1 00:00:01,639 --> 00:00:04,320 Speaker 1: Justice looks like, how do we use these tools in 2 00:00:04,360 --> 00:00:09,800 Speaker 1: a joyful, uplifting manner versus just being reactive to the 3 00:00:09,920 --> 00:00:11,040 Speaker 1: next times. 4 00:00:15,760 --> 00:00:17,480 Speaker 2: There are No Girls on the Internet. As a production 5 00:00:17,520 --> 00:00:25,200 Speaker 2: of iHeartRadio and Unbossed Creative, I'm brigitat and this is 6 00:00:25,200 --> 00:00:30,280 Speaker 2: there are No Girls on the Internet. This week, President 7 00:00:30,280 --> 00:00:33,520 Speaker 2: Biden signed an executive order to create some safeguards around 8 00:00:33,520 --> 00:00:37,120 Speaker 2: the use of AI. This comes after black women women 9 00:00:37,240 --> 00:00:40,760 Speaker 2: like doctor Joy Blomwini, founder of the Algorithmic Justice League 10 00:00:40,960 --> 00:00:43,800 Speaker 2: and author of the new book of Masking AI, have 11 00:00:43,880 --> 00:00:46,440 Speaker 2: been speaking up about the ways that technology like AI 12 00:00:46,880 --> 00:00:50,040 Speaker 2: has already harmed marginalized communities and what needs to be 13 00:00:50,040 --> 00:00:53,080 Speaker 2: done to stop it now. That last part is key 14 00:00:53,320 --> 00:00:56,800 Speaker 2: because even though her groundbreaking research has been critical to 15 00:00:56,880 --> 00:01:01,280 Speaker 2: understanding technology harms, doctor Blomweni's vision the future of technology 16 00:01:01,480 --> 00:01:05,959 Speaker 2: is optimistic, blending poetry and technology. She asks, what is 17 00:01:06,000 --> 00:01:08,680 Speaker 2: our collective, just and joyful vision for the future. 18 00:01:12,080 --> 00:01:15,560 Speaker 1: Hello, my name is doctor Joy Bloomwini. I'm the founder 19 00:01:15,640 --> 00:01:20,559 Speaker 1: of the Algorithmic Justice League and the author of Unmasking AI. 20 00:01:21,319 --> 00:01:24,200 Speaker 1: My pronouns are she and hers. 21 00:01:24,720 --> 00:01:27,720 Speaker 2: So I've heard you call yourself a poet of code 22 00:01:27,760 --> 00:01:29,880 Speaker 2: which is awesome. What do you mean by that? 23 00:01:30,680 --> 00:01:34,240 Speaker 1: So I am the daughter of an artist and a scientist, 24 00:01:34,360 --> 00:01:36,800 Speaker 1: So I do feel I've grown up with the arts 25 00:01:36,800 --> 00:01:41,800 Speaker 1: and science literally together, and so when I use the 26 00:01:41,959 --> 00:01:45,479 Speaker 1: term poet of Code, it's really to reflect those two 27 00:01:46,280 --> 00:01:50,040 Speaker 1: sensibilities which inform my work. So there's a major part 28 00:01:50,080 --> 00:01:54,080 Speaker 1: of it which is storytelling and humanizing what's going on 29 00:01:54,440 --> 00:01:59,520 Speaker 1: with evocative audits, you know, and portrayals. And then there's 30 00:01:59,560 --> 00:02:03,600 Speaker 1: another aspect of it that is getting into the analytical 31 00:02:03,720 --> 00:02:08,240 Speaker 1: technical pieces of what it means to evaluate a machine 32 00:02:08,320 --> 00:02:13,680 Speaker 1: learning system or other types of AI applications. So the 33 00:02:13,720 --> 00:02:19,240 Speaker 1: Poet of Code is very much indicating my origins as 34 00:02:19,320 --> 00:02:22,240 Speaker 1: the daughter of an artist and a scientist. 35 00:02:22,880 --> 00:02:24,880 Speaker 2: Do you think that that the way that you approach 36 00:02:24,919 --> 00:02:27,920 Speaker 2: the work has really helped bring more folks into it 37 00:02:27,960 --> 00:02:31,400 Speaker 2: Because I've been interested and invested in conversations about tech 38 00:02:31,440 --> 00:02:35,799 Speaker 2: for a very long time. I did not care about slash, 39 00:02:36,000 --> 00:02:39,640 Speaker 2: maybe even fully understand the implications around bias and things 40 00:02:39,680 --> 00:02:43,560 Speaker 2: like AI until you and so you had this way 41 00:02:43,560 --> 00:02:47,040 Speaker 2: of really making it visible, really making it poetic, really 42 00:02:47,080 --> 00:02:48,760 Speaker 2: making me understand what was at stake. 43 00:02:49,160 --> 00:02:49,760 Speaker 1: Do you think that. 44 00:02:50,000 --> 00:02:53,160 Speaker 2: Part of why that is why folks feel so drawn 45 00:02:53,160 --> 00:02:56,320 Speaker 2: to your work is because you make it so poetic, So, 46 00:02:56,720 --> 00:03:00,000 Speaker 2: you know, story based really help people understand like where 47 00:03:00,160 --> 00:03:00,919 Speaker 2: they fit into it. 48 00:03:01,639 --> 00:03:05,160 Speaker 1: I think there is that element. So as I was 49 00:03:05,200 --> 00:03:09,679 Speaker 1: doing my research at MIT that involves publishing research papers, 50 00:03:10,160 --> 00:03:12,240 Speaker 1: and as fun as those are, you know, that's a 51 00:03:12,440 --> 00:03:17,480 Speaker 1: very small community that will likely read those types of papers. 52 00:03:17,520 --> 00:03:19,280 Speaker 1: So I wanted to say, how do I go from 53 00:03:19,320 --> 00:03:24,239 Speaker 1: the performance metrics of evaluating an AI system to something 54 00:03:24,320 --> 00:03:28,400 Speaker 1: like performance art. Why does this even matter? How do 55 00:03:28,400 --> 00:03:31,640 Speaker 1: we get to the heart of all of these numbers. 56 00:03:31,720 --> 00:03:35,800 Speaker 1: So if we see bias in a system and we 57 00:03:35,960 --> 00:03:38,920 Speaker 1: quantify it, that's only part of the story. The other 58 00:03:39,000 --> 00:03:41,240 Speaker 1: part of the story is what does that mean for 59 00:03:41,320 --> 00:03:48,240 Speaker 1: someone who could experience algorithmic discrimination, algorithmic rature, or exploitation. 60 00:03:48,440 --> 00:03:51,720 Speaker 1: And that's where the storytelling has to come in and 61 00:03:51,760 --> 00:03:56,360 Speaker 1: it did for me. When I was a student at MIT, 62 00:03:57,080 --> 00:04:00,960 Speaker 1: I had an opportunity to do research that showed large 63 00:04:01,000 --> 00:04:07,120 Speaker 1: skin type gaps and gender gaps with the accuracy of 64 00:04:07,240 --> 00:04:10,960 Speaker 1: different gender classification systems. So these are AI systems that 65 00:04:11,000 --> 00:04:13,160 Speaker 1: look at a photo of your face and try to 66 00:04:13,200 --> 00:04:16,440 Speaker 1: guess your gender. Where could that go wrong? Well, so 67 00:04:16,640 --> 00:04:21,159 Speaker 1: we decided to do a bit of an evaluation. And 68 00:04:21,279 --> 00:04:23,280 Speaker 1: after we ran the numbers and we showed there were 69 00:04:23,400 --> 00:04:28,279 Speaker 1: large gaps and biases documented, I wanted to show people 70 00:04:28,520 --> 00:04:29,600 Speaker 1: why it mattered. 