WEBVTT - Facial Recognition Machinery, Part 1 0:00:03.000 --> 0:00:04.960 Welcome Stuff to Blow your Mind. A production of I 0:00:05.000 --> 0:00:14.040 Heart Radios has to works. Hey, welcome to Stuff to 0:00:14.080 --> 0:00:16.720 Blow your mind. My name is Robert Lamb and I'm 0:00:16.800 --> 0:00:19.360 Joe McCormick, and I want to start off today by 0:00:19.400 --> 0:00:23.520 proposing a scenario for you to imagine. What if your 0:00:23.680 --> 0:00:29.240 body included a personal search bar. I know that sounds weird, 0:00:29.280 --> 0:00:31.000 that might be hard to picture, but just try to 0:00:31.040 --> 0:00:35.440 imagine your body. Your physical body has a digital interface 0:00:35.640 --> 0:00:39.280 that maybe anybody within a hundred feet of you can access. 0:00:39.960 --> 0:00:44.879 What if your physical body included a searchable database of 0:00:44.920 --> 0:00:48.960 pretty much everything you ever did or posted on the Internet. 0:00:49.600 --> 0:00:51.960 So whether you're out at a bar with your friends, 0:00:52.400 --> 0:00:54.680 or you're sitting on the subway on your way to work, 0:00:54.920 --> 0:00:57.480 or you're sitting in your car in traffic, or taking 0:00:57.480 --> 0:01:00.640 part in a protest march, or working out of the gym, 0:01:00.800 --> 0:01:03.880 or you're on a date, whatever, anybody who could see 0:01:03.960 --> 0:01:07.680 you would instantly have the ability to look up your 0:01:07.800 --> 0:01:11.480 personal information. They could find out your name, your resume, 0:01:11.959 --> 0:01:16.200 your contact info, your workplace, home address, maybe find all 0:01:16.200 --> 0:01:18.959 publicly available photos of you that are out there on 0:01:19.040 --> 0:01:23.400 Instagram or whatever. Maybe everything you've ever posted on Twitter, 0:01:23.600 --> 0:01:27.399 or on Facebook or whatever other social media, and more broadly, 0:01:27.840 --> 0:01:31.800 basically just everything you've done on the internet, purchasing history, 0:01:31.920 --> 0:01:35.800 search history, as well as your location history anywhere you've 0:01:35.840 --> 0:01:39.520 physically taken your phone with GPS enabled. I think most 0:01:39.560 --> 0:01:43.440 of us would probably recoil in horror at the idea 0:01:43.560 --> 0:01:46.800 that we would ever lose the ability to be anonymous 0:01:46.959 --> 0:01:50.440 in a public place. But perhaps the horrifying part of 0:01:50.480 --> 0:01:54.280 imagining this scenario is that I think what I'm describing 0:01:54.400 --> 0:01:58.480 is not only fairly plausible in some preliminary ways, this 0:01:58.560 --> 0:02:02.400 is already the case, at least in principle. All the 0:02:02.480 --> 0:02:08.000 foundations and support structure of this horrible hypothetical world are laid, 0:02:08.760 --> 0:02:10.919 and really all that's left to do is just kind 0:02:10.919 --> 0:02:13.400 of tighten the screws on it. And one thing we 0:02:13.400 --> 0:02:15.720 know about this world is that there there's no shortage 0:02:15.720 --> 0:02:19.120 of would be screw tighteners out there, Nope, especially if 0:02:19.120 --> 0:02:21.360 you can make some money by tightening screws, which you 0:02:21.680 --> 0:02:24.959 very often can. Uh So, yeah, we we should already 0:02:24.960 --> 0:02:27.880 know that there is very little privacy in the modern 0:02:27.919 --> 0:02:32.320 technology sphere. Our phones, our social media accounts are advertiser 0:02:32.360 --> 0:02:34.320 i d s, which are used to track us across 0:02:34.360 --> 0:02:38.280 the Internet. These sources are already used to create profiles 0:02:38.320 --> 0:02:42.400 in which the disparate types of our personal identifying information 0:02:42.720 --> 0:02:45.400 get correlated with each other and used to serve us 0:02:45.440 --> 0:02:49.600 ads or manipulate us on social media. But the leap 0:02:49.720 --> 0:02:54.320 into physical space, where all of our information is easily 0:02:54.400 --> 0:02:57.880 linked now to our physical body wherever we are, whether 0:02:57.919 --> 0:03:01.400 we like it or not, is the frontier that's currently 0:03:01.440 --> 0:03:05.560 being pushed and at a very rapid pace. Now there 0:03:05.600 --> 0:03:08.160 are multiple ways to make this link, of course, you know, 0:03:08.280 --> 0:03:11.560 so linking our digital profiles and all the the associated 0:03:11.639 --> 0:03:14.560 data with our physical bodies. A very simple one would 0:03:14.560 --> 0:03:17.200 be with the tracking devices that pretty much all of 0:03:17.280 --> 0:03:20.280 us carry with us at all times, the mac addresses 0:03:20.280 --> 0:03:23.120 on our phones and our mobile devices. But one of 0:03:23.160 --> 0:03:27.080 the most powerful developing techniques is for the technosphere to 0:03:27.200 --> 0:03:31.320 recognize you in physical space the same way that your 0:03:31.320 --> 0:03:34.720 friends and your family do by your face. Now, one 0:03:34.720 --> 0:03:36.880 thing to keep in mind about this is that, of course, 0:03:37.000 --> 0:03:40.520 just because say you're thermostat in your house can recognize 0:03:40.520 --> 0:03:44.200 your face, that is not necessarily in and of itself 0:03:44.240 --> 0:03:47.560 a bad thing. In some cases that could be very helpful, 0:03:47.680 --> 0:03:49.960 or even it could be seen as a way to uh, 0:03:50.360 --> 0:03:54.680 you know, to safeguard the temperature of your home, uh, 0:03:54.720 --> 0:03:56.120 that sort of thing, like you could have some sort 0:03:56.160 --> 0:03:59.800 of a security feature. And as we as we proceed 0:04:00.040 --> 0:04:02.280 through these episodes, we're going to try and keep that 0:04:02.320 --> 0:04:05.080 in mind. We're gonna we're gonna try not to color 0:04:05.120 --> 0:04:11.960 the technology as inherently vile or inherently uh prone to misuse. 