1 00:00:04,440 --> 00:00:10,400 Speaker 1: Welcome to tech Stuff, a production from iHeartRadio. 2 00:00:12,400 --> 00:00:14,680 Speaker 2: Hey there, and welcome to tech Stuff. I'm your host, 3 00:00:14,760 --> 00:00:15,520 Speaker 2: John that Strickland. 4 00:00:15,520 --> 00:00:18,720 Speaker 1: I'm an executive producer with iHeart Podcasts and how the 5 00:00:18,800 --> 00:00:22,400 Speaker 1: tech are you? It's time for another classics episode. This 6 00:00:22,440 --> 00:00:27,080 Speaker 1: episode originally published on May nineteenth, twenty seventeen. It is 7 00:00:27,160 --> 00:00:34,600 Speaker 1: called the National Facial Recognition Database. Pretty, I would say, 8 00:00:34,680 --> 00:00:40,280 Speaker 1: controversial topic. Well, let's listen in now. Before I dive 9 00:00:40,320 --> 00:00:41,919 Speaker 1: into the topic, I want to make a couple of 10 00:00:41,920 --> 00:00:46,760 Speaker 1: things very clear at the very beginning. First is I'm biased. 11 00:00:47,880 --> 00:00:50,720 Speaker 1: I think the use of facial recognition software is problematic 12 00:00:50,880 --> 00:00:55,800 Speaker 1: even if you have regulations in place. But I'm mostly 13 00:00:55,840 --> 00:01:00,480 Speaker 1: talking about unregulated use because really we haven't a establish 14 00:01:00,600 --> 00:01:03,760 Speaker 1: the rules and policies to guide the use of facial 15 00:01:03,800 --> 00:01:08,199 Speaker 1: recognition software in a law enforcement context. So that's problem 16 00:01:08,280 --> 00:01:10,680 Speaker 1: Number one is I have a very strong opinion about 17 00:01:10,680 --> 00:01:13,120 Speaker 1: this and I'm not going to shy away from that. 18 00:01:15,560 --> 00:01:20,600 Speaker 1: It's really unjustifiable to have unregulated use of facial recognition 19 00:01:20,920 --> 00:01:24,679 Speaker 1: software in law enforcement contexts. So I want to make 20 00:01:24,720 --> 00:01:26,720 Speaker 1: that clear out of the gate that I have this bias, 21 00:01:27,360 --> 00:01:30,360 Speaker 1: and if that's an issue, that's fair, But at least 22 00:01:30,400 --> 00:01:32,880 Speaker 1: I'm being honest, right, I'm not presenting this as if 23 00:01:32,920 --> 00:01:39,319 Speaker 1: it's completely objective, unbiased information. I own this. You don't 24 00:01:39,360 --> 00:01:43,280 Speaker 1: have to tell me. I know it already. Next, this 25 00:01:43,400 --> 00:01:47,680 Speaker 1: is largely going to be a US centric discussion so 26 00:01:47,760 --> 00:01:51,280 Speaker 1: that I can talk about details. But please know that 27 00:01:51,320 --> 00:01:53,600 Speaker 1: there are a lot of these types of systems all 28 00:01:53,640 --> 00:01:56,600 Speaker 1: over the world, not just in the United States, and 29 00:01:56,680 --> 00:01:59,040 Speaker 1: a lot of these places have similar issues to the 30 00:01:59,080 --> 00:02:01,280 Speaker 1: ones I'm going to be talking about here in the US. 31 00:02:01,960 --> 00:02:06,320 Speaker 1: I'll just be focusing more on US stories to make 32 00:02:06,360 --> 00:02:09,720 Speaker 1: specific points because this is where I live, and now 33 00:02:09,760 --> 00:02:12,560 Speaker 1: to explain what I'm actually talking about here. So, back 34 00:02:12,600 --> 00:02:16,440 Speaker 1: in twenty ten, the FBI undertook a project that cost 35 00:02:16,639 --> 00:02:20,399 Speaker 1: more than an estimated one point two billion dollars that's 36 00:02:20,440 --> 00:02:24,160 Speaker 1: billion with a B to replace what was called the 37 00:02:24,280 --> 00:02:29,560 Speaker 1: Integrated Automated Fingerprint System or IAFS. Now, if I had 38 00:02:29,560 --> 00:02:32,720 Speaker 1: been in place since nineteen ninety nine, and I've talked 39 00:02:32,760 --> 00:02:40,200 Speaker 1: about fingerprints in a previous episode, IAFS was an attempt 40 00:02:40,400 --> 00:02:47,240 Speaker 1: to create a usye database of fingerprint records so that 41 00:02:47,600 --> 00:02:50,000 Speaker 1: if you were investigating a crime and you had lifted 42 00:02:50,040 --> 00:02:55,480 Speaker 1: some prints from the crime, you could end up consulting 43 00:02:56,120 --> 00:02:58,600 Speaker 1: this database and see if there are any matches in 44 00:02:58,639 --> 00:03:02,960 Speaker 1: place to give you any leads on your investigation. The 45 00:03:03,160 --> 00:03:06,519 Speaker 1: twenty ten project the FBI undertook was meant to vastly 46 00:03:06,680 --> 00:03:11,000 Speaker 1: expand that capability by adding a lot more data to 47 00:03:11,240 --> 00:03:14,799 Speaker 1: the database, not just fingerprints, but other stuff as well, 48 00:03:15,520 --> 00:03:19,360 Speaker 1: and the new system is called the Next Generation Identification 49 00:03:19,639 --> 00:03:24,880 Speaker 1: or NGI. It includes not just fingerprints, but other biographical 50 00:03:25,080 --> 00:03:30,200 Speaker 1: data and biometrics information, including face recognition technology. So a 51 00:03:30,200 --> 00:03:34,680 Speaker 1: lot of images are included in this particular database. So 52 00:03:34,720 --> 00:03:37,880 Speaker 1: as part of this project, the FBI incorporated the Interstate 53 00:03:38,120 --> 00:03:42,960 Speaker 1: Photo System or IPS, so you have NGI IPS it 54 00:03:43,080 --> 00:03:47,040 Speaker 1: typically is how it's referred to now. That system includes 55 00:03:47,080 --> 00:03:51,080 Speaker 1: images from police cases as well as photos from civil 56 00:03:51,240 --> 00:03:56,440 Speaker 1: civic sources that are not necessarily related to crimes. That's 57 00:03:56,480 --> 00:03:59,480 Speaker 1: not the only way the FBI can scan for a 58 00:03:59,560 --> 00:04:03,800 Speaker 1: match of a photograph they've taken that relates to a 59 00:04:03,800 --> 00:04:07,119 Speaker 1: case in some way to this massive database, but more 60 00:04:07,120 --> 00:04:10,880 Speaker 1: on that in a little bit now. The general process 61 00:04:10,920 --> 00:04:14,800 Speaker 1: of searching for a match follows a pretty simple pattern, 62 00:04:14,840 --> 00:04:18,479 Speaker 1: although the details can be vastly different depending upon what 63 00:04:18,680 --> 00:04:23,200 Speaker 1: facial recognition software you are using at the time. So 64 00:04:23,480 --> 00:04:26,680 Speaker 1: you first start with an image related to a case, 65 00:04:27,160 --> 00:04:30,520 Speaker 1: and this is called the probe photo. It is the 66 00:04:30,560 --> 00:04:34,280 Speaker 1: one you are probing for lack of a better term, 67 00:04:35,920 --> 00:04:39,919 Speaker 1: you don't know the identity of the person in the photograph, typically, 68 00:04:40,120 --> 00:04:42,400 Speaker 1: or at least you might have suspicions, but you don't 69 00:04:42,440 --> 00:04:44,839 Speaker 1: necessarily know for sure. So you've got a picture of 70 00:04:44,880 --> 00:04:50,000 Speaker 1: an unknown person in this photograph. You then scan that 71 00:04:50,040 --> 00:04:53,800 Speaker 1: photo and you use facial recognition software to analyze the 72 00:04:53,839 --> 00:04:56,800 Speaker 1: picture and to try and find a match in this 73 00:04:57,000 --> 00:04:59,880 Speaker 1: larger database. It starts searching all of the images in 74 00:04:59,880 --> 00:05:02,920 Speaker 1: the database looking for any that might be a potential match. 75 00:05:03,560 --> 00:05:06,440 Speaker 1: Depending upon the system and the policies that are in use, 76 00:05:06,880 --> 00:05:10,080 Speaker 1: you could end up with a single photo return to you. 77 00:05:10,080 --> 00:05:13,080 Speaker 1: You could end up with dozens of photos, so these 78 00:05:13,120 --> 00:05:16,520 Speaker 1: would all be potential matches with different degrees of certainty 79 00:05:16,640 --> 00:05:19,680 Speaker 1: for a match. You might remember in episodes I've talked 80 00:05:19,680 --> 00:05:22,640 Speaker 1: about things like IBM's Watson that would come up with 81 00:05:22,800 --> 00:05:26,360 Speaker 1: answers to a question and assign a value to each 82 00:05:26,400 --> 00:05:29,719 Speaker 1: potential answer, and the one that had the highest value, 83 00:05:30,680 --> 00:05:34,640 Speaker 1: assuming it's above a certain threshold, would be submitted as 84 00:05:34,720 --> 00:05:37,000 Speaker 1: the answer. So it's not so much that the computer 85 00:05:37,080 --> 00:05:40,600 Speaker 1: quote unquote knows it has a match. It suspects a 86 00:05:40,640 --> 00:05:43,960 Speaker 1: match based upon a certain percentage as long as it's 87 00:05:44,040 --> 00:05:48,320 Speaker 1: over a threshold of certainty, or you might end up 88 00:05:48,360 --> 00:05:50,960 Speaker 1: with no photos at all. If no match was found 89 00:05:51,080 --> 00:05:55,640 Speaker 1: or nothing ended up being above that threshold, the system 90 00:05:55,720 --> 00:05:58,440 Speaker 1: might say, I couldn't match this photo with anyone who's 91 00:05:58,480 --> 00:06:04,080 Speaker 1: in the database. A study performed by researchers at Georgetown 92 00:06:04,160 --> 00:06:09,840 Speaker 1: University found that one in every two American adults has 93 00:06:09,880 --> 00:06:14,560 Speaker 1: their face captured in an image database that is accessible 94 00:06:14,600 --> 00:06:19,200 Speaker 1: by various law enforcement agencies, including but not limited to 95 00:06:19,360 --> 00:06:23,080 Speaker 1: the IPS. In fact, the IPS has a small number 96 00:06:23,200 --> 00:06:27,800 Speaker 1: of photos compared to the overall number represented by databases 97 00:06:27,839 --> 00:06:33,520 Speaker 1: across the US. Now, this involves agencies at all different levels, federal, state, 98 00:06:33,640 --> 00:06:40,800 Speaker 1: even tribal law for Native American tribes. That ends up 99 00:06:40,800 --> 00:06:45,520 Speaker 1: being about one hundred and seventeen million people in these databases, 100 00:06:46,240 --> 00:06:50,160 Speaker 1: many of whom, in fact large percentage of whom have 101 00:06:50,279 --> 00:06:54,200 Speaker 1: no criminal background whatsoever. Their images are also in these databases, 102 00:06:54,600 --> 00:06:58,679 Speaker 1: and this raises some big concerns about privacy and also accountability. 103 00:06:59,000 --> 00:07:01,880 Speaker 1: So in today's episode, we're going to explore how facial 104 00:07:02,040 --> 00:07:07,480 Speaker 1: recognition software works, as well as talk about the implementation 105 00:07:08,120 --> 00:07:12,000 Speaker 1: for law enforcement and the reaction to this technology, and 106 00:07:12,040 --> 00:07:14,840 Speaker 1: will probably listen to me get upset and a little 107 00:07:14,840 --> 00:07:18,400 Speaker 1: head up about the whole thing in general. All right, 108 00:07:18,760 --> 00:07:22,200 Speaker 1: So first, before we leap into the mess of law enforcement, 109 00:07:22,400 --> 00:07:26,640 Speaker 1: because it is a mess, that's just a fact, let's 110 00:07:26,680 --> 00:07:30,880 Speaker 1: talk first about the technology itself. When did facial recognition 111 00:07:31,000 --> 00:07:34,720 Speaker 1: software get started and how does it work? Well, it's 112 00:07:34,760 --> 00:07:38,600 Speaker 1: related to computer vision, which is a subset of artificial 113 00:07:38,640 --> 00:07:42,280 Speaker 1: intelligence research. If you look at artificial intelligence, a lot 114 00:07:42,320 --> 00:07:45,240 Speaker 1: of people simplify that by meaning, oh, this is so 115 00:07:45,320 --> 00:07:47,840 Speaker 1: that you can teach computers how to think like people. 116 00:07:48,320 --> 00:07:51,520 Speaker 1: But that's actually a very specific definition of a very 117 00:07:51,560 --> 00:07:54,840 Speaker 1: specific type of artificial intelligence. When you really look at 118 00:07:54,840 --> 00:07:57,960 Speaker 1: AI and you break it out, it involves a lot 119 00:07:57,960 --> 00:08:01,440 Speaker 1: of subsets of abilities. One of those is the ability 120 00:08:01,440 --> 00:08:07,240 Speaker 1: for machines to analyze imagery and be able to determine 121 00:08:07,280 --> 00:08:10,600 Speaker 1: what that imagery represents. In a way, you could argue 122 00:08:10,640 --> 00:08:16,680 Speaker 1: it's teaching computers how to understand pictures. It's also really challenging, 123 00:08:17,280 --> 00:08:20,360 Speaker 1: and this is one of the object lessons that I 124 00:08:20,520 --> 00:08:25,320 Speaker 1: use to teach people how Artificial intelligence is really tricky. 125 00:08:25,360 --> 00:08:28,920 Speaker 1: It requires more than just pure processing power. I mean, 126 00:08:28,960 --> 00:08:32,720 Speaker 1: processing power is important, but you can't solve all of 127 00:08:32,800 --> 00:08:36,600 Speaker 1: AI's problems just by throwing more processors at it. You 128 00:08:36,679 --> 00:08:39,720 Speaker 1: have to figure out from a software level how to 129 00:08:39,880 --> 00:08:43,599 Speaker 1: leverage that processing power in a way that gives computers 130 00:08:43,640 --> 00:08:48,800 Speaker 1: this ability to identify stuff based upon imagery. So a 131 00:08:48,840 --> 00:08:52,199 Speaker 1: computer might be able to perform far more mathematical operations 132 00:08:52,240 --> 00:08:55,640 Speaker 1: per second than even the cleverest of humans, but without 133 00:08:55,679 --> 00:08:58,040 Speaker 1: the right software, they can't identify the picture of a 134 00:08:58,080 --> 00:09:01,960 Speaker 1: seagull compared to say, a semi truck. You have to 135 00:09:02,040 --> 00:09:05,600 Speaker 1: teach the computer how to do this. So let's say 136 00:09:05,640 --> 00:09:08,360 Speaker 1: you develop a program that can analyze an image and 137 00:09:08,440 --> 00:09:13,720 Speaker 1: break it down into simple data to describe that image, 138 00:09:13,960 --> 00:09:17,520 Speaker 1: and then you essentially teach a computer what a coffee 139 00:09:17,559 --> 00:09:19,959 Speaker 1: mug looks like. You take a picture of a coffee mug, 140 00:09:20,600 --> 00:09:23,960 Speaker 1: you feed it to a computer, and you essentially say 141 00:09:24,280 --> 00:09:30,960 Speaker 1: this data represents a coffee mug. You then would have 142 00:09:31,120 --> 00:09:36,200 Speaker 1: to try and train the computer on what that actually means. 