71 00:04:30,120 --> 00:04:33,320 Speaker 2: It does matter to all of us, whether you spend 72 00:04:33,360 --> 00:04:35,359 Speaker 2: a lot of time thinking about it or not. This 73 00:04:35,560 --> 00:04:38,960 Speaker 2: kind of technology is becoming more and more commonplace, despite 74 00:04:39,000 --> 00:04:40,919 Speaker 2: the fact that it doesn't work so well on women 75 00:04:41,040 --> 00:04:45,080 Speaker 2: or people with darker complexions, setting us up to disproportionately 76 00:04:45,160 --> 00:04:49,440 Speaker 2: experience harm from its use. In gender shades doctor Blomini's 77 00:04:49,480 --> 00:04:52,719 Speaker 2: groundbreaking research, she was among the first to uncover the 78 00:04:52,800 --> 00:04:57,719 Speaker 2: gender and racial biases that plague facial recognition technology. But ever, 79 00:04:57,800 --> 00:05:02,520 Speaker 2: the poet Doctor blomini spoken poem ai ain't ti a woman? 80 00:05:03,000 --> 00:05:06,440 Speaker 2: Really brings the problem to life? Or AI misgenders and 81 00:05:06,480 --> 00:05:10,440 Speaker 2: misidentifies famous black women in history like Michelle Obama, who 82 00:05:10,520 --> 00:05:14,839 Speaker 2: facial recognition recognizes as a young man wearing a two pay. 83 00:05:15,360 --> 00:05:18,840 Speaker 1: And so from that gender shades research project came the 84 00:05:18,960 --> 00:05:23,520 Speaker 1: art piece that is AI ain't I a woman? Michelle 85 00:05:23,520 --> 00:05:27,159 Speaker 1: Obama unabashed and unafraid to wear her crown of history, 86 00:05:27,440 --> 00:05:30,599 Speaker 1: yet her crown seems a mystery to systems unsure of 87 00:05:30,680 --> 00:05:33,520 Speaker 1: her hair, a wig of a fond a too pay 88 00:05:33,640 --> 00:05:36,440 Speaker 1: maybe not? Are there no words for our brains. 89 00:05:36,200 --> 00:05:36,880 Speaker 3: And our logs? 90 00:05:37,320 --> 00:05:42,040 Speaker 1: Where I show tech companies you've probably heard of failing 91 00:05:42,160 --> 00:05:47,160 Speaker 1: on the iconic women of people like Oprah Winfrey, Serena Williams, 92 00:05:47,279 --> 00:05:51,520 Speaker 1: Michelle Obama, historic figures like sojourn nor truth. Hence the 93 00:05:51,560 --> 00:05:55,640 Speaker 1: title AI AIN'TI a woman? And I saw that when 94 00:05:55,720 --> 00:06:00,719 Speaker 1: I shared that poem. It's a video poem in all 95 00:06:00,839 --> 00:06:06,719 Speaker 1: kinds of spaces, right you, defense ministers, you know, kids 96 00:06:06,760 --> 00:06:12,200 Speaker 1: in middle school. It touched people's humanity in a way 97 00:06:12,240 --> 00:06:16,480 Speaker 1: that the research couldn't, And that for me was really 98 00:06:16,839 --> 00:06:21,000 Speaker 1: an important moment because for a long time I felt 99 00:06:21,000 --> 00:06:25,760 Speaker 1: that I couldn't bring my art into my research because 100 00:06:25,800 --> 00:06:28,920 Speaker 1: it might not be taken as seriously or it might 101 00:06:29,040 --> 00:06:33,520 Speaker 1: lessen its impact. And I found just the opposite. When 102 00:06:33,520 --> 00:06:37,240 Speaker 1: you humanize what's going on, it extends the reach of 103 00:06:37,279 --> 00:06:39,599 Speaker 1: the people who feel they have a place in the 104 00:06:39,640 --> 00:06:44,279 Speaker 1: conversation about AI or even like, oh, this is how 105 00:06:44,360 --> 00:06:48,440 Speaker 1: it could matter to me, not some abstract Oh there's 106 00:06:48,480 --> 00:06:50,800 Speaker 1: discrimination or tech can be harmful. 107 00:06:51,880 --> 00:06:55,760 Speaker 2: These harms aren't abstract or theoretical. They're very real. And 108 00:06:55,760 --> 00:06:59,200 Speaker 2: they're already happening. We talked about Portia Woodrif on the 109 00:06:59,200 --> 00:07:02,480 Speaker 2: podcast before. She was heavily pregnant when she was falsely 110 00:07:02,560 --> 00:07:05,480 Speaker 2: arrested and held for hours and needed to be hospitalized 111 00:07:05,800 --> 00:07:09,840 Speaker 2: after being falsely arrested when police facial recognition misidentified her 112 00:07:09,880 --> 00:07:12,280 Speaker 2: as a suspect that a carjacking she had nothing to 113 00:07:12,320 --> 00:07:15,600 Speaker 2: do with. And she's not the only one. Back in 114 00:07:15,640 --> 00:07:19,400 Speaker 2: twenty twenty, Robert Williams, a black man, became the first 115 00:07:19,440 --> 00:07:22,680 Speaker 2: documented case of a person being falsely arrested thanks to 116 00:07:22,720 --> 00:07:26,720 Speaker 2: the use of faulty facial recognition technology. Robert was arrested 117 00:07:26,720 --> 00:07:30,080 Speaker 2: in front of his daughters after facial recognition mismatched his 118 00:07:30,160 --> 00:07:33,160 Speaker 2: driver's license photo to someone who stole watches from a 119 00:07:33,160 --> 00:07:36,200 Speaker 2: Shanola store in Detroit, but Robert had nothing to do 120 00:07:36,240 --> 00:07:39,000 Speaker 2: with it. Tools like turn it in that are used 121 00:07:39,000 --> 00:07:42,280 Speaker 2: to detect students cheating by turning in AI generated assignments 122 00:07:42,600 --> 00:07:46,760 Speaker 2: routinely falsely accused of students of plagiarizing. According to the markup, 123 00:07:46,960 --> 00:07:49,560 Speaker 2: the technology is much more likely to generate a false 124 00:07:49,600 --> 00:07:52,880 Speaker 2: positive for international students and students who are non native 125 00:07:52,920 --> 00:07:56,720 Speaker 2: English speakers. A group of Stanford Computer scientists found that 126 00:07:56,840 --> 00:08:00,000 Speaker 2: seven different AI detectors flagged writing by non native speed 127 00:08:00,000 --> 00:08:03,480 Speaker 2: makers as AI generated sixty one percent of the time 128 00:08:03,880 --> 00:08:07,000 Speaker 2: on about twenty percent of the papers. That incorrect assessment 129 00:08:07,040 --> 00:08:11,280 Speaker 2: was unanimous. Meanwhile, those same detectors almost never made such 130 00:08:11,320 --> 00:08:15,280 Speaker 2: mistakes when assessing the writing of native English speakers. Obviously, 131 00:08:15,520 --> 00:08:19,280 Speaker 2: these kinds of accusations could throw vulnerable students academic careers 132 00:08:19,280 --> 00:08:22,880 Speaker 2: into turmoil. The people like Porscha and the international students 133 00:08:23,080 --> 00:08:26,080 Speaker 2: speak up when they've experienced harms because of faulty technology, 134 00:08:26,440 --> 00:08:29,920 Speaker 2: So are the powers that be listening? Do their experiences 135 00:08:30,000 --> 00:08:32,360 Speaker 2: matter as much as the companies trying to make money 136 00:08:32,400 --> 00:08:34,040 Speaker 2: from rolling out this technology do. 137 00:08:35,160 --> 00:08:38,760 Speaker 1: But an example like Porscha Woodruff falsely arrested due to 138 00:08:38,920 --> 00:08:43,840 Speaker 1: AI powered facial recognition. She was eight months pregnant, sitting 139 00:08:43,880 --> 00:08:48,080 Speaker 1: in a jail cell, having contractions for crime. Now she 140 00:08:48,320 --> 00:08:52,080 Speaker 1: was being held for a crime she didn't commit. And 141 00:08:52,120 --> 00:08:54,480 Speaker 1: so when you hear those stories, the stories of who 142 00:08:54,559 --> 00:08:57,440 Speaker 1: I like to call the X coded, you start to 143 00:08:57,520 --> 00:09:00,719 Speaker 1: pay attention, right or maybe it's your kids and they 144 00:09:00,760 --> 00:09:04,240 Speaker 1: got flagged for cheating. Turns out they didn't actually cheat. 