0:04:12.360 --> 0:04:15.680 But the story of technology is that it is both 0:04:16.200 --> 0:04:19.040 light and dark. Well, yeah, and that I think there 0:04:19.080 --> 0:04:22.760 is a big difference between something being inherently vile versus 0:04:22.839 --> 0:04:25.359 prone to misuse. There are a lot of things that 0:04:25.440 --> 0:04:29.440 are that are created with perfectly good intentions in mind. 0:04:29.839 --> 0:04:31.960 You can think of lots of great reasons to do 0:04:32.080 --> 0:04:35.840 most of the worst technological stuff imaginable. You know, we've 0:04:35.880 --> 0:04:38.400 talked before on the show. We make no secret our 0:04:38.800 --> 0:04:41.120 distaste for a lot of things about social media. But 0:04:41.160 --> 0:04:43.200 of course, you know, you can see the good side 0:04:43.240 --> 0:04:46.200 of something like Facebook, Yeah, or very broadly, you can 0:04:46.240 --> 0:04:51.080 think of something like banking. Banking in many respects allows 0:04:51.200 --> 0:04:55.040 humans to do amazing things things they wouldn't or otherwise 0:04:55.040 --> 0:04:59.160 be able to do. Uh, to buy larger uh, you know, 0:04:59.440 --> 0:05:01.599 to buy a property, you know, to buy a vehicle, 0:05:01.760 --> 0:05:04.039 to start a business, start a business, and so forth. 0:05:04.080 --> 0:05:06.880 But at the same time, very terrible things have been 0:05:06.920 --> 0:05:10.960 done and are still being done under the broad tent 0:05:11.120 --> 0:05:14.680 of banking. Yeah, and we can help better protect ourselves 0:05:14.720 --> 0:05:17.680 from those outcomes by better understanding banking so that we 0:05:17.720 --> 0:05:20.839 can regulate it properly, which doesn't always get done, but 0:05:20.920 --> 0:05:23.200 you know that that, at least in theory, that is 0:05:23.240 --> 0:05:25.839 the way to protect yourself. And so today we're gonna 0:05:25.880 --> 0:05:28.240 be taking that same point of view to a series 0:05:28.240 --> 0:05:33.320 of episodes about facial recognition science and technology. And this 0:05:33.400 --> 0:05:35.240 is a subject I've wanted to talk about for a 0:05:35.240 --> 0:05:40.159 while because obviously this is an issue of increasing significance today. 0:05:40.720 --> 0:05:44.000 But actually, just after we landed on this topic, I 0:05:44.080 --> 0:05:46.640 came across a brand new story in The New York 0:05:46.680 --> 0:05:50.240 Times that serves as a really good anchor on why 0:05:50.279 --> 0:05:54.240 this issue is incredibly relevant today. And this article is 0:05:54.240 --> 0:05:57.640 called The Secretive Company that Might end Privacy as we 0:05:57.720 --> 0:06:00.400 know it by cashmir Hill, published in the New York 0:06:00.440 --> 0:06:06.120 Times January. Yeah. It uh. It concerns a small technology 0:06:06.160 --> 0:06:10.360 company called clear View AI, which just the title of 0:06:10.440 --> 0:06:12.920 that it sounds like it could be fine, right, sounds 0:06:13.040 --> 0:06:17.120 very transparent, doesn't clear View AI, But of course this 0:06:17.240 --> 0:06:22.080 is a facial recognition based artificial intelligence company. Uh So, 0:06:22.080 --> 0:06:24.440 so what is clear View in their own words? What 0:06:24.480 --> 0:06:27.680 do they say about themselves? Uh? To read from their 0:06:27.720 --> 0:06:31.719 marketing materials quote. Clear View is a new research tool 0:06:31.920 --> 0:06:36.480 used by law enforcement agencies to identify perpetrators and victims 0:06:36.480 --> 0:06:40.360 of crimes. Clear Views technology has helped law enforcement track 0:06:40.400 --> 0:06:44.760 down hundreds of at large criminals, including pedophiles, terrorists, and 0:06:44.839 --> 0:06:48.159 sex traffickers. It is also used to help exonerate the 0:06:48.200 --> 0:06:51.880 innocent and identify the victims of crimes, including child sex 0:06:51.880 --> 0:06:55.000 abuse and financial fraud. Now, on the surface of things, 0:06:55.040 --> 0:06:58.880 that sounds absolutely air tight, right. It describes the technology 0:06:58.920 --> 0:07:03.400 that is used by the appropriate agencies to protect the 0:07:03.480 --> 0:07:08.360 innocent and to go after the guilty. But then again, 0:07:09.080 --> 0:07:11.200 that can be used to sell a lot of things 0:07:11.680 --> 0:07:15.160 in the world, of course. So what they advertise is 0:07:15.200 --> 0:07:19.520 that this app helps law enforcement identify perpetrators and protect 0:07:19.640 --> 0:07:22.160 victims of crime, and of course, in some cases that 0:07:22.240 --> 0:07:25.040 may very well be true. Obviously, it would be pointless 0:07:25.120 --> 0:07:29.120 to deny that facial recognition technology the ability to take 0:07:29.120 --> 0:07:32.160 a picture of somebody's face and then find out tons 0:07:32.160 --> 0:07:34.160 of stuff about who they are and how you can 0:07:34.200 --> 0:07:36.760 find them. You know, be pointless to deny that in 0:07:36.800 --> 0:07:40.000 many cases that would be useful and beneficial to law enforcement. 