143 00:09:36,440 --> 00:09:39,640 Speaker 1: The computer does not now know what a coffee mug is. 144 00:09:40,600 --> 00:09:44,560 Speaker 1: It will recognize that specific mug in that specific orientation 145 00:09:44,840 --> 00:09:48,640 Speaker 1: under those specific lighting conditions, assuming that you've designed the 146 00:09:48,640 --> 00:09:54,000 Speaker 1: algorithm properly. But it's way more tricky than that. What 147 00:09:54,080 --> 00:09:56,680 Speaker 1: if in the image that you fed the computer, the 148 00:09:56,720 --> 00:10:00,840 Speaker 1: coffee mugs handle was facing to the left with respect 149 00:10:00,840 --> 00:10:04,040 Speaker 1: of the viewer, but in a future picture the handle 150 00:10:04,160 --> 00:10:06,120 Speaker 1: is off to the right instead of to the left, 151 00:10:06,200 --> 00:10:08,880 Speaker 1: or it's turned around so you can't see the handle 152 00:10:08,920 --> 00:10:11,480 Speaker 1: at all. It's behind the coffee mug. Well, if the 153 00:10:11,559 --> 00:10:14,600 Speaker 1: mug is bigger or smaller, or a different shape, well 154 00:10:14,600 --> 00:10:18,200 Speaker 1: if it's a different color. Image recognition is tough because 155 00:10:18,240 --> 00:10:23,280 Speaker 1: computers don't immediately associate different objects within the same category 156 00:10:23,960 --> 00:10:28,400 Speaker 1: as being the same thing. So if you teach me, Jonathan, 157 00:10:28,920 --> 00:10:31,240 Speaker 1: what a coffee mug is, and you show me a 158 00:10:31,280 --> 00:10:34,840 Speaker 1: couple of different examples saying, this is a coffee mug, 159 00:10:34,880 --> 00:10:37,080 Speaker 1: but this is also a coffee mug, even though it's 160 00:10:37,080 --> 00:10:39,640 Speaker 1: a different size and different shape and a different color, 161 00:10:40,320 --> 00:10:42,600 Speaker 1: I'll catch on pretty quickly and it won't take very 162 00:10:42,600 --> 00:10:46,000 Speaker 1: many coffee mugs for me to figure out. All Right, 163 00:10:46,040 --> 00:10:48,920 Speaker 1: I got the basic idea of what a coffee mug is. 164 00:10:49,040 --> 00:10:52,280 Speaker 1: I know what the concept of coffee mug is now, 165 00:10:53,000 --> 00:10:56,800 Speaker 1: But computers aren't like that. You have to feed them 166 00:10:57,000 --> 00:11:00,600 Speaker 1: thousands of images, both of coffee mugs and of not 167 00:11:01,000 --> 00:11:04,199 Speaker 1: coffee mugs, so that the computer starts to be able 168 00:11:04,240 --> 00:11:08,920 Speaker 1: to pick out the various features that are the essence 169 00:11:09,120 --> 00:11:12,520 Speaker 1: of a coffee mug versus things that are not related 170 00:11:12,600 --> 00:11:16,400 Speaker 1: to being a coffee mug. It takes hours and hours 171 00:11:16,400 --> 00:11:18,960 Speaker 1: and hours of work of training these computers to do it, 172 00:11:19,000 --> 00:11:22,520 Speaker 1: so it's a non trivial task, and this is true 173 00:11:22,559 --> 00:11:28,600 Speaker 1: of all types of image recognition, including facial recognition. Now, 174 00:11:28,600 --> 00:11:34,480 Speaker 1: to get around that problem, you end up sending thousands, 175 00:11:34,559 --> 00:11:38,080 Speaker 1: countless thousands, millions maybe of images of what you're interested 176 00:11:38,120 --> 00:11:40,480 Speaker 1: in while you're training the computer. And the nice thing 177 00:11:40,559 --> 00:11:43,920 Speaker 1: is computers can process this information very very quickly, so 178 00:11:44,040 --> 00:11:49,520 Speaker 1: while it takes a lot, it doesn't take relatively that long, 179 00:11:50,040 --> 00:11:52,959 Speaker 1: it's not as laborious a process as it could be 180 00:11:53,120 --> 00:11:58,319 Speaker 1: if computers were slower at analyzing information. So you might 181 00:11:58,360 --> 00:12:01,560 Speaker 1: remember a story that kind of illustrates the point. Back 182 00:12:01,600 --> 00:12:05,520 Speaker 1: in twenty twelve, there was a network of sixteen thousand 183 00:12:05,520 --> 00:12:11,080 Speaker 1: computers that analyzed ten million images, and as a result, 184 00:12:11,080 --> 00:12:13,880 Speaker 1: it could do the most important task any computer connected 185 00:12:13,880 --> 00:12:16,840 Speaker 1: to the Internet should be expected to do. It could 186 00:12:16,840 --> 00:12:20,440 Speaker 1: then identify cat videos because it now knew what a 187 00:12:20,440 --> 00:12:24,200 Speaker 1: cat was, or at least the features that define catness. 188 00:12:25,080 --> 00:12:27,600 Speaker 1: Catness as in the essence of being a cat, not 189 00:12:27,720 --> 00:12:31,520 Speaker 1: a character from Hunger Games. Even then, there were times 190 00:12:31,520 --> 00:12:33,600 Speaker 1: when a computer would get it wrong. Either it would 191 00:12:33,640 --> 00:12:35,880 Speaker 1: not identify a cat as being a cat, or it 192 00:12:35,920 --> 00:12:38,880 Speaker 1: would misidentify something else as being a cat because its 193 00:12:38,920 --> 00:12:41,360 Speaker 1: features were close enough to cat like for it to 194 00:12:41,520 --> 00:12:46,360 Speaker 1: fool the computer algorithm. A major breakthrough in facial recognition 195 00:12:46,400 --> 00:12:48,800 Speaker 1: algorithms happened way back in two thousand and one. That's 196 00:12:48,840 --> 00:12:52,360 Speaker 1: when Paul Viola and Michael Jones unveiled an algorithm for 197 00:12:52,440 --> 00:12:56,080 Speaker 1: face detection, and it worked in real time, which meant 198 00:12:56,160 --> 00:12:59,600 Speaker 1: that it could recognize a face that it would appear 199 00:12:59,679 --> 00:13:02,839 Speaker 1: on a webcam. And by recognized, I mean it recognized 200 00:13:02,880 --> 00:13:06,840 Speaker 1: that it was a face. It didn't assign an identity 201 00:13:07,360 --> 00:13:11,480 Speaker 1: to the face. It didn't say, Oh, that's Bob, It said, oh, 202 00:13:11,679 --> 00:13:13,600 Speaker 1: that is a face that is in front of the 203 00:13:13,600 --> 00:13:19,000 Speaker 1: webcam right now. The algorithm soon found its way into 204 00:13:19,120 --> 00:13:24,040 Speaker 1: open CV, which is an open source computer vision framework, 205 00:13:24,559 --> 00:13:28,079 Speaker 1: and the open source approach allowed other programmers to dive 206 00:13:28,120 --> 00:13:31,000 Speaker 1: into that code and to make changes and improvements, and 207 00:13:31,040 --> 00:13:36,360 Speaker 1: it helped a rapid prototyping of facial recognition software to 208 00:13:36,520 --> 00:13:40,000 Speaker 1: Other computer scientists who helped advance computer vision further were 209 00:13:40,080 --> 00:13:44,160 Speaker 1: Bill Triggs and Navnit de Lal, who published a paper 210 00:13:44,160 --> 00:13:48,439 Speaker 1: in two thousand and five about the histograbs of oriented gradients. Now, 211 00:13:48,480 --> 00:13:51,360 Speaker 1: that was an approach that looked at gradient orientation in 212 00:13:51,400 --> 00:13:53,800 Speaker 1: parts of an image, and essentially it describes the process 213 00:13:53,840 --> 00:13:56,960 Speaker 1: of viewing an image with attention to edge directions and 214 00:13:57,000 --> 00:14:01,200 Speaker 1: intensity gradients. That's a complicated way of saying the technique 215 00:14:01,240 --> 00:14:04,320 Speaker 1: looks at the totality of a person, and then a 216 00:14:04,400 --> 00:14:07,640 Speaker 1: machine learning algorithm determines whether or not that is actually 217 00:14:07,679 --> 00:14:11,600 Speaker 1: a person or not a person. A bit later, computer 218 00:14:11,640 --> 00:14:15,520 Speaker 1: scientists began pairing computer vision algorithms with deep learning and 219 00:14:15,679 --> 00:14:21,800 Speaker 1: convolutional neural networks or CNNs. To go into this would 220 00:14:21,840 --> 00:14:25,040 Speaker 1: require an episode all by itself. Neural networks are fascinating, 221 00:14:25,080 --> 00:14:28,160 Speaker 1: but they're also pretty complicated, and I've got a whole 222 00:14:28,240 --> 00:14:31,400 Speaker 1: lot of topics to cover today, so we can't really 223 00:14:31,440 --> 00:14:34,160 Speaker 1: dive into it. You can think of an artificial neural 224 00:14:34,200 --> 00:14:38,600 Speaker 1: network as designing a computer system that processes information in 225 00:14:38,640 --> 00:14:41,240 Speaker 1: a way that's similar to the way our brains do. 226 00:14:41,720 --> 00:14:44,440 Speaker 1: The computers are not thinking, but they are able to 227 00:14:44,480 --> 00:14:50,040 Speaker 1: process information in a way that mimics how we process information, 228 00:14:50,520 --> 00:14:55,040 Speaker 1: or a semi close approximation thereof that's a really kind 229 00:14:55,080 --> 00:14:57,000 Speaker 1: of weak way of describing it. But again, to really 230 00:14:57,000 --> 00:15:03,360 Speaker 1: go into detail will require a full episode all by itself. Typically, 231 00:15:03,640 --> 00:15:08,200 Speaker 1: facial recognition software uses feature extraction to look for patterns 232 00:15:08,200 --> 00:15:11,680 Speaker 1: in an image relating to facial features. In other words, 233 00:15:11,680 --> 00:15:15,440 Speaker 1: it searches for features that resemble a face, the elements 234 00:15:15,440 --> 00:15:19,280 Speaker 1: you would expect to be present in a typical face, 235 00:15:19,680 --> 00:15:24,280 Speaker 1: So eyes, nose, a mouth, that would be major ones. Right. 236 00:15:24,560 --> 00:15:27,800 Speaker 1: Then the software starts to estimate the relationships between those 237 00:15:27,840 --> 00:15:32,120 Speaker 1: different elements. How wide are the eyes, how far apart 238 00:15:32,160 --> 00:15:34,040 Speaker 1: are they from each other, How wide is the nose, 239 00:15:34,680 --> 00:15:37,640 Speaker 1: how long is the jawline, what shape are the cheekbones? 240 00:15:39,000 --> 00:15:43,000 Speaker 1: These sort of elements all play a part as points 241 00:15:43,000 --> 00:15:48,240 Speaker 1: of data, and different facial recognition software packages weight these 242 00:15:48,280 --> 00:15:52,600 Speaker 1: features in a different way. So it's not like I 243 00:15:52,640 --> 00:15:56,040 Speaker 1: could say all facial recognition software looks at these four 244 00:15:56,080 --> 00:15:59,880 Speaker 1: points of data as its primary source. It varies depending 245 00:15:59,920 --> 00:16:02,680 Speaker 1: on upon the algorithm that's been designed by various companies, 246 00:16:03,480 --> 00:16:04,880 Speaker 1: and part of the problem that we're going to talk 247 00:16:04,880 --> 00:16:09,200 Speaker 1: about is that law enforcement across the United States they 248 00:16:09,200 --> 00:16:13,360 Speaker 1: are not relying on a single facial recognition software approach. 249 00:16:13,640 --> 00:16:17,400 Speaker 1: Different agencies have different vendors that they work with, So 250 00:16:18,640 --> 00:16:22,080 Speaker 1: just because one might work very well doesn't necessarily mean 251 00:16:22,120 --> 00:16:25,400 Speaker 1: it's competitors work just as well. And that's part of 252 00:16:25,400 --> 00:16:28,400 Speaker 1: the problem. Now, all of these little points of data 253 00:16:28,400 --> 00:16:32,120 Speaker 1: I'm talking about, these notle points and how they relate 254 00:16:32,160 --> 00:16:35,640 Speaker 1: to one another, all of that gets boiled down into 255 00:16:35,720 --> 00:16:39,160 Speaker 1: a numeric code that you could think of as a 256 00:16:39,200 --> 00:16:42,280 Speaker 1: face print. This is supposed to be a representation of 257 00:16:42,320 --> 00:16:47,800 Speaker 1: the unique set of data that is a compilation of 258 00:16:47,880 --> 00:16:53,480 Speaker 1: all of these different points boiled down into numeric information itself. 259 00:16:56,080 --> 00:16:57,400 Speaker 1: Then what you would do is you would have a 260 00:16:57,480 --> 00:17:02,520 Speaker 1: database of face So if you wanted to find a match, 261 00:17:02,920 --> 00:17:05,920 Speaker 1: you would feed the image you have, the probe image 262 00:17:06,080 --> 00:17:09,919 Speaker 1: into this database, and the facial recognition software would analyze 263 00:17:09,960 --> 00:17:13,240 Speaker 1: the probe photo. It would end up assigning this numeric 264 00:17:13,400 --> 00:17:16,480 Speaker 1: value and would start looking through the database for other 265 00:17:16,560 --> 00:17:20,040 Speaker 1: numeric values that were as similar to that probe one 266 00:17:20,160 --> 00:17:25,800 Speaker 1: as possible and start returning those images as potential matches 267 00:17:26,119 --> 00:17:29,560 Speaker 1: or candidates. They tend to use the word candidate photos. 268 00:17:30,640 --> 00:17:33,199 Speaker 1: Otherwise you'll either get no match at all or you 269 00:17:33,240 --> 00:17:35,800 Speaker 1: get a false positive. You will end up getting an 270 00:17:35,800 --> 00:17:39,919 Speaker 1: image of someone who looks like the person whose image 271 00:17:39,920 --> 00:17:44,480 Speaker 1: you submitted, but is not the same person. That does happen, 272 00:17:45,240 --> 00:17:48,200 Speaker 1: And that's the basic way that facial recognition software works. 273 00:17:49,119 --> 00:17:51,439 Speaker 1: But keep in mind, different vendors use all their own 274 00:17:51,480 --> 00:17:54,560 Speaker 1: specific approaches, like I said, and some could be less 275 00:17:54,600 --> 00:17:58,840 Speaker 1: accurate than others. Some might be accurate for specific ethnicities 276 00:17:58,880 --> 00:18:01,600 Speaker 1: and not as accurate as other ones. That's a huge problem, 277 00:18:03,240 --> 00:18:08,000 Speaker 1: so it gets complicated. Even when I'm talking in more 278 00:18:08,040 --> 00:18:10,960 Speaker 1: general terms, you have to remember that there are a 279 00:18:10,960 --> 00:18:18,359 Speaker 1: lot of specific incidents and specific implementations of facial recognition 280 00:18:18,480 --> 00:18:23,080 Speaker 1: software that have their own issues. So I'm gonna be 281 00:18:23,160 --> 00:18:24,920 Speaker 1: as general as I can. I'm not going to call 282 00:18:24,960 --> 00:18:29,280 Speaker 1: out any particular facial recognition software vendors out there. I'm 283 00:18:29,280 --> 00:18:32,879 Speaker 1: more going to talk about the overall issues that various 284 00:18:32,960 --> 00:18:38,040 Speaker 1: organizations have had as they've looked into this topic. Now, 285 00:18:38,080 --> 00:18:40,879 Speaker 1: there are plenty of applications for facial recognition that have 286 00:18:40,920 --> 00:18:43,119 Speaker 1: nothing to do with identifying a person. I mentioned that 287 00:18:43,200 --> 00:18:45,840 Speaker 1: earlier that there was the one for a webcam that 288 00:18:45,880 --> 00:18:48,320 Speaker 1: could identify when a face was in front of the webcam. 