145 00:09:04,280 --> 00:09:10,240 Speaker 1: English is their second language. But some chat GPT detection system, 146 00:09:10,679 --> 00:09:14,120 Speaker 1: right is flagging them as cheating. And so I do 147 00:09:14,200 --> 00:09:18,640 Speaker 1: think those stories are what helps people see that this 148 00:09:18,720 --> 00:09:23,640 Speaker 1: is a conversation that requires their voice. And it's so 149 00:09:23,760 --> 00:09:26,640 Speaker 1: easy to think it's like I have a PhD from MIT, 150 00:09:26,800 --> 00:09:28,720 Speaker 1: but I was doing all this before. Right, You don't 151 00:09:28,720 --> 00:09:33,640 Speaker 1: have to have this type of in depth technical background 152 00:09:34,080 --> 00:09:36,960 Speaker 1: to have a voice and to have an important perspective, 153 00:09:37,080 --> 00:09:40,439 Speaker 1: because if you know you're being hard, that's enough. Yeah. 154 00:09:40,520 --> 00:09:42,800 Speaker 2: A big part of what we aim to do here 155 00:09:42,960 --> 00:09:45,200 Speaker 2: is to help people understand that you might not be 156 00:09:45,240 --> 00:09:48,120 Speaker 2: an engineer, you might not have a doctorate, but you 157 00:09:48,160 --> 00:09:50,319 Speaker 2: are the expert of your experience, and you use this 158 00:09:50,360 --> 00:09:52,520 Speaker 2: technology every day or it's being used on you, and 159 00:09:52,559 --> 00:09:56,880 Speaker 2: so you innately have a perspective that is valuable and 160 00:09:56,960 --> 00:10:00,079 Speaker 2: worth sharing and worth hearing and worth centering about how 161 00:10:00,120 --> 00:10:01,760 Speaker 2: that technology has impacted you. 162 00:10:06,120 --> 00:10:26,240 Speaker 3: Let's take a quick break. 163 00:10:17,920 --> 00:10:23,960 Speaker 2: At er back. People who are traditionally marginalized, like women 164 00:10:24,080 --> 00:10:27,840 Speaker 2: and black women, are made invisible by technology. Every single 165 00:10:27,920 --> 00:10:32,360 Speaker 2: day we're faced with silencing, erasure, and hostility. So is 166 00:10:32,400 --> 00:10:34,720 Speaker 2: that one of the reasons why the technology that these 167 00:10:34,720 --> 00:10:38,920 Speaker 2: spaces build also can't really see us? Do you ever 168 00:10:39,040 --> 00:10:42,160 Speaker 2: feel I mean this stay with me here. I think 169 00:10:42,160 --> 00:10:46,920 Speaker 2: that there is like a general hostility toward marginalized people, 170 00:10:47,000 --> 00:10:50,600 Speaker 2: like toward black women. And I think in technology and 171 00:10:50,640 --> 00:10:53,600 Speaker 2: I sometimes feel that the technology that is being made 172 00:10:53,960 --> 00:10:58,160 Speaker 2: in turn mirrors that same hostility, mirrors that same erasure. 173 00:10:58,200 --> 00:11:02,439 Speaker 2: And so because these facial recognition is not being tested 174 00:11:02,600 --> 00:11:05,120 Speaker 2: or trained out us, you know, diverse data sets or 175 00:11:05,160 --> 00:11:08,680 Speaker 2: whatever it in turn erases us, do you feel that 176 00:11:08,720 --> 00:11:12,520 Speaker 2: there's like a that is kind of because of this 177 00:11:12,720 --> 00:11:17,240 Speaker 2: underlying hostility toward people who are traditionally marginalized in the space. 178 00:11:18,240 --> 00:11:25,160 Speaker 1: I actually don't think it's a intentional underlying hostility, which 179 00:11:25,240 --> 00:11:31,600 Speaker 1: makes it even more dangerous. So well intentioned people right 180 00:11:32,440 --> 00:11:37,640 Speaker 1: collecting data, doing what they think is good science, good 181 00:11:37,720 --> 00:11:41,880 Speaker 1: machine learning, can still create harmful systems. And this is 182 00:11:41,920 --> 00:11:44,960 Speaker 1: what I learned after we did different audits, and I 183 00:11:44,960 --> 00:11:48,240 Speaker 1: would go talk to the teams behind some of these systems, right. 184 00:11:48,280 --> 00:11:51,360 Speaker 1: They are nice people, you know, try to send the 185 00:11:51,440 --> 00:11:56,680 Speaker 1: kids to college, right. And it was actually interesting to 186 00:11:56,720 --> 00:11:59,320 Speaker 1: me because as much as I was wanting to humanize 187 00:11:59,760 --> 00:12:03,040 Speaker 1: the people who are excoded, who are harmed by AI systems, 188 00:12:03,280 --> 00:12:04,920 Speaker 1: part of what I try to do in the book 189 00:12:04,960 --> 00:12:08,120 Speaker 1: as well is also humanize the people who are creating 190 00:12:08,160 --> 00:12:11,920 Speaker 1: the systems and where things go wrong. But to your 191 00:12:12,000 --> 00:12:17,720 Speaker 1: greater point, it's not an intentional hostility. Sometimes it is 192 00:12:18,120 --> 00:12:22,200 Speaker 1: a profound and harmful ignorance to not even think to 193 00:12:22,440 --> 00:12:27,040 Speaker 1: ask certain questions or to test the system in particular ways. 194 00:12:27,120 --> 00:12:30,080 Speaker 1: That's part of what the research was about. We asked 195 00:12:30,080 --> 00:12:33,719 Speaker 1: what happens when we put an intersectional lens on the 196 00:12:33,760 --> 00:12:37,599 Speaker 1: way in which we analyze the performance of AI systems, 197 00:12:37,960 --> 00:12:41,760 Speaker 1: and just doing that right opened up new areas of 198 00:12:41,840 --> 00:12:45,840 Speaker 1: conversations where before people would just look at the overall score, 199 00:12:46,800 --> 00:12:49,800 Speaker 1: and that gave us a false sense of progress within 200 00:12:49,880 --> 00:12:54,280 Speaker 1: the space. Because we were testing these systems on benchmark 201 00:12:54,360 --> 00:12:56,840 Speaker 1: data sets to see how well they do, and then 202 00:12:56,880 --> 00:12:59,280 Speaker 1: you would look at the benchmarks, some of the benchmarks 203 00:12:59,280 --> 00:13:03,480 Speaker 1: would be over eighty percent, lighter skinned individuals over seventy percent, 204 00:13:03,559 --> 00:13:07,520 Speaker 1: you know, people identified as men. And if that is 205 00:13:07,679 --> 00:13:12,720 Speaker 1: your benchmark of success, you're already not going to see 206 00:13:12,840 --> 00:13:16,680 Speaker 1: how you're failing. And so when we created a more 207 00:13:16,679 --> 00:13:20,360 Speaker 1: inclusive data set, et cetera, it allowed us to see 208 00:13:20,440 --> 00:13:25,160 Speaker 1: that the promises of potentially well intentioned people weren't even 209 00:13:25,240 --> 00:13:28,360 Speaker 1: panning out But there's even more to that, because I 210 00:13:28,440 --> 00:13:31,440 Speaker 1: think with some of these conversations it can seem like 211 00:13:31,480 --> 00:13:37,360 Speaker 1: it's a very technical problem with a technical solution. Data 212 00:13:37,440 --> 00:13:41,080 Speaker 1: didn't detect, you know, system didn't detect the face, make 213 00:13:41,120 --> 00:13:45,400 Speaker 1: it more inclusive, whatever else it is. But the problem 214 00:13:45,480 --> 00:13:50,839 Speaker 1: is accurate systems can be abuse. We've seen facial recognition 215 00:13:51,440 --> 00:13:56,440 Speaker 1: systems deployed at protests right, which we know can lead 216 00:13:56,480 --> 00:13:59,400 Speaker 1: to chilling effects if you know you're under surveillance for 217 00:14:00,160 --> 00:14:04,240 Speaker 1: daring to exercise your First Amendment rights to say this 218 00:14:04,320 --> 00:14:10,720 Speaker 1: is not correct. Accurate systems create tools for state surveillance. 