0:07:40.400 --> 0:07:42.800 But it is also, of course just so easy to 0:07:42.840 --> 0:07:45.360 see how a tool like this could be terrible, both 0:07:45.400 --> 0:07:48.760 in its successes and in its failures. So, first of all, 0:07:48.760 --> 0:07:51.800 of course, it could fail in catastrophic ways, say with 0:07:51.840 --> 0:07:54.760 like false matches when police are looking for a perpetrator, 0:07:55.520 --> 0:07:57.280 but of course you know that's something that can happen 0:07:57.320 --> 0:08:01.080 with human witnesses too, right, But then it could also 0:08:01.280 --> 0:08:05.400 be used effectively if it correctly identifies people too amazingly 0:08:05.480 --> 0:08:09.480 insidious ends. So how does it work. It's actually it 0:08:09.560 --> 0:08:13.000 sounds pretty simple in terms of its user interface. Specifically, 0:08:13.000 --> 0:08:16.440 what this tool does is it matches an input photo 0:08:16.520 --> 0:08:21.200 of a face with a huge database of existing photos 0:08:21.240 --> 0:08:24.840 scraped from the internet, and then it will provide links 0:08:24.880 --> 0:08:28.080 to the places that those images from the internet were 0:08:28.160 --> 0:08:32.080 originally found. So very simple example, I take a photo 0:08:32.080 --> 0:08:34.440 of you, and then I feed it into the tool, 0:08:34.800 --> 0:08:37.160 and then it comes back with other photos of you 0:08:37.800 --> 0:08:40.560 and links to the places where those photos were found, 0:08:40.640 --> 0:08:43.920 maybe your Facebook page, a YouTube video, you're in and 0:08:43.960 --> 0:08:47.080 so forth, and so when it works, it will provide 0:08:47.120 --> 0:08:51.400 a direct link between your anonymous face from a picture 0:08:51.960 --> 0:08:55.559 and the digital locations where all of your personal information 0:08:55.600 --> 0:08:58.720 may be logged online. Now, we're not going to look 0:08:58.760 --> 0:09:01.760 too far under the hood of exactly how the you know, 0:09:01.840 --> 0:09:06.199 the underlying technology of this sort of thing works right now. Uh, 0:09:06.440 --> 0:09:09.280 there are a number of ways that facial recognition algorithms 0:09:09.320 --> 0:09:11.880 can actually work, but a very common way is using 0:09:11.920 --> 0:09:15.840 neural networks. Yes, and uh and and for this For 0:09:15.840 --> 0:09:19.480 for like a nice succinct description of how this works, 0:09:20.040 --> 0:09:23.160 I'd like to refer to Max tag marks most recent 0:09:23.160 --> 0:09:26.040 book Life three point oh, which deals at length with 0:09:26.280 --> 0:09:31.000 AI and the potential the threats posed by by AI 0:09:31.040 --> 0:09:33.839 in the future. It's a it's a wonderful book. But 0:09:34.160 --> 0:09:38.560 in this this one section, he's just summarizing how this 0:09:38.640 --> 0:09:41.760 kind of facial recognition works. Uh. And he writes, uh 0:09:41.840 --> 0:09:45.600 that neural networks have been quote trained to input numbers 0:09:45.640 --> 0:09:50.200 representing the probability that the image depicts various people. Here, 0:09:50.360 --> 0:09:54.480 each artificial neuron and on on an illustration they're depicted 0:09:54.559 --> 0:09:58.360 as circles. Computes a weighted sum of the number sent 0:09:58.440 --> 0:10:02.880 into it via connections or lines in this image from above, 0:10:02.960 --> 0:10:06.720 applies a simple function and passes the results downward each 0:10:06.760 --> 0:10:11.760 subsequent layer of computing. Higher level features. Typical face recognition 0:10:11.800 --> 0:10:15.640 network contain hundreds of thousands of urons. Uh. The figure 0:10:15.640 --> 0:10:19.240 shows merely a handful for clarity, so uh. In the 0:10:19.600 --> 0:10:22.960 the visual representation the tech mark includes here you see 0:10:23.200 --> 0:10:26.560 all these circles and interconnected lines representing how the you know, 0:10:26.559 --> 0:10:30.480 the neural network is functioning. But then it begins, then 0:10:30.520 --> 0:10:33.240 we apply this to the facial features. It starts with 0:10:33.240 --> 0:10:35.959 with sort of general features and sort of blurry shapes, 0:10:36.040 --> 0:10:39.040 and then to more specific features, and then tying those 0:10:39.040 --> 0:10:43.120 features together and then eventually getting to an output probability 0:10:43.400 --> 0:10:47.079 of actual facial matches. Yeah. And as with many other 0:10:47.160 --> 0:10:50.000 neural networks that are trained on large data sets, to 0:10:50.280 --> 0:10:54.200 you know, match values together or produce an output, why 0:10:54.360 --> 0:10:58.400 