289 00:18:48,440 --> 00:18:51,720 Speaker 1: This wasn't to identify anybody. It was again just to say, yes, 290 00:18:51,760 --> 00:18:55,080 Speaker 1: there's somebody looking into the webcam at this moment, which 291 00:18:55,119 --> 00:18:57,560 Speaker 1: by itself can be useful and have nothing to do 292 00:18:57,600 --> 00:19:01,400 Speaker 1: with identification. There are plenty of digital cameras out there 293 00:19:01,480 --> 00:19:06,199 Speaker 1: and camera phone apps that can identify when there's a 294 00:19:06,280 --> 00:19:10,000 Speaker 1: face looking at the camera, and again it's not necessarily 295 00:19:10,000 --> 00:19:12,440 Speaker 1: to identify that person, but rather to say, oh, well, 296 00:19:12,720 --> 00:19:15,880 Speaker 1: this is a face. The camera is most likely trying 297 00:19:15,880 --> 00:19:18,520 Speaker 1: to focus on this person, so let's make this person 298 00:19:18,560 --> 00:19:21,560 Speaker 1: the point of focus and not focus on something in 299 00:19:21,600 --> 00:19:25,119 Speaker 1: the background like a tree that's fifty yards back. Instead, 300 00:19:25,160 --> 00:19:28,240 Speaker 1: let's focus on the person who's in the foreground. So 301 00:19:28,280 --> 00:19:33,560 Speaker 1: that's pretty handy, and again there's nothing particularly problematic from 302 00:19:33,600 --> 00:19:36,320 Speaker 1: an identification standpoint, because that's not the purpose of it. 303 00:19:38,119 --> 00:19:41,919 Speaker 1: But then you also have other implementations, like on social media, 304 00:19:42,200 --> 00:19:45,240 Speaker 1: which allow you to do things like tag people based 305 00:19:45,320 --> 00:19:50,080 Speaker 1: upon an algorithm recognizing a person. So Facebook is a 306 00:19:50,080 --> 00:19:52,560 Speaker 1: great example of this. Right, if you upload a picture 307 00:19:52,560 --> 00:19:55,960 Speaker 1: of one of your Facebook friends onto Facebook chances are 308 00:19:56,000 --> 00:19:59,080 Speaker 1: it's giving you a suggestion to tag that photo with 309 00:19:59,280 --> 00:20:04,600 Speaker 1: the specific in mind. That may not be that problematic either, 310 00:20:05,200 --> 00:20:08,640 Speaker 1: depending upon how your friend feels about pictures being uploaded 311 00:20:08,640 --> 00:20:13,959 Speaker 1: to Facebook. Some people are very cautious about that, and 312 00:20:14,240 --> 00:20:16,399 Speaker 1: of course you know, I always recommend you talk to 313 00:20:16,440 --> 00:20:20,200 Speaker 1: anybody before you start tagging folks on Facebook photos, just 314 00:20:20,240 --> 00:20:22,679 Speaker 1: to make sure they're fine with it. I say that 315 00:20:22,720 --> 00:20:25,320 Speaker 1: as a person who has done it, and then notice 316 00:20:25,359 --> 00:20:27,440 Speaker 1: that some of my tags got removed by the people 317 00:20:27,480 --> 00:20:30,920 Speaker 1: I tagged later on, which taught me I should probably 318 00:20:31,000 --> 00:20:35,760 Speaker 1: ask first, rather than give them the feeling that they 319 00:20:35,760 --> 00:20:39,080 Speaker 1: need to go and remove a tag or two. We've 320 00:20:39,080 --> 00:20:43,120 Speaker 1: also seen examples of this simple implementation of facial recognition 321 00:20:43,200 --> 00:20:49,080 Speaker 1: going awry. Google's street View will blur out faces, for example, 322 00:20:49,560 --> 00:20:53,520 Speaker 1: in an effort to protect people's identity while street view 323 00:20:53,520 --> 00:20:56,399 Speaker 1: cars are out and about taking images. This makes sense. 324 00:20:56,680 --> 00:20:58,920 Speaker 1: Let's say that you are in a part of town 325 00:20:59,280 --> 00:21:01,840 Speaker 1: that you normally would not be in. For whatever reason, 326 00:21:02,040 --> 00:21:05,080 Speaker 1: you might not want your picture to be included on 327 00:21:05,119 --> 00:21:07,800 Speaker 1: Google street View, so that whenever anyone looks at that 328 00:21:07,880 --> 00:21:11,200 Speaker 1: street for that point forward, they see your face on there, 329 00:21:11,800 --> 00:21:15,639 Speaker 1: you know, coming out of I don't know a Wendy's. 330 00:21:15,880 --> 00:21:19,320 Speaker 1: Maybe you are a manager for burger King that would 331 00:21:19,320 --> 00:21:23,359 Speaker 1: look bad, or you know, lots of other reasons that 332 00:21:23,520 --> 00:21:26,600 Speaker 1: obviously can spring to mind as well. You don't want 333 00:21:26,640 --> 00:21:32,200 Speaker 1: to violate someone's privacy. But Google StreetView would also blur 334 00:21:32,320 --> 00:21:35,400 Speaker 1: out images that were not real people faces, like images 335 00:21:35,440 --> 00:21:38,440 Speaker 1: on billboards or murals. Sometimes if it had a person's 336 00:21:38,480 --> 00:21:40,920 Speaker 1: face on a mural, the face would be blurred out, 337 00:21:40,920 --> 00:21:42,960 Speaker 1: even though it's not a real person, it's just a 338 00:21:43,000 --> 00:21:47,320 Speaker 1: painting or In September twenty sixteen, c Neet reported on 339 00:21:47,359 --> 00:21:49,760 Speaker 1: an incident in which Google street View blurred out the 340 00:21:49,760 --> 00:21:53,840 Speaker 1: face of a cow. So Google was being very thoughtful 341 00:21:53,960 --> 00:22:00,439 Speaker 1: to protect that cow's privacy. But what about matching faces 342 00:22:00,440 --> 00:22:04,679 Speaker 1: to identities? So in some cases, again seemingly harmless if 343 00:22:04,720 --> 00:22:07,280 Speaker 1: you want to tag your friends, but when it comes 344 00:22:07,320 --> 00:22:10,600 Speaker 1: to law enforcement, things get a bit sticky, particularly as 345 00:22:10,640 --> 00:22:13,160 Speaker 1: you learn more about the specifics. And we'll talk about 346 00:22:13,160 --> 00:22:16,240 Speaker 1: that in just a second, but first let's take a 347 00:22:16,320 --> 00:22:28,119 Speaker 1: quick break to thank our sponsor. All right, let's first 348 00:22:28,160 --> 00:22:33,720 Speaker 1: start with the FBI's Interstate Photos System, or IPS, because 349 00:22:33,800 --> 00:22:37,399 Speaker 1: this one has perhaps the least controversial elements to it 350 00:22:37,440 --> 00:22:40,679 Speaker 1: when you really look at it, it's still problematic, but 351 00:22:40,880 --> 00:22:45,560 Speaker 1: not nearly as much as the larger picture. The system 352 00:22:45,600 --> 00:22:51,880 Speaker 1: contains both images from criminal cases like mugshots and things 353 00:22:51,960 --> 00:22:54,920 Speaker 1: of that nature, but it also includes some photos from 354 00:22:55,080 --> 00:23:00,880 Speaker 1: civil sources like ID applications, that kind of thing. When 355 00:23:00,920 --> 00:23:04,760 Speaker 1: the Government Accountability Office or GAO, they're gonna be a 356 00:23:04,760 --> 00:23:08,720 Speaker 1: lot of acronyms and initializations or initialisms, I should say 357 00:23:08,800 --> 00:23:11,520 Speaker 1: in this episode, so I apologize for that. But Government 358 00:23:11,600 --> 00:23:16,480 Speaker 1: Accountability Office they did a study on this matter just 359 00:23:16,920 --> 00:23:20,680 Speaker 1: in twenty sixteen, so not that long ago. They published 360 00:23:20,680 --> 00:23:24,480 Speaker 1: its report on facial recognition software use among law enforcements, 361 00:23:24,520 --> 00:23:28,919 Speaker 1: specifically the FBI because they're a federal agency, so they 362 00:23:28,960 --> 00:23:33,719 Speaker 1: were concerned with the federal use of this. The database 363 00:23:34,040 --> 00:23:37,240 Speaker 1: contained about thirty million photos at the time of the 364 00:23:37,359 --> 00:23:41,680 Speaker 1: GAO study, so thirty million pictures are in this database. 365 00:23:42,119 --> 00:23:46,440 Speaker 1: Most of those images came from eighteen thousand different law 366 00:23:46,520 --> 00:23:50,720 Speaker 1: enforcement agencies at all levels of government, that includes the 367 00:23:50,720 --> 00:23:55,840 Speaker 1: tribal law enforcement offices. About seventy percent of all the 368 00:23:55,840 --> 00:24:01,119 Speaker 1: photos in the database were mugshots. More than of the 369 00:24:01,119 --> 00:24:06,080 Speaker 1: photos in that database are from criminal cases, so that 370 00:24:06,119 --> 00:24:10,040 Speaker 1: means that less than twenty percent were from civil sources. 371 00:24:10,920 --> 00:24:15,120 Speaker 1: In addition to that, there were some cases, plenty of them, 372 00:24:15,480 --> 00:24:19,760 Speaker 1: where the database had images of people both from a 373 00:24:19,760 --> 00:24:23,520 Speaker 1: civil source and from a criminal source. So I'll give 374 00:24:23,560 --> 00:24:27,320 Speaker 1: you a theoretical example. Let's say that sometime in the 375 00:24:27,320 --> 00:24:33,840 Speaker 1: past I got nabbed by the cops for grand theft 376 00:24:33,840 --> 00:24:37,840 Speaker 1: auto because I play that game. But let's say that 377 00:24:37,880 --> 00:24:40,080 Speaker 1: I stole a car, which we already know is a 378 00:24:40,160 --> 00:24:43,920 Speaker 1: complete fabrication because I don't even drive. But let's say 379 00:24:43,920 --> 00:24:47,199 Speaker 1: I stole a car, and that I had moved the 380 00:24:47,200 --> 00:24:51,440 Speaker 1: car across state lines. It became a federal case. Therefore, 381 00:24:51,960 --> 00:24:55,920 Speaker 1: my criminal information is included. My mugshot would be included 382 00:24:56,160 --> 00:25:03,879 Speaker 1: in this particular database. On related note, my ID also 383 00:25:04,720 --> 00:25:08,639 Speaker 1: is in that database as a civil image, not as 384 00:25:08,640 --> 00:25:11,800 Speaker 1: a criminal image. Well, in my case, they would tie 385 00:25:11,840 --> 00:25:15,920 Speaker 1: those two images together because they refer to the same 386 00:25:16,000 --> 00:25:19,200 Speaker 1: person and I had been involved in a criminal act. 387 00:25:20,119 --> 00:25:23,080 Speaker 1: So while I would have an image in there from 388 00:25:23,119 --> 00:25:26,440 Speaker 1: a civil source, it would be filed under the criminal 389 00:25:26,480 --> 00:25:28,720 Speaker 1: side of things. This is important when we get to 390 00:25:28,960 --> 00:25:32,920 Speaker 1: how the probes work. Now, let's say you have been 391 00:25:33,400 --> 00:25:37,840 Speaker 1: perfectly law abiding this whole time, and that your ID 392 00:25:39,000 --> 00:25:41,840 Speaker 1: is also in this database, but it's just under the 393 00:25:41,840 --> 00:25:45,200 Speaker 1: civil side of things. Since you don't have any criminal background, 394 00:25:45,840 --> 00:25:49,600 Speaker 1: it's not connected to anything on the criminal side, So 395 00:25:49,760 --> 00:25:54,160 Speaker 1: when it comes to probes using the IPS, your information 396 00:25:54,680 --> 00:25:59,520 Speaker 1: will not be referenced because the FBI policy is when 397 00:25:59,520 --> 00:26:03,720 Speaker 1: it's running these potential matches with a photo that's been 398 00:26:03,800 --> 00:26:06,800 Speaker 1: gathered as part of the evidence for an ongoing investigation, 399 00:26:07,520 --> 00:26:11,639 Speaker 1: they can only consult the criminal side, not the civil side, 400 00:26:12,280 --> 00:26:15,960 Speaker 1: with the exception of any civil photos that are connected 401 00:26:16,000 --> 00:26:20,280 Speaker 1: to a criminal case, as in my example, those are 402 00:26:20,359 --> 00:26:23,399 Speaker 1: fair game. So it might run a match and it 403 00:26:23,480 --> 00:26:27,520 Speaker 1: turns out that my photo for my state given identification 404 00:26:27,680 --> 00:26:31,840 Speaker 1: card is a better match than the mugshot is. That's 405 00:26:31,880 --> 00:26:34,520 Speaker 1: going to be fine because those two things were both 406 00:26:34,560 --> 00:26:37,119 Speaker 1: attached to a criminal file in the first place. But 407 00:26:37,560 --> 00:26:39,800 Speaker 1: let's say that it would have matched up against you 408 00:26:40,600 --> 00:26:43,159 Speaker 1: since you didn't have a criminal background, and since the 409 00:26:43,200 --> 00:26:46,760 Speaker 1: only record in there was a civil source, the match 410 00:26:46,760 --> 00:26:50,080 Speaker 1: would completely skip over you. It wouldn't return your picture 411 00:26:50,680 --> 00:26:54,600 Speaker 1: because your image is off limits in that particular use 412 00:26:56,359 --> 00:27:00,520 Speaker 1: very important because it's an effort to try and make 413 00:27:00,600 --> 00:27:07,600 Speaker 1: sure this facial recognition technology is focusing just on the 414 00:27:07,640 --> 00:27:13,560 Speaker 1: criminal side, not putting law abiding citizens in danger of 415 00:27:13,600 --> 00:27:19,320 Speaker 1: being pulled up in a virtual lineup, at least not 416 00:27:19,440 --> 00:27:23,479 Speaker 1: using that approach. That's the problem is that that's not the 417 00:27:23,480 --> 00:27:25,920 Speaker 1: only way the FBI runs searches. In fact, that might 418 00:27:25,920 --> 00:27:29,119 Speaker 1: not be the primary way the FBI runs searches when 419 00:27:29,119 --> 00:27:32,600 Speaker 1: they're looking for a match to a photo that was 420 00:27:32,600 --> 00:27:36,840 Speaker 1: taken as part of evidence gathering in pursuing a case. 421 00:27:40,320 --> 00:27:42,959 Speaker 1: But let's say that you are an FBI agent and 422 00:27:43,000 --> 00:27:45,720 Speaker 1: you've got a photo, a probe photo, and you want 423 00:27:45,760 --> 00:27:49,240 Speaker 1: to run it for a match. What's the procedure. You 424 00:27:49,240 --> 00:27:53,879 Speaker 1: would send off your request to the NGI dash Ips Department, 425 00:27:54,440 --> 00:27:57,840 Speaker 1: and you would have to indicate how many potential photographs 426 00:27:57,880 --> 00:28:02,600 Speaker 1: you want back, how many candidates do you want. You 427 00:28:02,600 --> 00:28:08,080 Speaker 1: can choose between two candidate photos and fifty candidate photos. 428 00:28:08,359 --> 00:28:10,879 Speaker 1: These are photos of different individuals, by the way, not 429 00:28:11,040 --> 00:28:14,040 Speaker 1: just here's a picture of Jonathan on the beach. Here's 430 00:28:14,040 --> 00:28:17,520 Speaker 1: a picture of Jonathan in the woods. No, it's more like, 431 00:28:17,600 --> 00:28:19,639 Speaker 1: here's a picture of Jonathan. Here's a picture of a 432 00:28:19,640 --> 00:28:22,200 Speaker 1: person who's not Jonathan, but also kind of matches this 433 00:28:22,400 --> 00:28:28,119 Speaker 1: particular probe photo you submitted. And here are forty eight others. 