219 00:14:11,280 --> 00:14:13,480 Speaker 1: So yes, you can say, well, our my phone tracks 220 00:14:13,520 --> 00:14:16,240 Speaker 1: me done. The other you can leave your phone at home. 221 00:14:16,880 --> 00:14:21,600 Speaker 1: Your face is a little bit harder. You know, some 222 00:14:21,720 --> 00:14:23,400 Speaker 1: people do put on a face, but you know what 223 00:14:23,440 --> 00:14:26,280 Speaker 1: I mean. You know what I mean, right, So I 224 00:14:26,360 --> 00:14:30,800 Speaker 1: think it's important to understand that even when we have 225 00:14:30,920 --> 00:14:35,360 Speaker 1: conversations about the accuracy of certain systems, and we should 226 00:14:35,440 --> 00:14:40,600 Speaker 1: have those conversations, accuracy is not enough to assure accountability 227 00:14:40,760 --> 00:14:41,440 Speaker 1: or equity. 228 00:14:42,160 --> 00:14:45,600 Speaker 2: Now, when we're talking about accountability, especially from tech companies, 229 00:14:45,800 --> 00:14:48,040 Speaker 2: it is so easy to get caught in a cycle 230 00:14:48,080 --> 00:14:50,440 Speaker 2: of name and shame where you point out all the 231 00:14:50,520 --> 00:14:53,440 Speaker 2: bad things that a specific company has done. And if 232 00:14:53,480 --> 00:14:56,000 Speaker 2: I'm being honest, I might have done that a time 233 00:14:56,120 --> 00:14:59,680 Speaker 2: or two on this very podcast. But doctor Bloomwini describes 234 00:14:59,720 --> 00:15:03,480 Speaker 2: their as less name and shame and more name and change. 235 00:15:03,680 --> 00:15:05,800 Speaker 2: They want to show companies what they're doing wrong so 236 00:15:05,840 --> 00:15:08,600 Speaker 2: they can change for the better. But this hasn't always 237 00:15:08,640 --> 00:15:11,760 Speaker 2: meant that those companies don't lash out when her teams 238 00:15:11,840 --> 00:15:15,560 Speaker 2: point out the harm that they've caused. I'm curious, how 239 00:15:15,720 --> 00:15:19,960 Speaker 2: have companies I won't say any names, but companies who 240 00:15:20,000 --> 00:15:22,840 Speaker 2: you have called out in your research or you know, 241 00:15:22,920 --> 00:15:25,360 Speaker 2: said like this is hey, this is what's going on. 242 00:15:25,440 --> 00:15:27,600 Speaker 2: How have they responded to your findings? 243 00:15:28,800 --> 00:15:32,360 Speaker 1: So overall, I take a name and change approach. So 244 00:15:32,520 --> 00:15:36,760 Speaker 1: the point of pointing out what's wrong isn't to shame 245 00:15:37,360 --> 00:15:41,160 Speaker 1: a company, is to say we can do better. Right, 246 00:15:42,000 --> 00:15:46,640 Speaker 1: And sometimes we have companies that are reactive. We have 247 00:15:46,720 --> 00:15:50,800 Speaker 1: companies that are proactive, and we have companies that are combative. 248 00:15:51,400 --> 00:15:55,480 Speaker 1: With the first set of research results that we released, 249 00:15:55,600 --> 00:16:00,520 Speaker 1: we saw more of the reactive stance, which is is oh, 250 00:16:00,840 --> 00:16:04,480 Speaker 1: now that there's a headline, right, we're gonna go. We 251 00:16:04,600 --> 00:16:08,280 Speaker 1: are on the problem or we have or we were 252 00:16:08,320 --> 00:16:11,480 Speaker 1: already working on the problem. You know, they're different ways, 253 00:16:12,720 --> 00:16:15,760 Speaker 1: but now it's a priority because it's making headlines. So 254 00:16:15,840 --> 00:16:19,320 Speaker 1: I saw that and the reactive approach tended to be 255 00:16:19,720 --> 00:16:23,640 Speaker 1: a technical approach, which is okay, there were these disparities, 256 00:16:23,720 --> 00:16:27,960 Speaker 1: so let's close them. We now have more accurate XYZ. Again, 257 00:16:28,040 --> 00:16:32,920 Speaker 1: accurate systems can be of use. Then we did experience 258 00:16:33,720 --> 00:16:39,520 Speaker 1: some combative responses. Right, so here we had a huge 259 00:16:39,600 --> 00:16:44,479 Speaker 1: tech company coming out and saying your research is misleading, 260 00:16:45,440 --> 00:16:49,120 Speaker 1: attempting to discredit the research of At that time, I 261 00:16:49,160 --> 00:16:52,440 Speaker 1: was a graduate student, and I was so fortunate, you 262 00:16:52,480 --> 00:16:56,920 Speaker 1: know that I had senior scholars and people well respected 263 00:16:56,960 --> 00:17:00,880 Speaker 1: in the AI industry who came to our offense, you know, 264 00:17:01,040 --> 00:17:04,560 Speaker 1: cheering prize winner, somebody who was literally the chief AI 265 00:17:04,640 --> 00:17:11,680 Speaker 1: scientist at that company, saying, what the research shows warrants 266 00:17:11,720 --> 00:17:16,640 Speaker 1: our attention, and this is research we should be elevating, 267 00:17:17,160 --> 00:17:21,199 Speaker 1: not dismissing, because it makes the field better as a hold, 268 00:17:21,680 --> 00:17:25,920 Speaker 1: if we can acknowledge our limitations, understand what's going wrong, 269 00:17:26,000 --> 00:17:29,119 Speaker 1: so we can build more robust systems. Because this doesn't 270 00:17:29,160 --> 00:17:31,160 Speaker 1: just deal with faces. Right, If you want to use 271 00:17:31,160 --> 00:17:35,600 Speaker 1: computer vision to help. Let's say, with medical diagnoses, you 272 00:17:35,640 --> 00:17:37,800 Speaker 1: want to make sure you understand where things can go 273 00:17:37,880 --> 00:17:40,680 Speaker 1: wrong so we can course correct for things to go right. 274 00:17:41,480 --> 00:17:45,240 Speaker 1: So we had the combative approach, the reactive approach, but 275 00:17:45,320 --> 00:17:50,720 Speaker 1: the approach I appreciate most is the proactive. Bro Okay, 276 00:17:51,200 --> 00:17:54,400 Speaker 1: we've heard there's some issues. Instead of waiting for someone 277 00:17:54,560 --> 00:17:57,919 Speaker 1: to drop the paper or the headline, what can we 278 00:17:58,040 --> 00:18:01,760 Speaker 1: be doing as a company now. And I've had the 279 00:18:01,800 --> 00:18:05,000 Speaker 1: opportunity to work with Procter and Gamble with Olay on 280 00:18:05,080 --> 00:18:09,200 Speaker 1: the Decode the Bias campaign. And when they came to 281 00:18:09,240 --> 00:18:12,120 Speaker 1: me and they asked for an algorithmic audit, I said, 282 00:18:13,080 --> 00:18:16,000 Speaker 1: given what you've described, your tool does and this was 283 00:18:16,040 --> 00:18:18,240 Speaker 1: a tool that would analyze your skin and give you 284 00:18:18,320 --> 00:18:22,199 Speaker 1: product recommendations. Ah, and how you train the tool on 285 00:18:22,240 --> 00:18:25,800 Speaker 1: a set of data. I suspect if we dig in there, 286 00:18:25,880 --> 00:18:28,800 Speaker 1: we're gonna find some bias. They're like, that's okay. If 287 00:18:28,840 --> 00:18:31,320 Speaker 1: you find bias, we'll do what we can to correct it. 288 00:18:31,880 --> 00:18:33,760 Speaker 1: I was like, and if you can't correct it, we'll 289 00:18:33,760 --> 00:18:36,399 Speaker 1: shut the system down. Like can I get this in writing? 