given input x uh you know, given a training method, 0:10:59.160 --> 0:11:01.880 there will often so. The way these things are trained 0:11:02.040 --> 0:11:03.840 is that you you know, you feed them a lot 0:11:03.880 --> 0:11:06.959 of examples of the kind of output you want, and 0:11:07.040 --> 0:11:11.000 slowly they refine their own rules internally. The rules that 0:11:11.040 --> 0:11:14.240 happened at each of these layers of neurons to manipulate 0:11:14.320 --> 0:11:17.240 numbers and values as they pass through the neural network 0:11:17.559 --> 0:11:20.560 in order to give the output that closely matches whatever 0:11:20.600 --> 0:11:23.360 you've trained it to come up with. But that that 0:11:23.440 --> 0:11:27.680 means that like you can train potentially an effective neural 0:11:27.720 --> 0:11:32.320 network without yourself really understanding very well exactly what's happening 0:11:32.320 --> 0:11:35.480 at each layer throughout throughout the network. Now, I think 0:11:35.480 --> 0:11:37.960 it is possible to like sort of get in there 0:11:38.000 --> 0:11:40.280 and try to dig into it and and see what's 0:11:40.280 --> 0:11:42.679 going on, if you've really got the time and expertise. 0:11:43.559 --> 0:11:47.440 But but it can be relatively opaque as far as 0:11:47.440 --> 0:11:51.320 computer programs go. It doesn't necessarily work like a normal 0:11:51.360 --> 0:11:54.679 computer program has lines of code that any programmer who 0:11:54.720 --> 0:11:57.040 knows the language can read through and figure out what's 0:11:57.040 --> 0:11:59.280 going on easily. But to come back to the cash 0:11:59.320 --> 0:12:03.400 Mirror Hill article and UH clear view AI uh. One 0:12:03.440 --> 0:12:06.520 thing that's important to point out is that this is 0:12:06.600 --> 0:12:10.839 by no means the first facial recognition app or tool, 0:12:11.000 --> 0:12:14.200 nor is it the first used by law enforcement. It's 0:12:14.240 --> 0:12:17.840 particular value that the thing that it's doing that's somewhat 0:12:17.880 --> 0:12:21.920 new is in its database of images, which again have 0:12:22.000 --> 0:12:27.040 been scraped from organic sources like Facebook and YouTube. Previously, 0:12:27.120 --> 0:12:31.720 law enforcement facial recognition matching programs were often weaker and 0:12:31.760 --> 0:12:36.400 more limited to smaller databases of government photos, say mug 0:12:36.440 --> 0:12:40.560 shots or driver's licenses. And of course there's the potential that, 0:12:40.640 --> 0:12:44.400 you know, smaller training or matching material will make any 0:12:44.440 --> 0:12:47.320 machine learning process weaker at at coming up with the 0:12:47.360 --> 0:12:50.800 results you want. Yeah. All this kind of ties in 0:12:50.840 --> 0:12:52.200 with what I often think of its kind of like 0:12:52.240 --> 0:12:55.800 this title pool illusion of the Internet, that feeling that 0:12:55.920 --> 0:12:58.600 a lot of us had I think still have sometimes 0:12:58.600 --> 0:13:01.320 but also especially early on, this feeling that we were 0:13:01.360 --> 0:13:04.920 engaging in something segmented from the general population. You know. 0:13:05.440 --> 0:13:07.560 But the thing about tide pools, of course, is that 0:13:07.640 --> 0:13:09.959 eventually the tide rolls in and then you realize that 0:13:10.400 --> 0:13:14.280 you're actually connected to the wider internet. Uh so you know, 0:13:14.360 --> 0:13:17.760 not just your friends or your family or your fandom, 0:13:17.840 --> 0:13:22.920 but also uh, you know, law enforcement, criminals, politics, all 0:13:23.600 --> 0:13:27.480 just churning around in the same grim ocean of numbing obscenity. 0:13:28.080 --> 0:13:30.640 I think that's a really excellent metaphor. Yeah, there is 0:13:30.960 --> 0:13:35.640 some somehow the Internet was very easily able to create 0:13:35.800 --> 0:13:40.280 a sense of isolated, walled off gardens that we were 0:13:40.280 --> 0:13:44.360 living in which were at the time totally public you know, um, 0:13:44.520 --> 0:13:48.280 early days of various social networking sites, fan forums, all 0:13:48.280 --> 0:13:50.280 that kind of thing, you know, whatever it was that 0:13:50.400 --> 0:13:52.920 gave people a sense that they were in a little 0:13:52.960 --> 0:13:57.079 private space, you know, that their little corner, their little room. 0:13:57.120 --> 0:13:59.800 But of course it's the Internet. What's happening there is public, 0:13:59.840 --> 0:14:02.800 and the consumers of what's happening there may be completely 0:14:02.840 --> 0:14:06.280 invisible to you. Right. So in this case, with the 0:14:06.760 --> 0:14:10.800 previous models of facial identification, the data sets they were 0:14:10.800 --> 0:14:14.400 depending on, we're basically title pools, like here's the title 0:14:14.400 --> 0:14:17.400 pool of of mug shots. Here's the driver's license, uh, 0:14:17.440 --> 0:14:19.800 title pool, and that's what we're feeding on. But basically 0:14:19.920 --> 0:14:22.160 clear View comes around, and this is a company that 0:14:22.320 --> 0:14:24.760 is saying, well, let's just use the whole ocean. What's 