434 00:28:28,400 --> 00:28:31,359 Speaker 1: The default is twenty, so if you don't change the 435 00:28:31,400 --> 00:28:35,320 Speaker 1: default at all, you will get back twenty images that 436 00:28:35,440 --> 00:28:39,800 Speaker 1: are potential candidates matching your probe photo, assuming that any 437 00:28:40,280 --> 00:28:43,440 Speaker 1: are found at all. It is possible that you submit 438 00:28:43,560 --> 00:28:46,160 Speaker 1: a probe photo and the system doesn't find any matches 439 00:28:46,160 --> 00:28:48,080 Speaker 1: at all, and which case you'll just get a null. 440 00:28:49,400 --> 00:28:52,960 Speaker 1: You might get less than what you asked for if 441 00:28:54,000 --> 00:28:58,520 Speaker 1: only a few had met the threshold for reliability. Now 442 00:28:58,640 --> 00:29:04,440 Speaker 1: we call them candidate photos because you're supposed to acknowledge 443 00:29:04,440 --> 00:29:07,960 Speaker 1: the fact that these are meant to help you pursue 444 00:29:07,960 --> 00:29:11,080 Speaker 1: a lead of inquiry. In a case, it is not 445 00:29:11,280 --> 00:29:17,960 Speaker 1: meant to be a source of positive identification of a suspect. 446 00:29:18,400 --> 00:29:21,200 Speaker 1: So in other words, you shouldn't run a facial recognition 447 00:29:21,680 --> 00:29:25,600 Speaker 1: software probe, get a result back and say that's our guy, 448 00:29:25,960 --> 00:29:29,320 Speaker 1: let's go pick him up. That's not enough. It's meant 449 00:29:29,320 --> 00:29:33,760 Speaker 1: to be the start of a line of inquiry, and 450 00:29:35,000 --> 00:29:36,640 Speaker 1: whether or not it gets used that way all the 451 00:29:36,680 --> 00:29:39,600 Speaker 1: time is another matter. But the purpose of calling it 452 00:29:39,720 --> 00:29:43,840 Speaker 1: candidate photo is to remind everyone this is not meant 453 00:29:43,840 --> 00:29:50,000 Speaker 1: to be proof of someone's guilt or innocence. The FBI 454 00:29:50,120 --> 00:29:54,040 Speaker 1: also allows certain state authorities to use this same database, 455 00:29:54,400 --> 00:29:59,840 Speaker 1: and different agencies have different preferences. So in the GAO 456 00:30:00,080 --> 00:30:02,720 Speaker 1: report that I talked about earlier, the authors noted that 457 00:30:02,800 --> 00:30:07,040 Speaker 1: law enforcement officials from Michigan, for example, would always ask 458 00:30:07,080 --> 00:30:10,960 Speaker 1: for the maximum number of candidate photos, particularly when they'd 459 00:30:11,000 --> 00:30:14,960 Speaker 1: use probe images that were of low quality. So let's 460 00:30:14,960 --> 00:30:18,600 Speaker 1: say you've got a picture captured from a security camera 461 00:30:18,840 --> 00:30:21,640 Speaker 1: and the lighting is pretty bad and perhaps the person 462 00:30:21,760 --> 00:30:24,720 Speaker 1: wasn't facing dead on into the camera. You might ask 463 00:30:24,760 --> 00:30:27,560 Speaker 1: for the maximum number of candidate photos to re return 464 00:30:27,640 --> 00:30:32,000 Speaker 1: to you, knowing that the image you submitted was low quality, 465 00:30:32,000 --> 00:30:37,960 Speaker 1: and therefore any match is only potentially going to be 466 00:30:38,040 --> 00:30:42,720 Speaker 1: the person you're actually looking for. And again, this is 467 00:30:42,760 --> 00:30:46,200 Speaker 1: all just to help you with the beginning of your investigation. 468 00:30:46,480 --> 00:30:50,000 Speaker 1: It's not meant to be the that's our guy moment 469 00:30:50,360 --> 00:30:55,720 Speaker 1: that you would see and say police procedural that would 470 00:30:55,720 --> 00:31:00,400 Speaker 1: appear on network television in primetime. The FBI I also 471 00:31:00,400 --> 00:31:04,360 Speaker 1: has a policy in that all returned candidate photos must 472 00:31:04,400 --> 00:31:08,440 Speaker 1: first be analyzed by human specialists before being passed on 473 00:31:08,600 --> 00:31:12,920 Speaker 1: to other law enforcement agencies. Up to that point, the 474 00:31:13,080 --> 00:31:16,800 Speaker 1: entire process is automatic, so you don't have people overseeing 475 00:31:16,960 --> 00:31:20,960 Speaker 1: the process once it's probing all of the database, but 476 00:31:21,000 --> 00:31:24,240 Speaker 1: once the results come in, human analysts, who are supposed 477 00:31:24,240 --> 00:31:26,479 Speaker 1: to be trained in this sort of thing, are supposed 478 00:31:26,520 --> 00:31:30,520 Speaker 1: to look at each of those returned candidates and determine 479 00:31:30,520 --> 00:31:34,800 Speaker 1: if whether or not they really do resemble the person 480 00:31:34,920 --> 00:31:37,440 Speaker 1: in the probe photo that was submitted in the first place, 481 00:31:37,480 --> 00:31:39,440 Speaker 1: and if they're not, they are not supposed to be 482 00:31:39,480 --> 00:31:43,240 Speaker 1: passed on any further down the chain. Now, so far, 483 00:31:43,320 --> 00:31:47,320 Speaker 1: this probably doesn't sound too problematic. The FBI has a 484 00:31:47,360 --> 00:31:50,480 Speaker 1: database containing both criminal and civil photographs, but when it 485 00:31:50,560 --> 00:31:53,120 Speaker 1: runs a probe, it can only use the criminal photos 486 00:31:53,280 --> 00:31:55,720 Speaker 1: or the civil ones that are attached to criminal files. 487 00:31:56,200 --> 00:31:58,720 Speaker 1: Candidate photos are supposed to only be used to help 488 00:31:58,720 --> 00:32:02,440 Speaker 1: start a line of inquiry, not to positively identify suspects, 489 00:32:02,680 --> 00:32:05,400 Speaker 1: and everything has to be reviewed by human being. That 490 00:32:05,520 --> 00:32:09,080 Speaker 1: sounds fairly reasonable. But even if you're mostly okay with 491 00:32:09,120 --> 00:32:12,600 Speaker 1: this approach, which still has some problems we'll talk about 492 00:32:12,600 --> 00:32:16,040 Speaker 1: in a bit, things get significantly more dicey as you 493 00:32:16,160 --> 00:32:20,640 Speaker 1: learn more about the FBI's policies. For example, they have 494 00:32:20,680 --> 00:32:25,680 Speaker 1: a unit called the Facial Analysis Comparison and Evaluation Services 495 00:32:25,840 --> 00:32:32,520 Speaker 1: or face FACE. This is a part of the Criminal 496 00:32:32,600 --> 00:32:37,400 Speaker 1: Justice Information Services Department CG. Rather I yeah, I can 497 00:32:37,440 --> 00:32:41,320 Speaker 1: spell justice with a G. It doesn't make sense. No, 498 00:32:41,520 --> 00:32:45,959 Speaker 1: the cjis. This is a department within the FBI, and 499 00:32:46,040 --> 00:32:48,960 Speaker 1: FACE can carry out a search far more wide reaching 500 00:32:49,320 --> 00:32:55,160 Speaker 1: than one that just uses the ngi IPS database. FACE 501 00:32:55,320 --> 00:33:00,320 Speaker 1: uses not only that database but also external databases when 502 00:33:00,360 --> 00:33:03,480 Speaker 1: conducting a search with a probe photo. So let's say again, 503 00:33:03,800 --> 00:33:06,640 Speaker 1: you're an FBI agent and you have an image that 504 00:33:06,680 --> 00:33:08,600 Speaker 1: you want to match. You want to find out who 505 00:33:08,640 --> 00:33:11,240 Speaker 1: this person is. Maybe it's just a person of interest, 506 00:33:11,600 --> 00:33:15,040 Speaker 1: doesn't even necessarily have to be a suspect. Could be that, hey, 507 00:33:15,080 --> 00:33:17,560 Speaker 1: maybe this person can tell us more about this thing 508 00:33:17,640 --> 00:33:22,800 Speaker 1: that happened later on. Well, you could follow the NGIIPS procedure, 509 00:33:22,840 --> 00:33:26,360 Speaker 1: which would focus on those criminal photographs, or you could 510 00:33:26,400 --> 00:33:31,760 Speaker 1: submit your image to face. Face then would search dozens 511 00:33:31,880 --> 00:33:37,840 Speaker 1: of databases holding more than four hundred eleven million photographs, 512 00:33:38,800 --> 00:33:43,880 Speaker 1: many of which are from civil sources. So NGIIPS has 513 00:33:44,080 --> 00:33:48,120 Speaker 1: thirty million, all of them together have four hundred eleven 514 00:33:48,160 --> 00:33:52,680 Speaker 1: million pictures. And again a lot of those pictures just 515 00:33:52,720 --> 00:34:02,080 Speaker 1: come from things like passport ID, driver's licenses, sometimes security clearances, 516 00:34:02,200 --> 00:34:05,239 Speaker 1: that sort of stuff. That's this database has a lot 517 00:34:05,280 --> 00:34:08,879 Speaker 1: of law abiding citizens who have no criminal record, and 518 00:34:09,040 --> 00:34:11,120 Speaker 1: the images have nothing to do with any sort of 519 00:34:11,120 --> 00:34:17,120 Speaker 1: criminal act, but they're in these databases. These external databases 520 00:34:17,400 --> 00:34:20,920 Speaker 1: belong to lots of different agencies, and both at the 521 00:34:20,920 --> 00:34:25,319 Speaker 1: federal level and state level. So you've got state police agencies, 522 00:34:25,760 --> 00:34:28,080 Speaker 1: You've got the Department of Defense, You've got the Department 523 00:34:28,080 --> 00:34:31,520 Speaker 1: of Justice, you have the Department of State, and again 524 00:34:31,560 --> 00:34:35,520 Speaker 1: it contains photos from licenses, passports, security ID cards, and more. 525 00:34:36,040 --> 00:34:38,800 Speaker 1: So your submission would then go to one of twenty 526 00:34:38,880 --> 00:34:43,240 Speaker 1: nine different biometric image specialists. They would take that probe 527 00:34:43,239 --> 00:34:46,080 Speaker 1: photo and run a scan through these various databases and 528 00:34:46,080 --> 00:34:49,360 Speaker 1: they would look for matches. Here's another problem. Each of 529 00:34:49,400 --> 00:34:53,280 Speaker 1: these systems has a different methodology for performing and returning 530 00:34:53,320 --> 00:34:58,560 Speaker 1: search results, which makes this even more complicated. For example, 531 00:34:59,200 --> 00:35:02,400 Speaker 1: I talked about how the ngi IPS system gives you 532 00:35:02,480 --> 00:35:06,799 Speaker 1: a return between two and fifty candidate photos. Right, Well, 533 00:35:06,840 --> 00:35:09,600 Speaker 1: the Department of State will return as many as eighty 534 00:35:09,640 --> 00:35:14,160 Speaker 1: eight candidate photos if they are all from visa applications 535 00:35:14,200 --> 00:35:17,880 Speaker 1: from people who are not US citizens. So you can 536 00:35:17,920 --> 00:35:21,960 Speaker 1: get up to eighty eight pictures from visa applicants, or 537 00:35:22,000 --> 00:35:26,960 Speaker 1: you could just get three images from US citizen passport applicants, 538 00:35:27,760 --> 00:35:30,800 Speaker 1: because that's a hard limit. They can only return three 539 00:35:30,840 --> 00:35:35,000 Speaker 1: candidate photos from US citizens who applied for passports, but 540 00:35:35,040 --> 00:35:38,640 Speaker 1: they can return up to eighty eight visa application photos. 541 00:35:40,080 --> 00:35:42,359 Speaker 1: The Department of Defense will will down all of their 542 00:35:42,440 --> 00:35:47,560 Speaker 1: candidates into a single entry. So, in other words, Diberna Defense, 543 00:35:47,640 --> 00:35:51,520 Speaker 1: if you query that database with your probe photo, you 544 00:35:51,560 --> 00:35:55,400 Speaker 1: will only get one image back, so they will call 545 00:35:55,600 --> 00:35:57,640 Speaker 1: all the other ones and give you the most likely 546 00:35:58,040 --> 00:36:00,400 Speaker 1: match out of all the ones that they find in 547 00:36:00,440 --> 00:36:07,200 Speaker 1: their search. Some states will do similar things where they 548 00:36:07,280 --> 00:36:11,160 Speaker 1: will narrow down which images they will return to you. 549 00:36:11,280 --> 00:36:13,040 Speaker 1: Some of them will just give you everything they've got. 550 00:36:13,440 --> 00:36:16,200 Speaker 1: Every match that comes up, they'll just return it back 551 00:36:16,360 --> 00:36:20,839 Speaker 1: to the FBI. So it's very complicated. You can't really 552 00:36:20,880 --> 00:36:25,560 Speaker 1: be sure what methods people are using to be certain 553 00:36:25,680 --> 00:36:29,360 Speaker 1: that the potential matches they have represent a good match, 554 00:36:29,440 --> 00:36:33,359 Speaker 1: a good chance that the person that they've returned is 555 00:36:33,440 --> 00:36:37,240 Speaker 1: actually the same one who is in the probe photo. 556 00:36:38,280 --> 00:36:41,920 Speaker 1: At any rate, you as an FBI agent, wouldn't get 557 00:36:42,080 --> 00:36:45,200 Speaker 1: all of these at all, all of these photos that 558 00:36:45,239 --> 00:36:47,480 Speaker 1: would come back, They would come back to that biometric 559 00:36:47,560 --> 00:36:51,000 Speaker 1: analyst over at face, So you send your request to 560 00:36:51,040 --> 00:36:54,600 Speaker 1: face face takes care of the rest. They get back 561 00:36:54,640 --> 00:36:57,680 Speaker 1: all these results. Then they go through the results they 562 00:36:57,719 --> 00:37:00,279 Speaker 1: get back and they whittle that down to one or 563 00:37:00,280 --> 00:37:03,080 Speaker 1: two candidate photos and they send those on to you, 564 00:37:03,360 --> 00:37:05,600 Speaker 1: the FBI agent. So by the time you get it, 565 00:37:05,760 --> 00:37:08,759 Speaker 1: you only see one or two out of the potentially 566 00:37:08,800 --> 00:37:13,480 Speaker 1: more than one hundred images that were returned on this search. 567 00:37:16,560 --> 00:37:20,480 Speaker 1: But you might ask, well, how frequently does this happen? 568 00:37:20,560 --> 00:37:23,840 Speaker 1: I mean, how often is the FBI looking at images, 569 00:37:23,920 --> 00:37:28,760 Speaker 1: including pictures of law abiding citizens in these virtual lineups. 570 00:37:28,760 --> 00:37:32,040 Speaker 1: It can't be that frequent, right, Well, again, according to 571 00:37:32,080 --> 00:37:37,360 Speaker 1: that GAO report, the FBI submitted two hundred fifteen thousand 572 00:37:37,560 --> 00:37:41,560 Speaker 1: searches between August twenty eleven, which is pretty much when 573 00:37:41,560 --> 00:37:45,080 Speaker 1: the program went into pilot mode and started to be 574 00:37:45,280 --> 00:37:50,600 Speaker 1: rolled out more widely, through December twenty fifteen two hundred 575 00:37:50,600 --> 00:37:54,960 Speaker 1: and fifteen thousand. From August twenty eleven to December twenty fifteen, 576 00:37:56,120 --> 00:38:00,960 Speaker 1: thirty six thousand of those searches were on state driver's 577 00:38:01,000 --> 00:38:05,480 Speaker 1: licensed databases. So it happens a lot thirty six thousand times. 