290 00:18:36,480 --> 00:18:39,880 Speaker 1: I never I never hear this. I never hear this. 291 00:18:40,400 --> 00:18:42,919 Speaker 1: And then my other question was if we did an audit, 292 00:18:42,960 --> 00:18:47,240 Speaker 1: could we publish the audit results, because that adds another 293 00:18:47,320 --> 00:18:50,399 Speaker 1: level of transparency, so it's not just up, we got checked, 294 00:18:50,400 --> 00:18:53,080 Speaker 1: but no one knows what happened. Right. They agreed to 295 00:18:53,160 --> 00:18:56,440 Speaker 1: all of those things we did and in fact find 296 00:18:56,480 --> 00:18:59,919 Speaker 1: bias as we thought would be there, and they actually 297 00:19:00,440 --> 00:19:04,360 Speaker 1: and the proactive Not only did they seek to be audited, 298 00:19:04,840 --> 00:19:09,840 Speaker 1: they also agree to a consented data promise. And this 299 00:19:10,000 --> 00:19:13,240 Speaker 1: was inspired by their skin promise. So when I first 300 00:19:13,240 --> 00:19:15,880 Speaker 1: started working with LA I was excited. And then they 301 00:19:15,880 --> 00:19:19,320 Speaker 1: told me, you know, when we do the campaigns, there 302 00:19:19,359 --> 00:19:25,200 Speaker 1: won't be any post production air brushing. What we capture 303 00:19:25,520 --> 00:19:29,280 Speaker 1: is what will show, right, you know, truth and advertising. 304 00:19:29,480 --> 00:19:32,600 Speaker 1: You want to think about people's body image and all 305 00:19:32,680 --> 00:19:34,600 Speaker 1: of that. And I'll be honest, I was a little 306 00:19:34,640 --> 00:19:38,639 Speaker 1: disappointed because, like, if you you just want to know, 307 00:19:38,720 --> 00:19:40,800 Speaker 1: you could be saved by the airbrush. 308 00:19:41,040 --> 00:19:41,359 Speaker 3: Right. 309 00:19:41,880 --> 00:19:46,680 Speaker 1: But because of that promise, which is a good promise, 310 00:19:46,760 --> 00:19:49,280 Speaker 1: I get where they're coming from. As the person on 311 00:19:49,320 --> 00:19:52,520 Speaker 1: the other side of the four K camera in your face, 312 00:19:53,200 --> 00:19:58,520 Speaker 1: you're asking, can we consider XYZ? But what I appreciated 313 00:19:58,560 --> 00:20:01,720 Speaker 1: about that is it made me even more disciplined with 314 00:20:02,000 --> 00:20:08,359 Speaker 1: my actual skincare regimen, and I also drank water, and 315 00:20:08,880 --> 00:20:11,840 Speaker 1: I did all the right things for vanity reasons. I 316 00:20:11,840 --> 00:20:17,239 Speaker 1: won't I did the right things for vanity reasons, but 317 00:20:17,600 --> 00:20:22,360 Speaker 1: I think about that with so that skincare that promise, right, 318 00:20:22,400 --> 00:20:26,680 Speaker 1: that was part of the inspiration for the consented data promise. 319 00:20:26,800 --> 00:20:29,080 Speaker 1: And just like when you make a promise and it's 320 00:20:29,119 --> 00:20:33,679 Speaker 1: a public commitment, that's the important part. Now there's a 321 00:20:33,680 --> 00:20:37,320 Speaker 1: little bit of accountability. Right. So now you are going 322 00:20:37,359 --> 00:20:40,440 Speaker 1: to bed early. Now you are drinking more water. Now 323 00:20:40,480 --> 00:20:43,960 Speaker 1: you are exercising five days of the week, which might 324 00:20:44,000 --> 00:20:47,200 Speaker 1: not have been the case before you made that public. 325 00:20:47,920 --> 00:20:51,439 Speaker 1: So that's an example of more proactive. So from the 326 00:20:51,520 --> 00:20:55,560 Speaker 1: reactive to the combative to the proactive, we've seen it all. 327 00:20:56,080 --> 00:21:02,480 Speaker 1: I think what I learned most from the bative response 328 00:21:03,240 --> 00:21:05,880 Speaker 1: was how much of a risk I was taking as 329 00:21:05,920 --> 00:21:11,719 Speaker 1: a young researcher to not only do their research, but 330 00:21:11,920 --> 00:21:15,040 Speaker 1: to name the companies and I'll name them now, right, 331 00:21:15,240 --> 00:21:19,000 Speaker 1: that I tested IBM, that I tested Microsoft, that I 332 00:21:19,119 --> 00:21:26,280 Speaker 1: tested Amazon, and because of their power, that meant I 333 00:21:26,480 --> 00:21:31,920 Speaker 1: was risking future opportunities and also other researchers watching how 334 00:21:32,000 --> 00:21:36,560 Speaker 1: I was treated, how my co authors were treated people 335 00:21:36,640 --> 00:21:39,920 Speaker 1: like doctor Gabru. They were also getting a sense of 336 00:21:40,359 --> 00:21:44,359 Speaker 1: what is possible When I look at research papers now 337 00:21:44,760 --> 00:21:49,040 Speaker 1: where people openly talk about algorithmic bias and algorithmic harms, 338 00:21:49,040 --> 00:21:53,720 Speaker 1: and people openly name the AI models or the tech companies. 339 00:21:53,760 --> 00:21:55,560 Speaker 1: That wasn't always the case. 340 00:21:56,320 --> 00:21:56,560 Speaker 3: Right. 341 00:21:57,160 --> 00:22:03,280 Speaker 1: A price was paid for this more robust conversation to happen. 342 00:22:05,920 --> 00:22:08,639 Speaker 2: Doctor Blomini is right, this all comes out of cost. 343 00:22:09,119 --> 00:22:12,520 Speaker 2: Doctor tim Net Gabrew, who she mentioned earlier, was a 344 00:22:12,520 --> 00:22:16,399 Speaker 2: co author on doctor Blomini's gender shades research. Doctor Cabrew 345 00:22:16,560 --> 00:22:19,200 Speaker 2: was once the technical co lead of the Ethical Artificial 346 00:22:19,200 --> 00:22:22,600 Speaker 2: Intelligence team at Google. While in that role, she worked 347 00:22:22,640 --> 00:22:25,040 Speaker 2: on a paper about the risks of large language models 348 00:22:25,160 --> 00:22:29,679 Speaker 2: from environmental impact to bias. Google demanded she withdraw the paper. 349 00:22:30,200 --> 00:22:33,639 Speaker 2: It got contentious, the conversation was hostile, and the whole 350 00:22:33,640 --> 00:22:38,240 Speaker 2: thing was highly gendered and racialized. Doctor Cabrew was belittled, discredited, 351 00:22:38,240 --> 00:22:40,840 Speaker 2: and harassed online, and it ended with her termination. 