0:14:24.760 --> 0:14:27.240 stopping is from using the whole ocean. So this is 0:14:27.240 --> 0:14:30.000 a company using the assets of various social media and 0:14:30.040 --> 0:14:32.960 in general visual media companies on the web to do 0:14:33.000 --> 0:14:35.920 the sorts of things that those companies have been loath 0:14:36.040 --> 0:14:38.160 to do, or if at least been you know, publicly 0:14:38.200 --> 0:14:41.360 opposed to doing. Because technically, as as pointed out in 0:14:41.440 --> 0:14:44.080 Hills New York Times, article. Um, you know, there is 0:14:44.120 --> 0:14:46.440 there is an argument that what they're doing here, what 0:14:46.440 --> 0:14:49.000 what this company, in any company that's that's engaging in 0:14:49.000 --> 0:14:51.280 this kind of like broad sampling, that they may be 0:14:51.840 --> 0:14:55.200 violating the terms of service for these various websites. Sure, yeah, 0:14:55.240 --> 0:14:59.040 automatically scraping imagery and data from Facebook. Say, I think 0:14:59.640 --> 0:15:01.960 there was at least the allegation that that could be 0:15:02.440 --> 0:15:05.840 a violation of Facebook's terms of service, but it didn't 0:15:05.880 --> 0:15:08.600 really seem to bother the clear view people right right, 0:15:08.640 --> 0:15:11.640 And I think when Hill reached out to Facebook representative, 0:15:11.640 --> 0:15:13.480 they said, well, we may look into that, and so 0:15:13.560 --> 0:15:17.160 it's kind of an open question. But you know, a 0:15:17.200 --> 0:15:19.280 lot of this also comes down to something we discussed 0:15:19.320 --> 0:15:23.280 in our our look at during uh Lanier's ten arguments 0:15:23.320 --> 0:15:27.040 for deleting your social media accounts right now, because it 0:15:27.080 --> 0:15:30.840 concerns your data, data that you have, in all likelihood 0:15:31.120 --> 0:15:34.680 given away to companies like Facebook, Twitter and others simply 0:15:35.000 --> 0:15:39.120 to be a part of the interconnectedness that they sold us. Now, 0:15:39.160 --> 0:15:41.400 a lot of the time when we're discussing such a data, 0:15:41.440 --> 0:15:45.080 we're discussing behavioral information, right your likes, your dislikes, your 0:15:45.080 --> 0:15:50.160 arousal patterns concerning posts and advertisements. But in addition to this, 0:15:50.240 --> 0:15:53.960 you also sold the devil your face. Yeah, I mean 0:15:54.040 --> 0:15:56.760 he he's the devil in this case is simply promising 0:15:56.800 --> 0:16:00.920 not to do anything unbecoming with your face. But but 0:16:01.040 --> 0:16:04.440 Hell is highly populated. Even if you have good reason 0:16:04.520 --> 0:16:07.960 to trust this particular devil to which you've already entrusted 0:16:08.000 --> 0:16:10.600 your face and perhaps the faces of your family members, 0:16:10.600 --> 0:16:14.920 your loved ones, deceased loved ones, your children. Uh, you know, 0:16:14.960 --> 0:16:18.960 there are countless others that they will make no such promises, 0:16:19.280 --> 0:16:21.640 steal your face right off your head. Yeah, and and 0:16:21.680 --> 0:16:24.120 they may have little concern for, you know, the values 0:16:24.160 --> 0:16:27.240 that were in place during the initial purchase. It's something 0:16:27.240 --> 0:16:29.200 that I think about coming up again and again with 0:16:29.520 --> 0:16:31.960 sharing data on the internet. So you share your data 0:16:32.000 --> 0:16:36.360 with a company, and you maybe trust that company today 0:16:36.400 --> 0:16:40.120 to protect your data. But what if so say that 0:16:40.200 --> 0:16:42.320 company hangs onto your data for a while and then 0:16:42.360 --> 0:16:44.880 they get new management that you don't trust as much, 0:16:44.920 --> 0:16:47.680 but they've got it, Uh, you can't get it back. 0:16:48.160 --> 0:16:50.760 Or maybe they have a security breach and somebody just 0:16:50.800 --> 0:16:54.240 happens to steal your data from them. It's like, well, 0:16:54.360 --> 0:16:56.920 you would have trusted the company maybe, but now somebody 0:16:56.920 --> 0:16:59.120 else has got it, and you can see how in 0:16:59.120 --> 0:17:02.960 a world like that, Uh, it could start to feel 0:17:03.200 --> 0:17:08.040 like maybe hopeless or pointless or futile for people who 0:17:08.080 --> 0:17:10.320 say are in a position to make money off of 0:17:10.359 --> 0:17:14.320 not being very careful about people's privacy. Uh. You know, 0:17:14.400 --> 0:17:17.800 it's like, you know, what's the point. Everything eventually gets 0:17:17.800 --> 0:17:20.640 out there anyway. And this kind of point of view 0:17:20.920 --> 0:17:23.400 was sort of articulated by some of the people quoted 0:17:23.520 --> 0:17:26.440 in Hill's article. For example, there's a figure named David 0:17:26.520 --> 0:17:30.240 Scalzo who was an early investor and investor in this company, 0:17:30.240 --> 0:17:33.439 Clear View AI uh, and Scalzo is quoted in the 0:17:33.560 --> 0:17:37.280 article saying, I've come to the conclusion that because information 0:17:37.400 --> 0:17:42.040 constantly increases, there's never going to be privacy. Laws have 0:17:42.160 --> 0:17:46.040 to determine what's legal, but you can't ban technology. Sure 0:17:46.200 --> 0:17:49.040 that might lead to a dystopian future or something, but 0:17:49.160 --> 0:17:53.840 you can't ban it. So, you know, this is it 0:17:53.960 --> 0:17:56.919 at once one of those statements that seems very pragmatic 0:17:56.960 --> 0:17:59.639 but also entirely self serving, because true the story of 0:17:59.640 --> 0:18:02.760 technolo alogy is that its advanced cannot really be stopped. 