578 00:38:05,560 --> 00:38:09,319 Speaker 1: Chances are if you are an adult in America, you 579 00:38:09,440 --> 00:38:12,120 Speaker 1: got like a coin flip situation that your image was 580 00:38:12,160 --> 00:38:14,759 Speaker 1: looked at at some time or another by an algorithm 581 00:38:15,000 --> 00:38:18,759 Speaker 1: comparing it to a probe photo in the pursuit of 582 00:38:18,880 --> 00:38:23,960 Speaker 1: information regarding a federal case or in some cases, state cases, 583 00:38:24,000 --> 00:38:28,920 Speaker 1: because the FBI has also allowed certain states law agencies 584 00:38:29,480 --> 00:38:34,560 Speaker 1: access to this approach. Now, according to the rules, the 585 00:38:34,640 --> 00:38:39,040 Speaker 1: FBI should have submitted some important documents to inform the 586 00:38:39,040 --> 00:38:43,680 Speaker 1: public of their policies and to lay down the regulations, 587 00:38:43,719 --> 00:38:46,760 Speaker 1: the rules, the processes that they would have to follow 588 00:38:47,120 --> 00:38:49,839 Speaker 1: in order for this to be fair, for it to 589 00:38:49,880 --> 00:38:53,360 Speaker 1: not encroach on your privacy or to violate civil liberties 590 00:38:53,440 --> 00:38:57,560 Speaker 1: or civil rights. Without those rules, the use of the 591 00:38:57,600 --> 00:39:03,439 Speaker 1: system is largely unread, which can lead to misuse, whether 592 00:39:03,480 --> 00:39:07,760 Speaker 1: it's intentional or otherwise. The Government Accountability Office specifically pointed 593 00:39:07,760 --> 00:39:11,520 Speaker 1: out two different types of notifications that the FBI either 594 00:39:11,800 --> 00:39:14,600 Speaker 1: failed to submit or was just very late in submitting. 595 00:39:15,040 --> 00:39:20,239 Speaker 1: The first is called a Privacy Impact assessment or PIA. Now, 596 00:39:20,280 --> 00:39:23,240 Speaker 1: as that name suggests, a PIA is meant to inform 597 00:39:23,280 --> 00:39:27,400 Speaker 1: the public about any potential conflicts with privacy with regards 598 00:39:27,440 --> 00:39:32,720 Speaker 1: to methods for collecting personal information. The FBI did submit 599 00:39:33,160 --> 00:39:37,120 Speaker 1: a PIA for its next generation system, but they did 600 00:39:37,120 --> 00:39:39,560 Speaker 1: it back in two thousand and eight when they first 601 00:39:39,680 --> 00:39:46,080 Speaker 1: launched the NGIIPS. According to the Government Accountability Office, the 602 00:39:46,120 --> 00:39:50,040 Speaker 1: FBI made enough significant changes to the system to warrant 603 00:39:50,120 --> 00:39:55,279 Speaker 1: another PIA that anytime you make a significant revision to 604 00:39:55,440 --> 00:39:59,880 Speaker 1: your personal information systems, you have to submit a new 605 00:40:00,719 --> 00:40:05,520 Speaker 1: because things have changed, and according to the GAO, the 606 00:40:05,640 --> 00:40:10,120 Speaker 1: FBI failed to do that for way too long. Now 607 00:40:10,239 --> 00:40:14,480 Speaker 1: ultimately the FBI would publish a new PIA, but by 608 00:40:14,480 --> 00:40:18,160 Speaker 1: that point, the Government Accountability Office said they had delayed 609 00:40:18,200 --> 00:40:23,040 Speaker 1: so long that it made it more problematic as a result, 610 00:40:23,120 --> 00:40:26,759 Speaker 1: because during the whole time that they were supposed to 611 00:40:27,000 --> 00:40:30,880 Speaker 1: have submitted this, they were actively using this system. It 612 00:40:30,920 --> 00:40:33,920 Speaker 1: wasn't like this was a system being tested. It was 613 00:40:34,040 --> 00:40:38,120 Speaker 1: actually being put to use in real cases. And that 614 00:40:38,280 --> 00:40:40,480 Speaker 1: kind of violates it, well, it doesn't. Kind of. It 615 00:40:40,560 --> 00:40:43,600 Speaker 1: violates a Privacy Act of nineteen seventy four, which states 616 00:40:44,040 --> 00:40:47,200 Speaker 1: that when you make these revisions, you're supposed to file 617 00:40:47,239 --> 00:40:52,600 Speaker 1: a PIA before you put it into use. According to 618 00:40:52,600 --> 00:40:56,200 Speaker 1: the GAO, the FBI failed to do so, and also 619 00:40:56,200 --> 00:40:59,960 Speaker 1: the longer you wait to file this the more entrenched though, 620 00:41:00,000 --> 00:41:04,040 Speaker 1: those uses come. So if you put a system in place, 621 00:41:05,160 --> 00:41:07,840 Speaker 1: you build everything out, you've actually taken the time to 622 00:41:07,880 --> 00:41:12,520 Speaker 1: do it, and then you publish a PIA any objections 623 00:41:12,520 --> 00:41:14,640 Speaker 1: that are raised, you could say, well, we've got a 624 00:41:14,680 --> 00:41:17,240 Speaker 1: system now, and it costs one point two billion dollars 625 00:41:17,280 --> 00:41:19,320 Speaker 1: to put it in place. It's going to cost more money, 626 00:41:19,560 --> 00:41:23,040 Speaker 1: taxpayer money for us to alter it, to remove it, 627 00:41:23,160 --> 00:41:27,920 Speaker 1: to change it. You could argue against any move to 628 00:41:28,280 --> 00:41:33,160 Speaker 1: amend the situation. And the GAO says, that's not playing 629 00:41:33,200 --> 00:41:40,759 Speaker 1: cricket or playing fair for my fellow Americans. So that's 630 00:41:40,800 --> 00:41:43,240 Speaker 1: a problem. But then there's another one. There's a second 631 00:41:43,239 --> 00:41:46,399 Speaker 1: type of report called a Systems of Records Notice or 632 00:41:46,640 --> 00:41:50,520 Speaker 1: sor in SORN. The Department of Justice was required to 633 00:41:50,560 --> 00:41:54,400 Speaker 1: submit a SORN upon the launch of NGIIPS, but didn't 634 00:41:54,400 --> 00:41:59,480 Speaker 1: do so until May fifth, twenty sixteen. The GAO criticized 635 00:41:59,520 --> 00:42:02,320 Speaker 1: both the FBI and the Department of Justice for failing 636 00:42:02,360 --> 00:42:04,680 Speaker 1: to inform the public of the nature of this technology 637 00:42:04,680 --> 00:42:09,520 Speaker 1: and how it might impact personal privacy. But wait, there's more. 638 00:42:10,239 --> 00:42:13,719 Speaker 1: The GAO report also accused the FBI of failing to 639 00:42:13,760 --> 00:42:16,680 Speaker 1: perform any audits to make certain the use of facial 640 00:42:16,719 --> 00:42:20,600 Speaker 1: recognition software isn't in violation of other policies, or even 641 00:42:20,680 --> 00:42:24,360 Speaker 1: to make sure it doesn't violate the Fourth Amendment rights 642 00:42:24,360 --> 00:42:26,759 Speaker 1: of US citizens. Now, for those of you who are 643 00:42:26,800 --> 00:42:29,920 Speaker 1: not US citizens, you might wonder what does this actually mean. Well, 644 00:42:29,920 --> 00:42:33,239 Speaker 1: the Fourth Amendment is supposed to protect us against unreasonable 645 00:42:33,239 --> 00:42:36,320 Speaker 1: search and seizure, and part of that means law enforcement 646 00:42:36,400 --> 00:42:39,719 Speaker 1: can't just demand to search you for no reason. And 647 00:42:39,880 --> 00:42:42,840 Speaker 1: some have argued that using facial recognition software without a 648 00:42:42,880 --> 00:42:49,160 Speaker 1: person's consent, using it invisibly and widespread essentially amounts to 649 00:42:50,160 --> 00:42:54,200 Speaker 1: crossing that line. Now, in the United States, we've got 650 00:42:54,239 --> 00:42:57,160 Speaker 1: plenty of examples of troublesome policies that seem to overstep 651 00:42:57,200 --> 00:43:00,520 Speaker 1: the bounds that are established by the Fourth Amendment. But 652 00:43:00,800 --> 00:43:04,920 Speaker 1: that's a tirade for an entirely different show, probably not 653 00:43:05,040 --> 00:43:06,759 Speaker 1: a tech stuff, maybe a stuff they don't want you 654 00:43:06,800 --> 00:43:09,520 Speaker 1: to know. There are a couple of laws in the 655 00:43:09,600 --> 00:43:11,800 Speaker 1: United States that are important to take note of here 656 00:43:12,239 --> 00:43:14,520 Speaker 1: besides that Fourth Amendment. One of them I just mentioned 657 00:43:14,520 --> 00:43:16,759 Speaker 1: the Privacy Act of nineteen seventy four, and the other 658 00:43:16,800 --> 00:43:19,399 Speaker 1: one is the e Government Act of two thousand and two. 659 00:43:20,080 --> 00:43:23,480 Speaker 1: The Privacy Act sets limitations on the collection, disclosure, and 660 00:43:23,640 --> 00:43:28,040 Speaker 1: use of personal information maintained in systems of records, including 661 00:43:28,040 --> 00:43:32,200 Speaker 1: the ones that law agencies use. The e Government Act 662 00:43:32,360 --> 00:43:35,279 Speaker 1: is the one that requires government agencies to conduct pias 663 00:43:35,840 --> 00:43:38,239 Speaker 1: to make certain that personal information is handled properly in 664 00:43:38,280 --> 00:43:41,680 Speaker 1: federal systems, and the GAO report alleges that the FBI 665 00:43:41,719 --> 00:43:46,360 Speaker 1: policy wasn't aligned with either of those. Now, part of 666 00:43:46,360 --> 00:43:48,680 Speaker 1: this accusation depends upon the fact that the FBI was 667 00:43:48,800 --> 00:43:53,120 Speaker 1: using face in investigations for years before they updated their SORN. 668 00:43:53,680 --> 00:43:57,520 Speaker 1: They're sworn. According to the Privacy Act, agencies must publish 669 00:43:57,560 --> 00:43:59,800 Speaker 1: a new SORN upon the establishment or revision of the 670 00:43:59,840 --> 00:44:02,200 Speaker 1: system of records. This is what I was talking about earlier, 671 00:44:02,200 --> 00:44:05,040 Speaker 1: except I think I said PIA earlier when actually I 672 00:44:05,080 --> 00:44:09,600 Speaker 1: met sor In. That's entirely my fault because I didn't 673 00:44:09,600 --> 00:44:11,759 Speaker 1: write in my notes and I was talking next to boraneously. 674 00:44:12,160 --> 00:44:15,880 Speaker 1: But SORN is what I should have said. The FBI 675 00:44:16,120 --> 00:44:19,400 Speaker 1: argued that it was continuously updating the database to refine 676 00:44:19,440 --> 00:44:23,720 Speaker 1: the system, but the GAO's argument was that you could 677 00:44:23,719 --> 00:44:27,839 Speaker 1: be continuously updating the system and argue, well, we don't 678 00:44:27,840 --> 00:44:31,400 Speaker 1: want to publish an sor in after every tiny revision 679 00:44:31,880 --> 00:44:36,920 Speaker 1: because it's wasteful and time consuming. The GAOS counter to 680 00:44:36,960 --> 00:44:39,480 Speaker 1: that is, yeah, but you were using this tool in 681 00:44:39,600 --> 00:44:43,960 Speaker 1: actual cases. If you were developing this, let's say, in 682 00:44:44,360 --> 00:44:46,960 Speaker 1: a department where you're not using real cases, you're just 683 00:44:48,560 --> 00:44:51,040 Speaker 1: gradually tweaking the system so that it's more and more 684 00:44:51,080 --> 00:44:55,480 Speaker 1: accurate in a controlled environment. That's one thing. But if 685 00:44:55,480 --> 00:44:59,720 Speaker 1: you're actively making use of the system in real world investigations, 686 00:45:00,360 --> 00:45:04,279 Speaker 1: you absolutely must adhere to these laws, because to do 687 00:45:04,440 --> 00:45:07,759 Speaker 1: otherwise is in violation to laws that are passing the 688 00:45:07,880 --> 00:45:12,040 Speaker 1: United States. So you can't have it both ways. You 689 00:45:12,080 --> 00:45:16,600 Speaker 1: can't continuously tweak a system and put it to official 690 00:45:16,719 --> 00:45:21,879 Speaker 1: use and not also file these reports. You could argue 691 00:45:21,880 --> 00:45:23,520 Speaker 1: the FBI was trying to have its cake and eat 692 00:45:23,560 --> 00:45:27,200 Speaker 1: it too, So the expression that I think I actually 693 00:45:27,280 --> 00:45:30,080 Speaker 1: use properly. All Right, we've got more to talk about, 694 00:45:30,560 --> 00:45:32,480 Speaker 1: but it's time for us to take another quick break 695 00:45:32,840 --> 00:45:44,399 Speaker 1: to thank our sponsor. All right, So, the Government Accountability 696 00:45:44,400 --> 00:45:48,920 Speaker 1: Office criticizes the FBI and various other agencies for failing 697 00:45:48,960 --> 00:45:52,759 Speaker 1: to establish the scope and use of its facial recognition technology. 698 00:45:52,760 --> 00:45:56,160 Speaker 1: But that's just the tip of the iceberg. Because the 699 00:45:56,239 --> 00:45:59,320 Speaker 1: GAO report goes on to make an equally troubling point 700 00:46:00,520 --> 00:46:03,640 Speaker 1: that the FBI had performed only a few studies on 701 00:46:03,640 --> 00:46:07,920 Speaker 1: how accurate these facial recognition systems were in the first place. So, 702 00:46:08,000 --> 00:46:10,680 Speaker 1: in other words, not only was this a poorly defined 703 00:46:10,719 --> 00:46:15,080 Speaker 1: and unregulated tool, but it's a tool of unknown accuracy 704 00:46:15,120 --> 00:46:19,879 Speaker 1: and precision, which is terrifying when you think about it now. 705 00:46:19,880 --> 00:46:23,520 Speaker 1: According to the report, the FBI did perform some initial 706 00:46:23,600 --> 00:46:28,560 Speaker 1: tests before they deployed the ngiibs, and then occasionally did 707 00:46:28,600 --> 00:46:31,600 Speaker 1: a couple of tests when they made some changes. But 708 00:46:32,400 --> 00:46:35,839 Speaker 1: there were problems with these tests. For one thing, they 709 00:46:35,840 --> 00:46:38,640 Speaker 1: were limited in scope and they didn't represent how the 710 00:46:38,680 --> 00:46:42,080 Speaker 1: system might be used out in the real world. When 711 00:46:42,120 --> 00:46:45,440 Speaker 1: they were actually running these tests, they ran on about 712 00:46:45,600 --> 00:46:49,319 Speaker 1: nine hundred thousand photographs in the database, so they took 713 00:46:49,320 --> 00:46:52,279 Speaker 1: a subset of the photos that they had. They took 714 00:46:52,360 --> 00:46:55,720 Speaker 1: nine hundred thousand of them, and they ran probe tests 715 00:46:56,360 --> 00:47:00,600 Speaker 1: using photos that they knew either were or were not 716 00:47:01,320 --> 00:47:05,359 Speaker 1: represented in that group of nine hundred thousand. However, you've 717 00:47:05,360 --> 00:47:08,760 Speaker 1: got to remember the full database is more than thirty 718 00:47:09,000 --> 00:47:13,680 Speaker 1: million images, so something that works on a smaller scale 719 00:47:13,760 --> 00:47:17,200 Speaker 1: may not work once you scale it up for another 720 00:47:17,640 --> 00:47:21,560 Speaker 1: The tests did not specify how often incorrect matches would 721 00:47:21,600 --> 00:47:25,799 Speaker 1: come back, so you didn't know how many false positives 722 00:47:26,200 --> 00:47:29,520 