:55:47,400 --> 00:55:48,320 and I hope you have a great day.