0:18:02.800 --> 0:18:04.959 You have to think ahead as best we can and 0:18:05.000 --> 0:18:07.600 prepare our laws and our moral code to dealing with 0:18:07.680 --> 0:18:10.639 emerging technologies. We've talked about this before, for instance, in 0:18:10.880 --> 0:18:15.200 as far as genetic technology is concerned, But this particular 0:18:15.280 --> 0:18:17.480 quote also sounds a lot like, Hey, it's gonna happen 0:18:17.520 --> 0:18:19.560 either way, so I might as well be the person 0:18:19.600 --> 0:18:21.879 to make some money off of it. It totally agree, 0:18:21.920 --> 0:18:24.919 I mean, I agree that it is difficult to stop 0:18:25.000 --> 0:18:28.199 technological progress. That if you know, if one group of 0:18:28.200 --> 0:18:30.800 people isn't working on it, maybe a less ethical group 0:18:30.800 --> 0:18:33.640 of people might be somewhere, But that's not an excuse 0:18:33.720 --> 0:18:37.240 to be the person who creates the synthetic supervirus that 0:18:37.359 --> 0:18:42.679 you know who like genetically engineers, captain trips flu or whatever. Like. Also, 0:18:42.840 --> 0:18:46.000 you could use this logic about almost any bad thing. 0:18:46.600 --> 0:18:48.600 It's kind of like saying, yeah, you know, there's no 0:18:48.640 --> 0:18:51.840 way to totally eliminate pollution. Some people are always going 0:18:51.880 --> 0:18:54.040 to find a way to pollute and harm the environment, 0:18:54.119 --> 0:18:56.159 so you might as well just go hog while just 0:18:56.240 --> 0:18:59.359 dump it all. Like. So, it's true that you can't 0:18:59.440 --> 0:19:03.480 stop everything harmful to the environment with regulation, but you 0:19:03.520 --> 0:19:06.760 can really slow it down. You can present major obstacles 0:19:06.760 --> 0:19:09.680 to the worst types of offenses. And likewise, I think 0:19:09.680 --> 0:19:13.240 it would be very difficult to completely stop the advancing 0:19:13.280 --> 0:19:17.720 capabilities of AI, including facial recognition, but you can certainly 0:19:17.840 --> 0:19:21.240 slow it. You can certainly limit its potentially harmful uses 0:19:21.359 --> 0:19:24.520 by banning those uses and punishing offenders. Now, on the 0:19:24.560 --> 0:19:27.160 other hand, you could think, well, yeah, you could do that, 0:19:27.200 --> 0:19:29.840 but this would be so helpful to law enforcement in 0:19:29.880 --> 0:19:33.480 some cases. You know, so would the ability to search 0:19:33.560 --> 0:19:36.239 any house you wanted without a warrant, right right, I mean, 0:19:36.280 --> 0:19:40.000 this is the same argument that has often been part 0:19:40.320 --> 0:19:46.000 of the reasoning for enhanced interrogation and torture, is that, well, 0:19:46.040 --> 0:19:48.240 it can help us get the bad guys, it can 0:19:48.280 --> 0:19:51.199 help us in this situation. And then also in all 0:19:51.240 --> 0:19:54.080 of these arguments are is also the the idea that well, 0:19:54.119 --> 0:19:56.640 if you have nothing to hide, if you are truly 0:19:56.720 --> 0:20:01.080 a good and supportive member of society, then what if 0:20:01.119 --> 0:20:02.840 you have to what do you have to worry about? Anyway? 0:20:03.440 --> 0:20:06.000 But but it kind of comes down to the data issue. Well, 0:20:06.080 --> 0:20:08.159 you trust the person who has your data now, but 0:20:08.240 --> 0:20:10.080 do you trust the person who have your data tomorrow? 0:20:10.119 --> 0:20:15.320 You trust the government of today, but governments change, yeah, 0:20:15.320 --> 0:20:18.679 I mean, and No, nobody actually in practice believes this 0:20:18.800 --> 0:20:20.920 what do you have to hide argument? It's just something 0:20:20.960 --> 0:20:23.280 you would say. I mean, like, if anybody ever says that, 0:20:23.400 --> 0:20:26.600 just immediately demand them to give you their email password. 0:20:27.000 --> 0:20:29.240 Me like, yeah, just let me read all your email. 