Speaker 1: were there because the FBI wasn't tracking false positives. They 723 00:47:29,560 --> 00:47:32,319 Speaker 1: were only concerned with how frequently they were getting a 724 00:47:32,440 --> 00:47:37,720 Speaker 1: match to an actual image. So the way they test 725 00:47:37,800 --> 00:47:41,319 Speaker 1: this is, you've got nine hundred thousand images, they've got 726 00:47:41,320 --> 00:47:44,360 Speaker 1: a probe image, They know for a fact that the 727 00:47:44,400 --> 00:47:47,719 Speaker 1: probe image is inside that database, and then they run 728 00:47:47,760 --> 00:47:51,160 Speaker 1: the search to see if the system sends that image back. 729 00:47:51,480 --> 00:47:55,279 Speaker 1: And their threshold was an eighty five percent detection rate 730 00:47:56,080 --> 00:47:59,160 Speaker 1: for a positive match. So, in other words, it went 731 00:47:59,239 --> 00:48:01,800 Speaker 1: like this, Let's say you need to conduct a test 732 00:48:01,880 --> 00:48:04,520 Speaker 1: of this system. This is one way you would determine 733 00:48:04,520 --> 00:48:07,439 Speaker 1: whether or not you had that eighty five percent detection rate. 734 00:48:08,719 --> 00:48:12,279 Speaker 1: Let's say you have one hundred probe photos that you've 735 00:48:12,320 --> 00:48:16,680 Speaker 1: taken of one person, and you know this person's face 736 00:48:16,960 --> 00:48:19,360 Speaker 1: is in that database. You know it's going to be 737 00:48:19,400 --> 00:48:24,160 Speaker 1: in among those nine hundred thousand or so images, So 738 00:48:24,239 --> 00:48:26,919 Speaker 1: then you submit your query. If you have an eighty 739 00:48:26,920 --> 00:48:29,960 Speaker 1: five percent detection rate, then eighty five of those probe 740 00:48:30,000 --> 00:48:33,160 Speaker 1: photos should come back with a match, and that match 741 00:48:33,239 --> 00:48:38,160 Speaker 1: should be the actual person you're looking for. That's what 742 00:48:38,200 --> 00:48:40,719 Speaker 1: they meant by an eighty five percent detection rate, that 743 00:48:40,800 --> 00:48:43,920 Speaker 1: eighty five percent of the time an image that is 744 00:48:43,960 --> 00:48:47,760 Speaker 1: in their database would be pulled due to a facial 745 00:48:47,800 --> 00:48:53,000 Speaker 1: recognition software search. Now, during this testing phase, the FBI 746 00:48:53,320 --> 00:48:57,720 Speaker 1: reported that they met this threshold. They used that subset 747 00:48:57,760 --> 00:49:00,640 Speaker 1: of actually was nine hundred and twenty six thousand photos 748 00:49:00,840 --> 00:49:03,480 Speaker 1: as their subset when they were testing it, and they 749 00:49:03,560 --> 00:49:06,440 Speaker 1: said that they had an eighty six percent detection rate, 750 00:49:06,520 --> 00:49:09,279 Speaker 1: So they actually were exceeding what they had set as 751 00:49:09,320 --> 00:49:12,720 Speaker 1: their threshold. But that just meant that eighty six percent 752 00:49:12,760 --> 00:49:15,200 Speaker 1: of the time, the actual match for a probe photos 753 00:49:15,239 --> 00:49:19,560 Speaker 1: showed up in a group of fifty candidate images, so 754 00:49:21,280 --> 00:49:23,680 Speaker 1: you would get forty nine other images that were not 755 00:49:23,920 --> 00:49:28,120 Speaker 1: your match. The match would be there eighty six percent 756 00:49:28,120 --> 00:49:32,600 Speaker 1: of the time along with forty nine other images. So 757 00:49:32,640 --> 00:49:34,840 Speaker 1: we know that the system works if you were asking 758 00:49:34,880 --> 00:49:38,520 Speaker 1: for the maximum number of candidates. Remember in the FBI system, 759 00:49:38,520 --> 00:49:40,840 Speaker 1: you can ask for between two and fifty, but fifty 760 00:49:40,960 --> 00:49:44,200 Speaker 1: is the max. But what happens if you ask for 761 00:49:44,280 --> 00:49:49,400 Speaker 1: fewer images? What if you said, no, I want twenty returns. 762 00:49:49,640 --> 00:49:54,280 Speaker 1: What's the accuracy, then the FBI can't tell you because 763 00:49:54,320 --> 00:49:57,200 Speaker 1: they do not know. According to the FBI, they did 764 00:49:57,239 --> 00:50:00,400 Speaker 1: not run tests to see what would happen if you 765 00:50:00,560 --> 00:50:03,799 Speaker 1: decrease the number of candidate photos you asked for. They 766 00:50:03,920 --> 00:50:07,640 Speaker 1: only ran tests on the maximum number of candidate photos. 767 00:50:08,760 --> 00:50:11,719 Speaker 1: And keep in mind, the default for any search is 768 00:50:11,840 --> 00:50:15,480 Speaker 1: twenty photos, so the default is less than what they tested, 769 00:50:15,520 --> 00:50:18,439 Speaker 1: and they never tried to see if the eighty six 770 00:50:18,480 --> 00:50:22,560 Speaker 1: percent detection rate held true at these lower numbers. That's 771 00:50:22,600 --> 00:50:27,800 Speaker 1: a huge issue. On top of that, the FBI didn't 772 00:50:27,800 --> 00:50:30,439 Speaker 1: go so far to determine how frequently its system would 773 00:50:30,480 --> 00:50:35,080 Speaker 1: return false positives to probes, so they never paid attention 774 00:50:35,200 --> 00:50:39,000 Speaker 1: to how many times they got responses that didn't reflect 775 00:50:39,120 --> 00:50:44,239 Speaker 1: and the actual identity. They didn't keep track of it. So, 776 00:50:44,280 --> 00:50:46,200 Speaker 1: according to the FBI, the purpose of the system is 777 00:50:46,239 --> 00:50:49,640 Speaker 1: to generate leads, not to positively identify persons of interest. 778 00:50:49,680 --> 00:50:51,840 Speaker 1: So it shouldn't come as a big surprise, or you 779 00:50:51,880 --> 00:50:55,800 Speaker 1: shouldn't even care if it returns a lot of false positives, 780 00:50:56,200 --> 00:51:00,360 Speaker 1: because hey, this technology isn't meant to be the smoking 781 00:51:00,440 --> 00:51:03,279 Speaker 1: gun that says, here's the evidence that will put this 782 00:51:03,320 --> 00:51:05,920 Speaker 1: person away. It's meant to just create a lead, So 783 00:51:06,200 --> 00:51:09,080 Speaker 1: why do you care how many false positives it returns? 784 00:51:10,000 --> 00:51:16,000 Speaker 1: As if being looped in on an official inquiry when 785 00:51:16,000 --> 00:51:18,600 Speaker 1: you had nothing to do with it isn't disruptive or 786 00:51:18,600 --> 00:51:22,440 Speaker 1: stressful or provoke anxiety. I don't know about you, guys, 787 00:51:23,000 --> 00:51:25,080 Speaker 1: but if I had a federal agent show up at 788 00:51:25,080 --> 00:51:29,080 Speaker 1: my door asking me weird questions about a case that 789 00:51:29,160 --> 00:51:32,880 Speaker 1: I had no connection to because my image had popped 790 00:51:32,960 --> 00:51:36,600 Speaker 1: up in one of these searches and I have nothing 791 00:51:36,640 --> 00:51:38,880 Speaker 1: to do with it, it just so happens that I 792 00:51:38,880 --> 00:51:41,840 Speaker 1: look enough like a photo that's being used in the 793 00:51:41,880 --> 00:51:45,640 Speaker 1: case to warrant this. I would probably find that pretty 794 00:51:45,640 --> 00:51:51,440 Speaker 1: disruptive in my life, so I would care about false positives. FBI, 795 00:51:51,719 --> 00:51:54,719 Speaker 1: at least according to this GAO report, apparently didn't think 796 00:51:54,760 --> 00:52:00,840 Speaker 1: it was that big a deal. Now, the GAO points 797 00:52:00,840 --> 00:52:02,879 Speaker 1: out that it is a big deal, and that they're 798 00:52:02,880 --> 00:52:06,120 Speaker 1: not the only ones to think so. The National Science 799 00:52:06,160 --> 00:52:09,840 Speaker 1: and Technology Council and the National Institute of Standards and 800 00:52:09,840 --> 00:52:13,800 Speaker 1: Technology both state then, in order to know how accurate 801 00:52:13,840 --> 00:52:17,160 Speaker 1: a system is, you need to know two pieces of information, 802 00:52:17,800 --> 00:52:20,799 Speaker 1: not just the detection rate, which the FBI claims is 803 00:52:20,840 --> 00:52:23,840 Speaker 1: eighty six percent at least when you're asking for fifty candidates, 804 00:52:24,840 --> 00:52:27,520 Speaker 1: but also the false positive rate. You have to know 805 00:52:27,640 --> 00:52:30,759 Speaker 1: both of them in order to understand how accurate a 806 00:52:30,840 --> 00:52:33,680 Speaker 1: system is, So only knowing one of those pieces of 807 00:52:33,680 --> 00:52:37,680 Speaker 1: information isn't enough to state this system is accurate or not. 808 00:52:38,200 --> 00:52:41,200 Speaker 1: You have to know both. So, not only does the 809 00:52:41,280 --> 00:52:44,840 Speaker 1: FBI not have a grasp on how accurate their system 810 00:52:44,920 --> 00:52:47,759 Speaker 1: is if you're asking for fewer than the maximum number 811 00:52:47,800 --> 00:52:51,280 Speaker 1: of candidates, they also don't know how often it returns 812 00:52:51,320 --> 00:52:54,600 Speaker 1: false positives. So the FBI has no way of knowing 813 00:52:54,640 --> 00:53:01,040 Speaker 1: how accurate this facial recognition software is that's being used 814 00:53:01,160 --> 00:53:06,880 Speaker 1: to actually further investigations for official investigations of the FBI 815 00:53:07,040 --> 00:53:11,080 Speaker 1: and also other state agencies that have access to the system, 816 00:53:11,920 --> 00:53:16,200 Speaker 1: That is beyond problematic. If you cannot say that the 817 00:53:16,239 --> 00:53:21,680 Speaker 1: system with any degree of certainty is above a certain 818 00:53:21,680 --> 00:53:25,399 Speaker 1: threshold of accuracy, why are you using it? Because? I mean, 819 00:53:25,440 --> 00:53:29,920 Speaker 1: it has the potential to dramatically impact people's lives and 820 00:53:30,280 --> 00:53:34,960 Speaker 1: potentially lead people down a pathway that could result in 821 00:53:35,520 --> 00:53:40,160 Speaker 1: false accusations and imprisonment. The person who is actually responsible 822 00:53:40,280 --> 00:53:42,799 Speaker 1: might totally get away with something because of this. This 823 00:53:42,840 --> 00:53:46,839 Speaker 1: is a real problem. And the thing is it might 824 00:53:46,880 --> 00:53:49,880 Speaker 1: be a perfectly accurate system, but we don't know that 825 00:53:50,239 --> 00:53:53,279 Speaker 1: because we haven't tested it. So until we test it, 826 00:53:53,320 --> 00:53:58,560 Speaker 1: we cannot just assume that it's accurate enough. That's not 827 00:53:58,600 --> 00:54:00,960 Speaker 1: when people's lives are at staate. This is where that 828 00:54:01,000 --> 00:54:04,960 Speaker 1: my bias doesn't so much creep in as it kicks 829 00:54:04,960 --> 00:54:07,680 Speaker 1: open the door and makes itself at home on your couch. 830 00:54:08,840 --> 00:54:15,719 Speaker 1: But I digress. The GAO report also goes into great 831 00:54:15,760 --> 00:54:20,719 Speaker 1: detail about how this accuracy really can have a clear 832 00:54:20,760 --> 00:54:24,000 Speaker 1: impact on people's privacy, their civil liberties, their civil rights. 833 00:54:24,600 --> 00:54:28,960 Speaker 1: They also cite the Electronic Frontier Foundation the EFF which 834 00:54:29,239 --> 00:54:31,279 Speaker 1: says that if a person is brought up as a 835 00:54:31,320 --> 00:54:34,680 Speaker 1: defendant in a case and it is revealed that they 836 00:54:34,680 --> 00:54:38,680 Speaker 1: were matched by a facial recognition system, it puts a 837 00:54:38,719 --> 00:54:41,600 Speaker 1: burden on the defendant to argue that they are not 838 00:54:41,800 --> 00:54:46,960 Speaker 1: the same person as was seen in a probe photo, 839 00:54:47,120 --> 00:54:49,040 Speaker 1: that they are not the same one that the system 840 00:54:49,080 --> 00:54:54,160 Speaker 1: has identified. And if you cannot reliably state how accurate 841 00:54:54,200 --> 00:54:57,080 Speaker 1: your system is because you don't know how frequently it 842 00:54:57,120 --> 00:55:01,160 Speaker 1: returns false positives, you have unfairly burned And the defendant, 843 00:55:01,880 --> 00:55:03,439 Speaker 1: Like if you were to say, if you're the FBI, 844 00:55:03,480 --> 00:55:05,879 Speaker 1: and you say, we have an eighty six percent detection rate, 845 00:55:06,400 --> 00:55:09,080 Speaker 1: but you don't admit, oh, by the way, we don't 846 00:55:09,080 --> 00:55:12,080 Speaker 1: know how many false positives we get on any given search. 847 00:55:12,840 --> 00:55:15,640 Speaker 1: The implication you have given is that we're pretty sure 848 00:55:15,680 --> 00:55:19,760 Speaker 1: that this is the right guy. And again they argue 849 00:55:19,760 --> 00:55:23,080 Speaker 1: that this is meant to be a point of inquiry, 850 00:55:23,600 --> 00:55:25,799 Speaker 1: but you could easily see how it could also be 851 00:55:25,880 --> 00:55:29,719 Speaker 1: used by a lawyer to argue that a defendant is 852 00:55:29,760 --> 00:55:32,960 Speaker 1: in fact the person responsible for a crime, and they 853 00:55:33,000 --> 00:55:37,640 Speaker 1: may not be. And because you don't know the accuracy 854 00:55:37,680 --> 00:55:42,240 Speaker 1: of the system, you can't using the system to argue 855 00:55:42,280 --> 00:55:48,480 Speaker 1: for it is irresponsible. There's no accountability there. Now. Not 856 00:55:48,520 --> 00:55:50,960 Speaker 1: only has the FBI failed to establish the accuracy of 857 00:55:51,000 --> 00:55:55,239 Speaker 1: its own NGIIPS system, it has also not assessed the 858 00:55:55,280 --> 00:55:58,880 Speaker 1: accuracy of all those external databases that are used whenever 859 00:55:58,880 --> 00:56:03,600 Speaker 1: they use the face approach. There are no accuracy requirements 860 00:56:03,640 --> 00:56:07,000 Speaker 1: for these agencies, so there's not like a threshold they 861 00:56:07,040 --> 00:56:09,560 Speaker 1: have to prove that they meet in order to be 862 00:56:09,640 --> 00:56:13,360 Speaker 1: part of this. That's a huge problem. While each agency 863 00:56:13,480 --> 00:56:17,000 Speaker 1: might be accurate with no testing procedure, in place, it's 864 00:56:17,040 --> 00:56:20,400 Speaker 1: impossible to be certain of that. And since these databases 865 00:56:20,440 --> 00:56:24,000 Speaker 1: include millions of people with no criminal background and they 866 00:56:24,040 --> 00:56:28,719 Speaker 1: all use different facial recognition software products, this is a 867 00:56:28,800 --> 00:56:31,440 Speaker 1: huge issue. You could be put in a virtual lineup 868 00:56:31,480 --> 00:56:34,560 Speaker 1: simply because you look enough like someone else that a 869 00:56:34,560 --> 