352 00:22:42,440 --> 00:22:46,280 Speaker 1: It cost us something for doctor Gabru, you know, it 353 00:22:46,359 --> 00:22:49,920 Speaker 1: cost her her job to speak up when she saw 354 00:22:50,000 --> 00:22:53,840 Speaker 1: some of the issues that we see in what they 355 00:22:53,880 --> 00:22:57,199 Speaker 1: call large language models, the type of AI systems that 356 00:22:57,320 --> 00:23:03,240 Speaker 1: will power chat GPT right, and so I do think 357 00:23:03,359 --> 00:23:08,399 Speaker 1: the timing of the different types of company responses also 358 00:23:09,720 --> 00:23:13,200 Speaker 1: made a difference in my own trajectory. The first response 359 00:23:13,240 --> 00:23:17,199 Speaker 1: I had when the gender Shades paper was published was 360 00:23:17,320 --> 00:23:20,520 Speaker 1: IBM invited me to their headquarters. You know, I spoke 361 00:23:20,600 --> 00:23:23,400 Speaker 1: to their team members. They actually had released a new 362 00:23:23,440 --> 00:23:26,199 Speaker 1: model by the time I was presenting that research, and 363 00:23:26,240 --> 00:23:29,080 Speaker 1: I could share what their results had been. And then 364 00:23:29,160 --> 00:23:32,399 Speaker 1: later we did our own study. So that was a 365 00:23:32,560 --> 00:23:36,120 Speaker 1: very different reception. That reception gave me home. I was like, Okay, 366 00:23:36,160 --> 00:23:40,439 Speaker 1: all right, let's work together. Amazon situation. I don't know, 367 00:23:40,840 --> 00:23:45,639 Speaker 1: I don't know how corporate anymore sort of thing. But 368 00:23:45,760 --> 00:23:50,280 Speaker 1: I'm being you know, I'm putting these as more extreme cases. 369 00:23:50,400 --> 00:23:54,320 Speaker 1: But the point being, we can't really just wait on 370 00:23:54,600 --> 00:23:57,520 Speaker 1: if a company's going to choose to be reactive, proactive, 371 00:23:57,600 --> 00:24:01,800 Speaker 1: or combative. What we really need are laws and regulations 372 00:24:02,400 --> 00:24:06,600 Speaker 1: that don't rely on the goodwill of companies. Yeah. 373 00:24:06,680 --> 00:24:09,479 Speaker 2: I mean, I have to ask, when you were this 374 00:24:09,600 --> 00:24:13,240 Speaker 2: young researcher naming these companies in your in your findings, 375 00:24:13,560 --> 00:24:16,280 Speaker 2: did you know that you were taking on such a 376 00:24:16,320 --> 00:24:19,600 Speaker 2: personal risk or were you like, oh wait, glad it 377 00:24:19,800 --> 00:24:24,520 Speaker 2: worked out. Glad people had my back. Did you know 378 00:24:24,680 --> 00:24:26,560 Speaker 2: that that was a risk that you were encourring and 379 00:24:26,600 --> 00:24:28,040 Speaker 2: did it anyway or did you sort of do you 380 00:24:28,080 --> 00:24:30,400 Speaker 2: sort of look back and think like, Wow, I'm really 381 00:24:30,400 --> 00:24:31,200 Speaker 2: glad that worked out. 382 00:24:32,000 --> 00:24:37,840 Speaker 1: I knew that once the research was published it would 383 00:24:37,880 --> 00:24:41,639 Speaker 1: be questioned, so before it was published, I actually sat 384 00:24:41,760 --> 00:24:42,960 Speaker 1: with a law clinic. 385 00:24:43,640 --> 00:24:43,880 Speaker 3: Right. 386 00:24:44,080 --> 00:24:47,440 Speaker 1: We went through what could be said, right, what might 387 00:24:47,560 --> 00:24:50,280 Speaker 1: actually put you in legal jeopardy, and so forth. So 388 00:24:50,880 --> 00:24:54,480 Speaker 1: I didn't go into the situation not thinking there might 389 00:24:54,560 --> 00:24:57,919 Speaker 1: be blowback. I was actually surprised with the first round 390 00:24:58,000 --> 00:25:01,600 Speaker 1: we prepared and they're like, oh, yeah, okay, these are issues. 391 00:25:01,640 --> 00:25:05,399 Speaker 1: Come to the headquarters XYZ, we released new models, et cetera. 392 00:25:06,880 --> 00:25:10,639 Speaker 1: The blowback that I got with the second paper is 393 00:25:10,720 --> 00:25:14,880 Speaker 1: what I had thought I might experience, but just the 394 00:25:14,920 --> 00:25:18,320 Speaker 1: magnitude of it I wasn't ready for. I remember what 395 00:25:18,480 --> 00:25:23,280 Speaker 1: the film Coded Bias available on Netflix. It shares part 396 00:25:23,320 --> 00:25:28,320 Speaker 1: of the story of graduate students Darnalgorithmic Justice League, and 397 00:25:28,359 --> 00:25:34,879 Speaker 1: examples of people experiencing real world AI harms. The people 398 00:25:34,920 --> 00:25:39,280 Speaker 1: to provide the insurance for that film were nervous because 399 00:25:39,320 --> 00:25:42,560 Speaker 1: we critiqued Amazon. It wasn't that Amazon had said anything. 400 00:25:42,680 --> 00:25:48,360 Speaker 1: It was just a acknowledgment of Amazon's power, right, And 401 00:25:48,520 --> 00:25:54,639 Speaker 1: so it didn't dawn on me just how powerful some 402 00:25:54,800 --> 00:25:58,840 Speaker 1: of these tech companies are. I remember being at an 403 00:26:00,000 --> 00:26:04,720 Speaker 1: international summit in Switzerland, and it was as if the 404 00:26:04,760 --> 00:26:09,000 Speaker 1: heads of the tech companies were heads of state, you know, 405 00:26:09,440 --> 00:26:16,400 Speaker 1: and so observing that closer made me really I was like, Oh, 406 00:26:16,440 --> 00:26:19,119 Speaker 1: I'm like, okay, I'm poking. It's like, okay, I'm poking 407 00:26:19,119 --> 00:26:22,880 Speaker 1: a dragon. I'm like, oh, it's a fire breathing dragon. Oh, 408 00:26:22,960 --> 00:26:24,320 Speaker 1: it's like a dragon dragon. 409 00:26:24,560 --> 00:26:26,000 Speaker 2: And like when you go to kill a bug and 410 00:26:26,040 --> 00:26:29,159 Speaker 2: you're like, oh, it's got wings, it flies, right. 411 00:26:29,680 --> 00:26:31,879 Speaker 1: So I knew, like, you know, it's not going to 412 00:26:31,920 --> 00:26:34,840 Speaker 1: be the best situation. But I don't think I was 413 00:26:35,000 --> 00:26:39,040 Speaker 1: fully prepared, though I thought I had prepared. 414 00:26:43,280 --> 00:26:43,760 Speaker 3: More after a. 415 00:26:43,800 --> 00:26:59,440 Speaker 2: Quick break, let's get right back into it. Even though 416 00:26:59,480 --> 00:27:03,080 Speaker 2: Amazon tried publicly discrediting doctor Blomini's work calling out the 417 00:27:03,119 --> 00:27:06,520 Speaker 2: harms of their facial recognition technology, in the end, they 418 00:27:06,560 --> 00:27:10,879 Speaker 2: conceded that technology wasn't exactly safe. In twenty twenty, they 419 00:27:10,880 --> 00:27:13,520 Speaker 2: announced a pause on allowing police to use the technology, 420 00:27:13,720 --> 00:27:18,680 Speaker 2: and eventually extended that pause indefinitely and correct me if 421 00:27:18,720 --> 00:27:23,320 Speaker 2: I'm wrong. But your work ended up with Amazon rolling 422 00:27:23,400 --> 00:27:27,000 Speaker 2: back some of the uses of their faulty facial recognition technology. 