0:20:29.240 --> 0:20:32.880 I mean, what's the problem, wild inspector? Right, I think 0:20:32.920 --> 0:20:36.560 you'll find everything in order. Uh. But yeah, so, I 0:20:36.600 --> 0:20:40.399 mean obviously, society has often decide to regulate police power 0:20:40.440 --> 0:20:43.800 in ways that I mean that are truly inconvenient to 0:20:43.880 --> 0:20:47.919 law enforcement, because they decide that in some cases there 0:20:47.960 --> 0:20:51.200 are types of privacy and other civil liberties that matter 0:20:51.359 --> 0:20:55.640 more than prosecuting offenses at the maximum efficiency. Yeah. Alright, Well, 0:20:55.640 --> 0:21:01.320 on that note, we're going to take a quick break. Okay, 0:21:01.359 --> 0:21:04.040 we're back. So again, we're in the middle of this 0:21:04.119 --> 0:21:07.880 first episode of this exploration we're doing of of facial 0:21:07.960 --> 0:21:11.439 recognition science and technology, and we've been talking about this 0:21:11.480 --> 0:21:13.920 New York Times article that just came out last week 0:21:14.040 --> 0:21:19.119 by Cashmere Hill about a facial recognition technology company called 0:21:19.200 --> 0:21:22.960 clear View AI. Now, broadly speaking, the two big advantages 0:21:23.440 --> 0:21:26.680 of clear View AI are that it first of all, 0:21:26.920 --> 0:21:30.280 pulls from an extensive database of images. So we're talking 0:21:30.400 --> 0:21:33.879 three billion photos in a database versus the four hundred 0:21:33.920 --> 0:21:37.879 and eleven million searchable through the FBI's database. These stats 0:21:37.920 --> 0:21:42.200 according to clear View marketing materials reviewed by Cashmere here 0:21:42.359 --> 0:21:44.960 Hill for that New York Times article. And then secondly, 0:21:45.240 --> 0:21:48.800 it boasts a robust enough facial recognition engine built up 0:21:49.040 --> 0:21:52.840 from academic work by others on artificial intelligence, image recognition 0:21:52.840 --> 0:21:56.520 and machine learning. UM that it and it and it 0:21:56.520 --> 0:22:00.440 it does not require high quality or complete facial images 0:22:00.480 --> 0:22:03.600 to produce matches. So like the you know, the I 0:22:03.600 --> 0:22:06.520 guess the ideal example would be you could have somebody 0:22:06.560 --> 0:22:09.639 going into a bank and robbing it with their face 0:22:09.760 --> 0:22:13.960 partially covered, and then this would potentially be able to 0:22:14.040 --> 0:22:17.120 match that partial face to a full face and say, 0:22:17.119 --> 0:22:19.720 a Facebook profile, at least according to what the company 0:22:19.760 --> 0:22:22.439 and some of its satisfied customers in law enforcement have 0:22:22.480 --> 0:22:26.280 been alleging. Yeah. Like one example of a successful match 0:22:26.359 --> 0:22:31.000 that Hill mentions h is um matching an individual to 0:22:31.160 --> 0:22:34.760 a face in a mirror in someone else's gym photo. 0:22:35.040 --> 0:22:38.439 What yeah, and uh, you know, details of the you know, 0:22:38.480 --> 0:22:41.600 presumed guilt or innocence of that particular individual aside, I 0:22:41.600 --> 0:22:44.600 think this is notable in that gyms are often considered 0:22:44.600 --> 0:22:47.960 to be photo taboo places, and it's certainly as far 0:22:48.000 --> 0:22:50.399 as like other people working out in the gym go. 0:22:50.640 --> 0:22:53.960 You know, this is I'm not, you know, an expert 0:22:54.000 --> 0:22:57.320 on jim etiquette, but it is my understanding that you 0:22:57.359 --> 0:23:00.400 shouldn't even you know, accidentally photographed someone else the gym. 0:23:00.640 --> 0:23:04.159 But obviously it does happen. Uh, it just it just 0:23:04.560 --> 0:23:06.720 you should not have your phone out snap and picks 0:23:06.720 --> 0:23:09.760 at the gym unless it's unless it's your private gym 0:23:09.800 --> 0:23:12.159 and you're the only person there, yeah, or unless it 0:23:12.160 --> 0:23:16.600 catches bad guys, because what do you have to hide. Um. 0:23:16.680 --> 0:23:19.399 That being said, the people at the company do admit 0:23:19.480 --> 0:23:21.960 that of course, like you know it, it still has flaws. 0:23:22.000 --> 0:23:24.800 There's still things that can't do. Yeah, Like, for instance, 0:23:25.000 --> 0:23:28.080 a lot of it is leaning on eye level photos, 0:23:28.080 --> 0:23:29.919 the kind of photos that you see and say a 0:23:29.960 --> 0:23:32.800 linked in profile photo, as opposed to the sort of 0:23:32.840 --> 0:23:36.520 ceiling level security camera footage that it is often involved 0:23:36.520 --> 0:23:39.720 in these scenarios, right right, uh yeah, I mean, And 0:23:39.800 --> 0:23:41.800 it's so it's running into the same kind of problems 0:23:41.800 --> 0:23:43.720 that we could talk about this later in the episode 0:23:43.960 --> 0:23:47.359 that human beings sometimes have with less familiar faces. I mean, 0:23:47.640 --> 0:23:51.760 this is a A known thing about human face perception 0:23:51.840 --> 0:23:55.439 and facial recognition within the brain is that, uh, we 0:23:55.520 --> 0:23:59.919 are much better at recognizing very familiar faces under unfair 0:24:00.040 --> 0:24:02.879 ruble conditions, like a partial face, a face it a 0:24:02.920 --> 0:24:06.040 weird angle, facing bad lighting. We can do that a 0:24:06.040 --> 0:24:08.400 lot better if it's a familiar face. Then if it's 0:24:08.400 --> 0:24:11.320 a relatively unfamiliar face, right, I mean, even things like 0:24:11.920 --> 0:24:14.479 our face in a mirror versus our face in a photo. 0:24:14.800 --> 0:24:17.159 You know, things like that can be distorting or if 0:24:17.200 --> 0:24:20.640 you or or more more directly, I find it with 0:24:20.840 --> 0:24:24.239 someone else's face reflected in a mirror. I'm definitely not 0:24:24.320 --> 0:24:27.119 used to seeing that, and that'll throw me off sometimes. 