00:56:38,120 Speaker 1: computer thinks you are in fact the same person. The 870 00:56:38,200 --> 00:56:42,040 Speaker 1: GAO report concludes with a host of recommendations for future actions, 871 00:56:42,760 --> 00:56:45,240 Speaker 1: including addressing the problem of the FBI being so slow 872 00:56:45,280 --> 00:56:48,960 Speaker 1: to publish those updated pias in a timely manner, and 873 00:56:49,160 --> 00:56:53,040 Speaker 1: create a means to assess each system's accuracy. The Department 874 00:56:53,080 --> 00:56:57,359 Speaker 1: of Justice read the report and then responded disagreeing with 875 00:56:57,400 --> 00:57:03,000 Speaker 1: several points that the GOAO report made, including arguing that 876 00:57:03,040 --> 00:57:06,200 Speaker 1: the FBI and the Department of Justice published information when 877 00:57:06,200 --> 00:57:08,759 Speaker 1: it made the most sense, when the system had been 878 00:57:08,800 --> 00:57:13,440 Speaker 1: tweaked and finalized. More or less. However, by that time, again, 879 00:57:14,080 --> 00:57:16,520 Speaker 1: they had been using that system for real world cases 880 00:57:16,880 --> 00:57:19,720 Speaker 1: throughout the entire process, So it seems to me to 881 00:57:19,760 --> 00:57:23,600 Speaker 1: be kind of a weak argument. You can't really say, like, hey, 882 00:57:23,640 --> 00:57:26,200 Speaker 1: it wasn't finished until then, that's when we published it. 883 00:57:26,840 --> 00:57:29,520 Speaker 1: If you also are saying, hey, we use that for 884 00:57:29,640 --> 00:57:34,520 Speaker 1: real zees to go after actual people. You can't have 885 00:57:34,640 --> 00:57:39,240 Speaker 1: it both ways and not maintain accountability at any rate. 886 00:57:42,320 --> 00:57:45,720 Speaker 1: So that kind of gets to the end of the 887 00:57:45,760 --> 00:57:48,520 Speaker 1: Government Accountability Office report, but that's not the end of 888 00:57:48,560 --> 00:57:52,040 Speaker 1: the story. In March twenty seventeen, Congress held some hearings 889 00:57:52,080 --> 00:57:55,880 Speaker 1: about this, and boy howdy, were some congress people very 890 00:57:55,920 --> 00:57:58,400 Speaker 1: upset with the FBI. On both sides of the aisle. 891 00:57:58,440 --> 00:58:02,600 Speaker 1: You had Democrats and Republicans really chastising the FBI for 892 00:58:02,720 --> 00:58:06,640 Speaker 1: their use of facial recognition software and arguing that it 893 00:58:06,680 --> 00:58:10,520 Speaker 1: could amount to an enormous invasion of privacy as well 894 00:58:10,560 --> 00:58:15,120 Speaker 1: as endangering the civil liberties of US citizens. So people 895 00:58:15,200 --> 00:58:20,960 Speaker 1: who have dramatically different political philosophies were agreeing on this point. 896 00:58:21,040 --> 00:58:24,000 Speaker 1: So it wasn't really a partisan issue in this case, 897 00:58:24,480 --> 00:58:26,960 Speaker 1: and it got pretty ugly, but probably not as ugly 898 00:58:27,000 --> 00:58:29,880 Speaker 1: as the Georgetown University report that was published in late 899 00:58:29,880 --> 00:58:34,320 Speaker 1: twenty sixteen. This is an amazing report. Both the Government 900 00:58:34,320 --> 00:58:38,160 Speaker 1: Accountability Office report and the Georgetown University report are available 901 00:58:38,280 --> 00:58:42,360 Speaker 1: for free online. I will warn you collectively, they're about 902 00:58:42,360 --> 00:58:46,760 Speaker 1: two hundred pages, so if you want some light reading 903 00:58:47,600 --> 00:58:50,280 Speaker 1: you can check it out. They are quite good, both 904 00:58:50,280 --> 00:58:52,520 Speaker 1: of them. And they're very accessible. Neither of them are 905 00:58:52,520 --> 00:58:57,000 Speaker 1: written in crazy legallees which will make it impossible to understand. 906 00:58:57,000 --> 00:59:00,840 Speaker 1: They're written in very plain English, as in the Georgetown 907 00:59:00,920 --> 00:59:03,640 Speaker 1: University report that was revealed that one in every two 908 00:59:03,680 --> 00:59:07,280 Speaker 1: American adults has their picture contained in a database connected 909 00:59:07,320 --> 00:59:11,560 Speaker 1: to law enforcement facial recognition systems. And this report goes 910 00:59:11,600 --> 00:59:15,040 Speaker 1: far beyond just that FBI to state all the way 911 00:59:15,040 --> 00:59:17,320 Speaker 1: down to state and local systems that are implementing their 912 00:59:17,320 --> 00:59:20,400 Speaker 1: own facial recognition databases, and many of them have no 913 00:59:20,560 --> 00:59:23,440 Speaker 1: understanding of how it might impact the civil liberties or 914 00:59:23,480 --> 00:59:27,200 Speaker 1: privacy of citizens. The report is the summary of a 915 00:59:27,240 --> 00:59:30,400 Speaker 1: study that lasted a full year with more than one 916 00:59:30,480 --> 00:59:34,000 Speaker 1: hundred records requests to various police departments. They looked at 917 00:59:34,040 --> 00:59:37,560 Speaker 1: fifty two different law enforcement agencies across the United States, 918 00:59:38,040 --> 00:59:40,960 Speaker 1: and the report assessed the risks to civil liberties and 919 00:59:41,000 --> 00:59:45,440 Speaker 1: civil rights because up until this report was filed, no 920 00:59:45,560 --> 00:59:48,560 Speaker 1: such study had been made, which is a huge problem. 921 00:59:48,880 --> 00:59:51,000 Speaker 1: You don't know the impact of the tool that you've 922 00:59:51,040 --> 00:59:54,360 Speaker 1: created until after it's been put in use for a while. 923 00:59:54,480 --> 00:59:58,440 Speaker 1: That's an issue. Ideally, you think all this out before 924 00:59:58,480 --> 01:00:02,320 Speaker 1: you implement the procedure and their findings were pretty upsetting. 925 01:00:03,000 --> 01:00:05,640 Speaker 1: For example, the report found that some agencies limit themselves 926 01:00:05,680 --> 01:00:08,960 Speaker 1: to using facial recognition within the framework of a targeted 927 01:00:09,080 --> 01:00:11,960 Speaker 1: and public use, such as using it on someone who 928 01:00:12,040 --> 01:00:16,760 Speaker 1: has been legally arrested or detained for a crime. And 929 01:00:16,840 --> 01:00:22,040 Speaker 1: in this case, you're talking about totally above board approach. 930 01:00:22,720 --> 01:00:27,600 Speaker 1: You're assuming that everyone is following the law as regards 931 01:00:27,600 --> 01:00:31,800 Speaker 1: to apprehending and charging a suspect with a crime, and 932 01:00:31,840 --> 01:00:36,080 Speaker 1: maybe that person is unwilling or unable to tell you 933 01:00:36,160 --> 01:00:39,160 Speaker 1: what their identity is, and in that case, you would 934 01:00:39,240 --> 01:00:42,720 Speaker 1: use this facial recognition software stuff in order to figure 935 01:00:42,760 --> 01:00:47,920 Speaker 1: out who you are dealing with. That's largely a legitimate 936 01:00:47,960 --> 01:00:53,080 Speaker 1: case the government. The Georgetown University study didn't say that's bad. 937 01:00:53,240 --> 01:00:56,600 Speaker 1: They actually said, no, that makes sense. It's targeted, it's public. 938 01:00:57,560 --> 01:01:02,600 Speaker 1: But you could have a more invisible approach, for example, 939 01:01:03,080 --> 01:01:06,400 Speaker 1: using facial recognition software in real time on a closed 940 01:01:06,520 --> 01:01:11,120 Speaker 1: circuit camera pointed at a city street, where you're literally 941 01:01:11,160 --> 01:01:14,680 Speaker 1: picking up people as they pass by. They're not people 942 01:01:14,720 --> 01:01:17,800 Speaker 1: of interest, they're just people going about their day. And 943 01:01:17,880 --> 01:01:21,120 Speaker 1: if you're running facial recognition software on such a feed, 944 01:01:21,800 --> 01:01:27,320 Speaker 1: you are potentially invading privacy and stepping on civil rights 945 01:01:27,360 --> 01:01:28,160 Speaker 1: and civil liberties. 946 01:01:28,840 --> 01:01:31,400 Speaker 2: Hey, it's modern day, Jonathan here just cutting end to 947 01:01:31,480 --> 01:01:34,160 Speaker 2: say we will have more about the National Facial Recognition 948 01:01:34,280 --> 01:01:36,240 Speaker 2: Database after this break. 949 01:01:45,440 --> 01:01:49,439 Speaker 1: So even if you were to argue that this real 950 01:01:49,440 --> 01:01:51,400 Speaker 1: time use where you're just looking at people as they 951 01:01:51,440 --> 01:01:53,280 Speaker 1: pass by, and maybe a little name pops up every 952 01:01:53,320 --> 01:01:55,760 Speaker 1: now and then as it as the system recognizes a 953 01:01:55,760 --> 01:01:59,680 Speaker 1: person that matches a file in the database, it's easy 954 01:01:59,680 --> 01:02:03,360 Speaker 1: to a scenario in which such a technology could be abused. 955 01:02:04,400 --> 01:02:09,120 Speaker 1: Either it picks up somebody mistakenly, it thinks it identifies someone, 956 01:02:09,400 --> 01:02:11,880 Speaker 1: but in fact it's a totally different person, and then 957 01:02:12,040 --> 01:02:17,440 Speaker 1: you end up establishing a person's location by mistake, like 958 01:02:17,480 --> 01:02:20,440 Speaker 1: it's not really where they were, but because the system 959 01:02:20,480 --> 01:02:24,000 Speaker 1: has identified a person as being at X place at 960 01:02:24,040 --> 01:02:29,240 Speaker 1: why time, you then have established supposedly that person's location, 961 01:02:30,000 --> 01:02:31,760 Speaker 1: when in fact that person might be across town or 962 01:02:31,840 --> 01:02:34,240 Speaker 1: not even in the same state. But it's because of 963 01:02:34,880 --> 01:02:38,120 Speaker 1: a misidentification in the system. That's one problem. But think 964 01:02:38,120 --> 01:02:42,320 Speaker 1: about this. Think of this is a scary scenario. Imagine 965 01:02:42,320 --> 01:02:45,360 Speaker 1: a situation in which a group of people are discriminated 966 01:02:45,400 --> 01:02:49,120 Speaker 1: against by a government agency. Let's say they have a 967 01:02:49,200 --> 01:02:54,840 Speaker 1: legitimate gripe. It's completely legitimate. They're victims of unfair treatment. 968 01:02:55,160 --> 01:02:57,520 Speaker 1: So a group of them at some of their allies 969 01:02:57,920 --> 01:03:01,280 Speaker 1: get together in a public place for peaceful protest, to 970 01:03:01,680 --> 01:03:06,440 Speaker 1: raise awareness of this issue and to confront the government 971 01:03:06,440 --> 01:03:10,960 Speaker 1: agencies that have discriminated against them. This is all perfectly 972 01:03:11,040 --> 01:03:14,160 Speaker 1: legal according to the US Constitution. They're not doing anything legal. 973 01:03:14,160 --> 01:03:19,000 Speaker 1: They're assembling on public grounds in order to practice free speech. 974 01:03:20,760 --> 01:03:24,400 Speaker 1: But it's not hard to imagine a government agency using 975 01:03:24,440 --> 01:03:27,040 Speaker 1: a camera with this sort of facial recognition software to 976 01:03:27,120 --> 01:03:29,760 Speaker 1: identify people who are in the crowd in order to 977 01:03:29,880 --> 01:03:33,520 Speaker 1: use that as leverage in the future for some purpose 978 01:03:33,640 --> 01:03:36,800 Speaker 1: or another, even if it's just to say we know 979 01:03:36,960 --> 01:03:40,640 Speaker 1: you were there, and to put that kind of pressure 980 01:03:40,720 --> 01:03:46,920 Speaker 1: on a person in order to essentially squelch people's freedom 981 01:03:46,920 --> 01:03:49,960 Speaker 1: of speech. So this is a First Amendment issue, not 982 01:03:50,000 --> 01:03:53,040 Speaker 1: just a Fourth Amendment issue. Now that might sound like 983 01:03:53,040 --> 01:03:56,960 Speaker 1: a dramatic scenario like something like Big brother Ish. It's orwellian, 984 01:03:57,920 --> 01:04:01,080 Speaker 1: but it's also entirely within the realm of possibility. From 985 01:04:01,120 --> 01:04:05,680 Speaker 1: a technological standpoint, there's nothing technologically oriented that would prevent 986 01:04:05,760 --> 01:04:08,240 Speaker 1: us from doing this or prevent an agency from doing this, 987 01:04:08,840 --> 01:04:12,280 Speaker 1: and even without the evil empire scenario in place, you 988 01:04:12,360 --> 01:04:15,280 Speaker 1: still have the problematic issue of treading on civil liberties 989 01:04:15,360 --> 01:04:19,800 Speaker 1: just by having such technology available and unregulated. You don't 990 01:04:19,840 --> 01:04:25,240 Speaker 1: have rules to guide this sort of stuff. The Georgetown 991 01:04:25,320 --> 01:04:29,040 Speaker 1: report found that only one agency out of the fifty 992 01:04:29,120 --> 01:04:35,000 Speaker 1: two that they looked at have a specific rule against 993 01:04:35,120 --> 01:04:39,760 Speaker 1: using facial recognition software to identify people participating in public 994 01:04:39,800 --> 01:04:44,400 Speaker 1: demonstrations or free speech in general. So only one agency 995 01:04:44,440 --> 01:04:47,560 Speaker 1: actually has rules against that. Now, that doesn't mean the 996 01:04:47,560 --> 01:04:51,760 Speaker 1: other fifty one agencies are regularly using this technology to 997 01:04:52,400 --> 01:04:56,720 Speaker 1: monitor acts of free speech, but it also doesn't mean 998 01:04:56,720 --> 01:04:59,560 Speaker 1: that they can't. They don't have rules against it. Only 999 01:04:59,600 --> 01:05:04,120 Speaker 1: one a agency out of the fifty two, people are 1000 01:05:04,120 --> 01:05:06,920 Speaker 1: being watched and identified without any connection to a crime. 