423 00:27:27,000 --> 00:27:30,440 Speaker 2: So ultimately, not only were you obviously vindicated, but that 424 00:27:30,560 --> 00:27:35,040 Speaker 2: work went on to create a somewhat like safer landscape 425 00:27:35,040 --> 00:27:37,520 Speaker 2: for everyone because Amazon had to step back and be like, Okay, 426 00:27:37,520 --> 00:27:40,840 Speaker 2: wait a minute, this technology maybe isn't really working that well. 427 00:27:41,520 --> 00:27:43,560 Speaker 1: They would not put it in those terms, but they 428 00:27:43,560 --> 00:27:48,720 Speaker 1: did make they didn't take other steps. So I will 429 00:27:48,760 --> 00:27:53,399 Speaker 1: say before Amazon, IBM actually said we are no longer 430 00:27:53,440 --> 00:27:57,639 Speaker 1: going to sell to police departments. And this was in 431 00:27:57,720 --> 00:28:01,199 Speaker 1: twenty twenty, right, so we also had the murder of 432 00:28:01,280 --> 00:28:05,800 Speaker 1: George Floyd happening at that time, and Microsoft said we 433 00:28:05,880 --> 00:28:10,159 Speaker 1: will not sell this until regulations are in place. And 434 00:28:10,200 --> 00:28:14,360 Speaker 1: then Amazon came third, you know, and they said we'll 435 00:28:14,400 --> 00:28:17,520 Speaker 1: halt it for a year, and then they extended that halt. 436 00:28:18,760 --> 00:28:19,000 Speaker 3: Right. 437 00:28:19,400 --> 00:28:22,800 Speaker 1: But this is to say there was an acknowledgment right 438 00:28:22,920 --> 00:28:25,600 Speaker 1: of the risk and the harms. Was it just risk 439 00:28:25,640 --> 00:28:27,679 Speaker 1: and harms to people, or risk and harms to the 440 00:28:27,760 --> 00:28:33,520 Speaker 1: company's reputations. It could be a combination of both. I'm 441 00:28:33,560 --> 00:28:40,680 Speaker 1: excited to share that this research led to numerous cities, 442 00:28:41,080 --> 00:28:46,000 Speaker 1: you know, incorporating some of the findings in their analysis 443 00:28:46,040 --> 00:28:49,840 Speaker 1: and in their statements for why they chose to enact 444 00:28:50,440 --> 00:28:55,080 Speaker 1: certain laws that restrict police use of facial recognition. It 445 00:28:55,200 --> 00:29:00,600 Speaker 1: also changed the conversation at the national and international level 446 00:29:00,720 --> 00:29:07,520 Speaker 1: that EUAI Act actually has a provision that would prevent 447 00:29:07,920 --> 00:29:11,680 Speaker 1: the use of life facial recognition in public spaces. When 448 00:29:11,720 --> 00:29:15,600 Speaker 1: I spoke with President Biden at the AI Realm Table 449 00:29:16,440 --> 00:29:20,200 Speaker 1: some time ago, this was top of mind. I shared 450 00:29:20,200 --> 00:29:23,440 Speaker 1: the story of Robert Williams being wrongfly arrested in front 451 00:29:23,480 --> 00:29:27,320 Speaker 1: of his two daughters. We talked about racial bias in 452 00:29:27,360 --> 00:29:30,320 Speaker 1: AI systems and other types of harms that can impact 453 00:29:30,400 --> 00:29:33,360 Speaker 1: many people because no one, trust me, no one is immute. 454 00:29:33,640 --> 00:29:39,360 Speaker 1: This isn't just other people's problems. And so to see 455 00:29:39,560 --> 00:29:44,440 Speaker 1: the reach of that work certainly made all of the 456 00:29:44,480 --> 00:29:51,520 Speaker 1: combative you know, responses and things like that somehow worth it. 457 00:29:52,280 --> 00:29:54,160 Speaker 2: And it really goes back to what you were saying 458 00:29:54,200 --> 00:29:58,120 Speaker 2: earlier about how systems don't have to be biased to 459 00:29:58,120 --> 00:30:01,120 Speaker 2: be misused, and like, don't know, I want to believe 460 00:30:01,200 --> 00:30:05,320 Speaker 2: in a tech landscape where companies with so much power 461 00:30:06,080 --> 00:30:09,200 Speaker 2: don't have to wait until something goes wrong, don't have 462 00:30:09,280 --> 00:30:11,400 Speaker 2: to wait until somebody is wrongfully arrested, don't have to 463 00:30:11,400 --> 00:30:14,640 Speaker 2: wait until they're called out to make things a little 464 00:30:14,640 --> 00:30:16,840 Speaker 2: bit safer and more equitable, Like, do you believe that 465 00:30:16,960 --> 00:30:21,640 Speaker 2: is possible where companies aren't just reacting, they're actually you know, 466 00:30:21,840 --> 00:30:25,120 Speaker 2: being proactive at wanting the technology that they deploy on 467 00:30:25,160 --> 00:30:27,280 Speaker 2: all of us to be more equitable. 468 00:30:28,480 --> 00:30:32,120 Speaker 1: I think that again, you do see companies taking on 469 00:30:32,280 --> 00:30:36,760 Speaker 1: the mantle of responsible AI. You'll have other companies like 470 00:30:36,800 --> 00:30:40,400 Speaker 1: Credo dot Ai that will have services that are meant 471 00:30:40,440 --> 00:30:44,280 Speaker 1: to help companies adopting AI systems do it within a 472 00:30:44,320 --> 00:30:49,720 Speaker 1: responsible way. You'll see companies hiring responsible AI leads, right, 473 00:30:49,960 --> 00:30:54,360 Speaker 1: So I definitely think there is an intention there. Where 474 00:30:54,640 --> 00:30:58,600 Speaker 1: I still push back a bit is self regulation is 475 00:30:58,640 --> 00:31:05,120 Speaker 1: always self interested, not surprising, So I do think real 476 00:31:05,160 --> 00:31:09,960 Speaker 1: accountability requires external accountability. And the other part that I 477 00:31:10,000 --> 00:31:14,440 Speaker 1: don't really see companies focused on so much is redress. 478 00:31:14,960 --> 00:31:18,160 Speaker 1: So there's a lot of conversation about being responsible in 479 00:31:18,240 --> 00:31:21,800 Speaker 1: terms of preventing future harms, but what about those who 480 00:31:21,840 --> 00:31:26,520 Speaker 1: have been harmed already? And I do think algorithmic redress 481 00:31:26,560 --> 00:31:32,479 Speaker 1: is oftentimes missing from this conversation of responsible AI. So 482 00:31:32,960 --> 00:31:35,480 Speaker 1: when I see the company stepping out to say and 483 00:31:35,600 --> 00:31:39,320 Speaker 1: we're doing redress, I might be convinced I haven't seen 484 00:31:39,360 --> 00:31:41,920 Speaker 1: it yet, though, prove me wrong. Prove me wrong. I 485 00:31:41,960 --> 00:31:42,600 Speaker 1: want to be wrong. 486 00:31:44,040 --> 00:31:46,840 Speaker 2: Well as somebody at the helm of an organization fighting 487 00:31:46,880 --> 00:31:49,200 Speaker 2: for algorithmic justice, what does justice look like? 488 00:31:49,960 --> 00:31:52,920 Speaker 1: Justice looks like you live in a world where data 489 00:31:53,160 --> 00:31:55,960 Speaker 1: is not destiny, where you're hue is in a cue 490 00:31:55,960 --> 00:32:00,719 Speaker 1: to dismiss your humanity, where you actually have data rights 491 00:32:00,800 --> 00:32:06,360 Speaker 1: and you can consent to how your information is used. 492 00:32:06,640 --> 00:32:09,320 Speaker 1: Justice looks like how do we use these tools in 493 00:32:09,360 --> 00:32:14,880 Speaker 1: a joyful, uplifting manner versus just being reactive to the 494 00:32:14,920 --> 00:32:18,920 Speaker 1: next time. So when I think of social justice, you 495 00:32:18,960 --> 00:32:23,000 Speaker 1: can't have social justice without algorithmic justice. Because if you're 496 00:32:23,000 --> 00:32:25,920 Speaker 1: saying we're pushing for gender equality, yet you have an 497 00:32:25,960 --> 00:32:30,400 Speaker 1: AI system that cuts out women's resumes, we didn't quite 498 00:32:30,560 --> 00:32:34,160 Speaker 1: think it right. You can't necessarily say, oh, we have 499 00:32:34,360 --> 00:32:40,320 Speaker 1: racial equality and then you're adopting bias facial recognition. That's 500 00:32:40,360 --> 00:32:43,200 Speaker 1: putting so far. The folks I've seen have all been 501 00:32:43,280 --> 00:32:49,480 Speaker 1: dark skinned like us, you know, into prison due to misidentification, 502 00:32:49,800 --> 00:32:53,680 Speaker 1: And so me for me right, Algorithmic justice is truly 503 00:32:53,840 --> 00:32:57,240 Speaker 1: being in that place where we can be our full 504 00:32:57,400 --> 00:33:03,720 Speaker 1: selves and not be targeted right or algorithmically placed as 505 00:33:03,880 --> 00:33:09,400 Speaker 1: other algorithmically erased, algorithmically exploited. And so that's the world 506 00:33:10,080 --> 00:33:12,800 Speaker 1: we fight for. Right we say, free the X coded, 507 00:33:13,240 --> 00:33:16,360 Speaker 1: and so this is algorithmic justice. 508 00:33:17,120 --> 00:33:20,680 Speaker 2: The book is unmasking Ai your namesake, Joy. I have 509 00:33:20,760 --> 00:33:23,960 Speaker 2: to tell you you are such a joyful person. Speaking 510 00:33:23,960 --> 00:33:26,840 Speaker 2: to you about this work is it just comes through 511 00:33:26,840 --> 00:33:29,440 Speaker 2: how much you care and how I don't know. I 512 00:33:29,440 --> 00:33:32,479 Speaker 2: have a lot of conversations about tech that are hard 513 00:33:32,760 --> 00:33:36,680 Speaker 2: and dark and grim, and your work is just so 514 00:33:37,000 --> 00:33:40,520 Speaker 2: the opposite. It asks what if, what's possible? What can 515 00:33:40,600 --> 00:33:42,960 Speaker 2: we do? How can things be better? Like it's just 516 00:33:43,040 --> 00:33:46,400 Speaker 2: really nice to see somebody leading the way with such 517 00:33:46,440 --> 00:33:52,160 Speaker 2: a joyful but justice rooted perspective. It's so refreshing. 518 00:33:52,960 --> 00:33:56,960 Speaker 1: Thank you, and it's so wonderful to be in conversation 519 00:33:57,280 --> 00:34:01,560 Speaker 1: with you. I love supporting Fi bad Ass. You said 520 00:34:01,600 --> 00:34:04,920 Speaker 1: I could cuss women, so this is great. 521 00:34:06,960 --> 00:34:10,000 Speaker 2: So how can folks learn more about the Algorithmic Justice League. 522 00:34:10,239 --> 00:34:13,560 Speaker 1: I'm so glad you asked. We do have ww dot 523 00:34:13,600 --> 00:34:17,200 Speaker 1: AJL dot org and so we invite people to be 524 00:34:17,520 --> 00:34:22,600 Speaker 1: agents of change and join the Algorithmic Justice League. We 525 00:34:22,680 --> 00:34:25,480 Speaker 1: have a library there as well, so if you're new 526 00:34:25,520 --> 00:34:28,600 Speaker 1: to this area and you are curious you know what 527 00:34:28,840 --> 00:34:33,560 Speaker 1: is even AI, We've created our resources for you so 528 00:34:33,640 --> 00:34:36,480 Speaker 1: you can be part of the conversation. And we also 529 00:34:36,600 --> 00:34:40,360 Speaker 1: have an ex Coded Experiences platform, So, like you were saying, 530 00:34:40,760 --> 00:34:43,359 Speaker 1: you are the expert of your own lived experience, and 531 00:34:43,400 --> 00:34:47,680 Speaker 1: we value that expertise, so we do campaigns where people 532 00:34:47,719 --> 00:34:51,440 Speaker 1: can tell their stories of being ex coded. For some people, 533 00:34:51,520 --> 00:34:56,040 Speaker 1: we just launched a campaign about facial recognition in airports, 534 00:34:56,400 --> 00:34:58,799 Speaker 1: so people are sharing if they saw signage, if they 535 00:34:58,880 --> 00:35:01,880 Speaker 1: knew they could opt out, and all of that actually 536 00:35:01,920 --> 00:35:07,440 Speaker 1: builds a database of stories that shows if the TSA 537 00:35:07,640 --> 00:35:10,719 Speaker 1: or others are actually doing what they say. They said 538 00:35:10,760 --> 00:35:13,640 Speaker 1: it was optional. I didn't even know I could opt out. 539 00:35:13,719 --> 00:35:17,279 Speaker 1: We have a disconnect, but we also have the data, right, 540 00:35:17,600 --> 00:35:22,160 Speaker 1: So I do think people as they are encountering various 541 00:35:22,160 --> 00:35:26,320 Speaker 1: AI systems and they have questions or stories to share, 542 00:35:26,920 --> 00:35:29,640 Speaker 1: AJL is that place they can go to. 543 00:35:30,719 --> 00:35:34,000 Speaker 2: So please check out AJL it is you're doing such 544 00:35:34,000 --> 00:35:35,920 Speaker 2: incredible work. Thank you so much for being here, and 545 00:35:35,960 --> 00:35:37,759 Speaker 2: just thank you for being you. Thank you for being 546 00:35:37,800 --> 00:35:40,399 Speaker 2: in the space. We need more people like you. 547 00:35:40,600 --> 00:35:41,640 Speaker 1: Thank you for having me. 548 00:35:43,040 --> 00:35:45,920 Speaker 2: Black women like doctor Blowini have been speaking truth to 549 00:35:46,000 --> 00:35:48,520 Speaker 2: power when it comes to AI, and it's critical that 550 00:35:48,520 --> 00:35:52,080 Speaker 2: the people who hold that power are listening. Her new book, 551 00:35:52,239 --> 00:35:54,680 Speaker 2: A Masking AI is poised to be one of the 552 00:35:54,680 --> 00:35:57,640 Speaker 2: most important books about technology of the year, and it 553 00:35:57,680 --> 00:36:00,520 Speaker 2: could not have come at a better time. Available now, 554 00:36:00,560 --> 00:36:02,279 Speaker 2: so I'll hope that you'll join me in reading it. 555 00:36:07,080 --> 00:36:09,200 Speaker 2: If you're looking for ways to support the show, check 556 00:36:09,200 --> 00:36:11,840 Speaker 2: out our merch store at tangoty dot com slash store. 557 00:36:13,040 --> 00:36:15,080 Speaker 2: Got a story about an interesting thing in tech, or 558 00:36:15,160 --> 00:36:17,000 Speaker 2: just want to say hi, You can read us at 559 00:36:17,000 --> 00:36:19,760 Speaker 2: Hello at tangodi dot com. You can also find transcripts 560 00:36:19,760 --> 00:36:22,200 Speaker 2: for today's episode at tengody dot com. There Are No 561 00:36:22,280 --> 00:36:24,319 Speaker 2: Girls on the Internet was created by me Bridget Tood. 562 00:36:24,719 --> 00:36:27,880 Speaker 2: It's a production of iHeartRadio and Unbossed Creative edited by 563 00:36:27,960 --> 00:36:32,000 Speaker 2: Joey pat Jonathan Strickland is our executive producer. Tari Harrison 564 00:36:32,080 --> 00:36:34,840 Speaker 2: is our producer and sound engineer. Michael Almada is our 565 00:36:34,880 --> 00:36:37,960 Speaker 2: contributing producer. I'm your host, bridget Tood. If you want 566 00:36:37,960 --> 00:36:40,360 Speaker 2: to help us grow rate, and review us on Apple Podcasts. 567 00:36:41,120 --> 00:36:43,959 Speaker 2: For more podcasts from iHeartRadio, check out the iHeartRadio app, 568 00:36:44,000 --> 00:36:46,360 Speaker 2: Apple Podcasts, or wherever you get your podcasts.