1001 01:05:06,960 --> 01:05:11,560 Speaker 1: In these cases, it's pretty terrifying. The Georgetown report also 1002 01:05:11,600 --> 01:05:13,800 Speaker 1: found that no state had yet passed a law to 1003 01:05:13,920 --> 01:05:18,880 Speaker 1: regulate police use of facial recognition software. No state in 1004 01:05:18,920 --> 01:05:21,800 Speaker 1: the US. They're fifty of them, and none of them 1005 01:05:21,840 --> 01:05:25,520 Speaker 1: have passed any regulations, any laws to regulate the use 1006 01:05:25,520 --> 01:05:29,200 Speaker 1: of facial recognition software. So without rules, how do you 1007 01:05:29,320 --> 01:05:32,360 Speaker 1: argue whether someone's misused or abused a system, you have 1008 01:05:32,400 --> 01:05:35,280 Speaker 1: to have rules so that you know what is allowed 1009 01:05:35,320 --> 01:05:38,440 Speaker 1: and what is not allowed. With no rules, the implication 1010 01:05:38,560 --> 01:05:43,280 Speaker 1: is that everything's allowed until it isn't. That's a huge 1011 01:05:43,440 --> 01:05:50,160 Speaker 1: dangerous problem. The report also pointed out that most of 1012 01:05:50,200 --> 01:05:53,920 Speaker 1: these agencies lacked any sort of methodology to ensure that 1013 01:05:54,120 --> 01:05:59,440 Speaker 1: the accuracy of their respective systems was decent. The report 1014 01:05:59,480 --> 01:06:03,000 Speaker 1: stated that of all the agencies they investigated, only two, 1015 01:06:03,880 --> 01:06:07,000 Speaker 1: the San Francisco Police Department and the South Sound nine 1016 01:06:07,040 --> 01:06:11,240 Speaker 1: to one one from Seattle, had made decisions about what 1017 01:06:11,440 --> 01:06:14,720 Speaker 1: facial recognition software they were going to incorporate in their 1018 01:06:15,320 --> 01:06:20,800 Speaker 1: office based off of accuracy rates. That was not a 1019 01:06:20,840 --> 01:06:24,040 Speaker 1: consideration for all of the other agencies, at least not 1020 01:06:24,120 --> 01:06:28,160 Speaker 1: the ones that they asked. Moreover, they report points out 1021 01:06:28,200 --> 01:06:30,840 Speaker 1: that facial recognition companies are also trying to have it 1022 01:06:30,920 --> 01:06:34,959 Speaker 1: both ways. So, for example, they cite a company called 1023 01:06:35,120 --> 01:06:39,600 Speaker 1: fat Face First. Now face First advertises that has a 1024 01:06:39,680 --> 01:06:45,360 Speaker 1: ninety five percent accuracy rate, but it simultaneously disclaims any 1025 01:06:45,440 --> 01:06:49,600 Speaker 1: liability for failing to meet that ninety five percent accuracy rate. 1026 01:06:51,120 --> 01:06:53,840 Speaker 1: So it's kind of like saying we guarantee these tires. 1027 01:06:53,880 --> 01:06:58,120 Speaker 1: Tires are not guaranteed not quite like that, but similar. 1028 01:06:59,040 --> 01:07:02,920 Speaker 1: So again, this is according to the Georgetown University report, 1029 01:07:03,720 --> 01:07:07,560 Speaker 1: that's a problem for a company to sell itself on 1030 01:07:08,360 --> 01:07:13,600 Speaker 1: a performance threshold, but then say, hey, you can't hold 1031 01:07:13,680 --> 01:07:15,920 Speaker 1: us to that performance threshold that we sold you on. 1032 01:07:16,840 --> 01:07:21,760 Speaker 1: That's a little dangerous there too. Then the report goes 1033 01:07:21,800 --> 01:07:24,040 Speaker 1: on to state that the human analysts, you know, the 1034 01:07:24,080 --> 01:07:27,040 Speaker 1: ones I was talking about earlier, that supposed to be 1035 01:07:27,160 --> 01:07:32,560 Speaker 1: a safeguard. Human analysts are supposed to take the images 1036 01:07:32,600 --> 01:07:37,080 Speaker 1: that are returned by these automated systems and manually review 1037 01:07:37,160 --> 01:07:39,360 Speaker 1: them to make sure that they do or do not 1038 01:07:39,640 --> 01:07:42,960 Speaker 1: match that probe photo. That was the whole thing to 1039 01:07:43,000 --> 01:07:47,480 Speaker 1: begin with. But it turns out, according to this report, 1040 01:07:47,560 --> 01:07:52,080 Speaker 1: those human analysts are not that accurate. In fact, they're 1041 01:07:52,160 --> 01:07:56,400 Speaker 1: no better than a coin flip. Literally. The report sites 1042 01:07:56,440 --> 01:07:59,640 Speaker 1: of study that showed that if analysts did not have 1043 01:08:00,120 --> 01:08:05,200 Speaker 1: highly specialized training, they would make the wrong decision for 1044 01:08:05,280 --> 01:08:09,000 Speaker 1: a potential match fifty percent of the time. Literally a 1045 01:08:09,000 --> 01:08:13,800 Speaker 1: coin flip. That's ridiculous. Now, the report found only eight 1046 01:08:13,880 --> 01:08:18,120 Speaker 1: agencies out of the fifty two used specialized personnel to 1047 01:08:18,200 --> 01:08:22,479 Speaker 1: review images. In other words, people who presumably have actually 1048 01:08:22,520 --> 01:08:26,320 Speaker 1: received that highly specialized training necessary to make more accurate 1049 01:08:26,360 --> 01:08:30,280 Speaker 1: decisions regarding these photos, and the report states that there's 1050 01:08:30,400 --> 01:08:34,800 Speaker 1: no formal training regime in place for examiners, which is 1051 01:08:34,800 --> 01:08:37,400 Speaker 1: a major problem for a system that's already in widespread use. 1052 01:08:37,680 --> 01:08:40,559 Speaker 1: So not only do you need highly specialized training, there's 1053 01:08:40,680 --> 01:08:47,240 Speaker 1: no formalized approach to give or receive that highly specialized training. 1054 01:08:48,240 --> 01:08:50,879 Speaker 1: So we know you need it, but we haven't developed 1055 01:08:50,920 --> 01:08:54,919 Speaker 1: the best practices to actually deliver upon that. So meanwhile, 1056 01:08:54,920 --> 01:08:58,400 Speaker 1: you've got human analysts who are making mistakes half the 1057 01:08:58,479 --> 01:09:02,200 Speaker 1: time while reviewing these photo And if you wonder if 1058 01:09:02,240 --> 01:09:06,559 Speaker 1: facial recognition systems would disproportionately affect some ethnicities over others, 1059 01:09:06,600 --> 01:09:10,880 Speaker 1: the answer to that is resounding and dismaying yes. The 1060 01:09:11,000 --> 01:09:15,200 Speaker 1: report found that African Americans would be affected more than 1061 01:09:15,479 --> 01:09:19,640 Speaker 1: other ethnicities. According to an FBI co authored study that 1062 01:09:19,720 --> 01:09:24,160 Speaker 1: was cited by this Georgetown University report, several facial recognition 1063 01:09:24,200 --> 01:09:28,600 Speaker 1: algorithms are less accurate for Black people than for other ethnicities, 1064 01:09:28,880 --> 01:09:32,439 Speaker 1: and there's no independent testing process to determine if there's 1065 01:09:32,520 --> 01:09:36,280 Speaker 1: a racial bias in any of these facial recognition systems, 1066 01:09:36,520 --> 01:09:39,839 Speaker 1: so no one has developed a test to make certain 1067 01:09:40,400 --> 01:09:45,320 Speaker 1: that it is in fact accurate despite a person's age, gender, 1068 01:09:45,479 --> 01:09:49,600 Speaker 1: or race, without being able to verify that it is 1069 01:09:49,720 --> 01:09:54,599 Speaker 1: accurate across all parameters, you have opened up an enormous 1070 01:09:54,640 --> 01:09:59,439 Speaker 1: can of worms, and you are disproportionately affecting people just 1071 01:09:59,479 --> 01:10:02,879 Speaker 1: because of the race, because your system does not address 1072 01:10:03,000 --> 01:10:07,719 Speaker 1: that properly. The report also points out that the information 1073 01:10:07,760 --> 01:10:10,519 Speaker 1: about the systems in use had not been generally available 1074 01:10:10,560 --> 01:10:13,000 Speaker 1: to the public. In fact, all of the fifty two 1075 01:10:13,080 --> 01:10:20,160 Speaker 1: agencies they contacted, only four had publicly available use policies. So, 1076 01:10:20,200 --> 01:10:23,040 Speaker 1: in other words, only four of the fifty two could 1077 01:10:23,080 --> 01:10:27,519 Speaker 1: tell you what their general policy was as far as 1078 01:10:27,520 --> 01:10:30,960 Speaker 1: facial recognition software goes. That's less than ten percent of 1079 01:10:31,120 --> 01:10:34,599 Speaker 1: all of the agencies they looked at, and only one 1080 01:10:34,640 --> 01:10:38,720 Speaker 1: of those agencies, which was San Diego's Association of Governments, 1081 01:10:38,960 --> 01:10:42,800 Speaker 1: had legislative approval for its policy. All the others were 1082 01:10:42,840 --> 01:10:46,000 Speaker 1: just self appointed policies that had not passed through any 1083 01:10:46,080 --> 01:10:50,360 Speaker 1: kind of official legislative support. Finally, the report asserted that 1084 01:10:50,720 --> 01:10:54,400 Speaker 1: most of these systems did not have an official audit 1085 01:10:54,479 --> 01:10:58,440 Speaker 1: process to determine if or when someone misuses the systems. 1086 01:10:59,080 --> 01:11:02,400 Speaker 1: Nine agencies were or that they did have a process, 1087 01:11:03,000 --> 01:11:06,840 Speaker 1: but only one provided Georgetown with any evidence that they 1088 01:11:06,840 --> 01:11:09,360 Speaker 1: had a working audit system, and that was the Michigan 1089 01:11:09,400 --> 01:11:12,080 Speaker 1: State Police, by the way, who said, we have an 1090 01:11:12,080 --> 01:11:15,240 Speaker 1: audit system, and here's proof that it actually works the 1091 01:11:15,240 --> 01:11:17,639 Speaker 1: way we said it did. So good on you, Michigan 1092 01:11:17,680 --> 01:11:20,320 Speaker 1: State for our having that system in place and being 1093 01:11:20,360 --> 01:11:24,960 Speaker 1: able to back it up now. The Georgetown University report 1094 01:11:25,040 --> 01:11:27,519 Speaker 1: also urged some major changes in the way law enforcement 1095 01:11:27,600 --> 01:11:30,840 Speaker 1: uses facial recognition, including an appeal to Congress to create 1096 01:11:31,000 --> 01:11:34,160 Speaker 1: clear regulations to define the parameters of when such a 1097 01:11:34,160 --> 01:11:37,479 Speaker 1: system could be used. They also called for companies to 1098 01:11:37,520 --> 01:11:42,200 Speaker 1: publish processes that test their products accuracy regardless of race, gender, 1099 01:11:42,240 --> 01:11:47,680 Speaker 1: and age, to remove that possibility of bias. And if 1100 01:11:47,720 --> 01:11:52,200 Speaker 1: we're being really super kind and generous toward law enforcement, 1101 01:11:52,640 --> 01:11:55,440 Speaker 1: we could say this is just another case where technology 1102 01:11:55,560 --> 01:11:58,920 Speaker 1: has clearly outpaced the law. We see that all the time, 1103 01:11:59,360 --> 01:12:05,160 Speaker 1: driverless artificial intelligence, lots of different technologies are advancing far 1104 01:12:05,240 --> 01:12:10,599 Speaker 1: faster than legislation can keep up with. All right, that's fair, 1105 01:12:10,960 --> 01:12:15,320 Speaker 1: we see it happen. However, it's particularly troublesome that this 1106 01:12:15,360 --> 01:12:18,960 Speaker 1: is happening within law enforcement that is already employing this 1107 01:12:19,040 --> 01:12:23,120 Speaker 1: technology before we've developed the policies to guide it. It's 1108 01:12:23,200 --> 01:12:26,439 Speaker 1: one thing to say someone's out here working on a 1109 01:12:26,520 --> 01:12:29,360 Speaker 1: driverless car, and we need to start thinking about how 1110 01:12:29,400 --> 01:12:32,519 Speaker 1: are we going to regulate that in the future. Maybe 1111 01:12:32,600 --> 01:12:35,280 Speaker 1: right now we say you aren't allowed to operate your 1112 01:12:35,320 --> 01:12:38,720 Speaker 1: driverless car until we figured this out. That's fair. It's 1113 01:12:38,720 --> 01:12:42,440 Speaker 1: another thing to say, there's this technology that could potentially 1114 01:12:42,640 --> 01:12:45,639 Speaker 1: impact people's lives and we're allowing law enforcement to use 1115 01:12:45,680 --> 01:12:48,400 Speaker 1: it while we try and figure out the rules. That's 1116 01:12:48,960 --> 01:12:53,720 Speaker 1: at best a problem. And as I said at the 1117 01:12:53,760 --> 01:12:56,040 Speaker 1: top of the show, I'm really just talking about the 1118 01:12:56,120 --> 01:12:59,439 Speaker 1: United States with particulars here, but this is happening all 1119 01:12:59,479 --> 01:13:02,519 Speaker 1: around the world. There are lots of governments around the 1120 01:13:02,520 --> 01:13:07,520 Speaker 1: world that are incorporating facial recognition software along with law enforcement. 1121 01:13:08,040 --> 01:13:12,760 Speaker 1: So while I'm using specific US examples in this podcast, 1122 01:13:13,360 --> 01:13:16,120 Speaker 1: the same is true for lots of other places. Of course, 1123 01:13:16,439 --> 01:13:19,360 Speaker 1: the laws that protect the citizens can be different from 1124 01:13:19,360 --> 01:13:22,680 Speaker 1: country to country, and in some cases there might not 1125 01:13:22,720 --> 01:13:26,839 Speaker 1: be very many outlets for citizens to voice their concern 1126 01:13:27,160 --> 01:13:30,200 Speaker 1: or it might even be dangerous to do so. But 1127 01:13:30,680 --> 01:13:32,559 Speaker 1: this is something I think we need to be aware of. 1128 01:13:32,800 --> 01:13:36,360 Speaker 1: I'm not generally the kind of person who tells you 1129 01:13:36,400 --> 01:13:38,960 Speaker 1: that you're being watched or you know, you should be paranoid. 1130 01:13:39,439 --> 01:13:42,360 Speaker 1: But I'm also not the person to just sit back 1131 01:13:42,400 --> 01:13:46,920 Speaker 1: and let something go on when I feel like it's 1132 01:13:47,000 --> 01:13:50,639 Speaker 1: potentially more of a problem than a solution. 1133 01:13:51,920 --> 01:13:54,040 Speaker 2: Well that was it for the episode I did on 1134 01:13:54,080 --> 01:13:57,160 Speaker 2: the National Facial Recognition Database back in twenty seventeen. It's 1135 01:13:57,200 --> 01:14:01,559 Speaker 2: a topic I should definitely revisit. Obviously, there's so much 1136 01:14:01,600 --> 01:14:05,639 Speaker 2: going on here. There are so many concerning things about it, 1137 01:14:05,680 --> 01:14:10,040 Speaker 2: from surveillance states, to privacy and security concerns, to the 1138 01:14:10,120 --> 01:14:13,200 Speaker 2: fact that we've been seeing lots of companies try and 1139 01:14:13,280 --> 01:14:18,040 Speaker 2: use facial recognition to match people against databases to varying 1140 01:14:18,080 --> 01:14:22,240 Speaker 2: degrees of success, and that for people of color in particular, 1141 01:14:22,520 --> 01:14:26,960 Speaker 2: those degrees of success are not good. And I think 1142 01:14:27,000 --> 01:14:29,240 Speaker 2: there's a lot that we need to talk about as 1143 01:14:29,320 --> 01:14:31,320 Speaker 2: far as this goes, when it comes to things like, 1144 01:14:31,439 --> 01:14:37,000 Speaker 2: you know, individual rights and authoritarian abuse of these kind 1145 01:14:37,040 --> 01:14:40,599 Speaker 2: of technologies, and I think we do need to have 1146 01:14:40,640 --> 01:14:43,040 Speaker 2: another update on this, so I will put that on 1147 01:14:43,080 --> 01:14:46,320 Speaker 2: my list. I hope that you are all well, and 1148 01:14:46,360 --> 01:14:55,600 Speaker 2: I'll talk to you again really soon. Tech stuff is 1149 01:14:55,600 --> 01:15:01,280 Speaker 2: an iHeartRadio production for more podcasts from iheartradiosit the iHeartRadio app, 1150 01:15:01,439 --> 01:15:04,599 Speaker 2: Apple Podcasts, or wherever you listen to your favorite shows.