0:24:27.320 --> 0:24:29.480 Do you ever do the inversion test? This is another 0:24:29.520 --> 0:24:31.879 weird quirk of of the brain, trying to see if 0:24:31.880 --> 0:24:36.399 you recognize photos of people's heads upside down. I know 0:24:36.480 --> 0:24:38.560 we've talked about that before in the podcast, but I 0:24:38.600 --> 0:24:41.160 haven't really put it to the test in my own life. 0:24:42.040 --> 0:24:44.400 There's a This is just a total side note. There's 0:24:44.400 --> 0:24:47.800 a very funny thing known as the Thatcher effect. It 0:24:47.840 --> 0:24:50.560 has to do with um, the fact that so if 0:24:50.600 --> 0:24:55.320 you look at somebody's head upside down, but with their 0:24:55.440 --> 0:24:58.280 eyes right side up. A lot of times people look 0:24:58.320 --> 0:25:00.399 at that and they don't even notice anything's wrong with 0:25:00.440 --> 0:25:03.359 the photo. So, like, the head is upside down, but 0:25:03.400 --> 0:25:06.080 the eyebrows are like over the eyes because the eyes 0:25:06.160 --> 0:25:09.879 are still in the correct orientation. Um. But then if 0:25:09.880 --> 0:25:11.879 you flip that whole thing to where the head is 0:25:11.960 --> 0:25:14.520 right side up but the eyes are upside down, it 0:25:14.600 --> 0:25:18.360 looks unbelievably grotesque, Like you will burst out, you'll make 0:25:18.440 --> 0:25:21.920 noise when you see it. Look it up. That's your effect. 0:25:22.119 --> 0:25:24.239 But anyway, back to the story about clear View. So 0:25:24.600 --> 0:25:28.320 the company claims its product finds matches for an input 0:25:28.320 --> 0:25:31.119 photo up to seventy percent of the time, and of 0:25:31.160 --> 0:25:33.280 course Hill notes in the article that we can't be 0:25:33.320 --> 0:25:37.840 sure how often false matches turn up. She quotes Claire Garvey, 0:25:37.960 --> 0:25:41.480 who is a researcher at Georgia University's Center on Privacy 0:25:41.480 --> 0:25:44.879 and Technology, who says, quote, we have no data to 0:25:44.960 --> 0:25:48.280 suggest this tool is accurate. The larger the database, the 0:25:48.359 --> 0:25:52.200 larger the risk of misidentification because of the Doppelganger effect. 0:25:52.520 --> 0:25:55.600 They're talking about a massive database of random people they've 0:25:55.600 --> 0:26:00.080 found on the Internet. The Doppelganger effect being not that 0:26:00.080 --> 0:26:05.440 that vengeful German spirits are are actually invading the database, 0:26:05.560 --> 0:26:09.000 but that there's just a going to be the larger 0:26:09.040 --> 0:26:11.360 the pool of people, the more people they're gonna look 0:26:11.640 --> 0:26:15.880 very much like like others. There's gonna be more similarity 0:26:15.920 --> 0:26:21.000 between an increasing a pool of individuals. Right. But at 0:26:21.040 --> 0:26:23.679 the same time, anecdotal reports from a number of law 0:26:23.800 --> 0:26:27.080 enforcement officers have claimed that this tool was effective at 0:26:27.119 --> 0:26:30.919 identifying real perpetrators from photos alone, and there have been 0:26:30.920 --> 0:26:34.719 plenty of other examples in recent years of supposedly effective 0:26:34.800 --> 0:26:38.239 facial recognition technology provided by other companies that have been 0:26:38.359 --> 0:26:43.480 used to UH to allegedly capture perpetrators of crimes done 0:26:43.520 --> 0:26:47.040 in public places in New York, in the UK, certainly 0:26:47.240 --> 0:26:50.240 you know, in in countries with a very strong surveillance state, 0:26:50.320 --> 0:26:52.360 like in China. And we can come back to more 0:26:52.400 --> 0:26:54.639 about that in later episodes, I think, But as a 0:26:54.640 --> 0:26:58.480 personal anecdote in this reported story, Hill at one point 0:26:58.520 --> 0:27:01.959 has the company's founder use the app on a picture 0:27:02.000 --> 0:27:05.200 of her, and she claims that the tool quote returned 0:27:05.280 --> 0:27:09.440 numerous results dating back a decade, including photos of myself 0:27:09.480 --> 0:27:12.720 that I had never seen before. When I used my 0:27:12.800 --> 0:27:15.200 hand to cover my nose, and the bottom of my face. 0:27:15.280 --> 0:27:19.000 The app still returned seven correct matches for me. So 0:27:19.400 --> 0:27:22.960 I think we can assume that failures, including both false 0:27:22.960 --> 0:27:26.800 and negatives and false positives are surely occurring at some rate. 0:27:27.240 --> 0:27:29.760 But it's also clear that this thing at least works 0:27:29.840 --> 0:27:32.760 some of the time. Yeah, and that's enough to help 0:27:32.760 --> 0:27:35.920 it get picked up by law enforcement. Also, it helps 0:27:35.920 --> 0:27:37.520 that there was a it seems like there's a pretty 0:27:37.520 --> 0:27:42.840 sizeable outreach campaign from the company to to to market 0:27:43.080 --> 0:27:46.120 the technology to law enforcement. Yes, and and we should say, 0:27:46.119 --> 0:27:48.000 I mean, we're not going to hash out everything they 0:27:48.040 --> 0:27:51.200 get into in the article, but the company has arrived 0:27:51.280 --> 0:27:54.720 at law enforcement as their primary customers. Before that, they 0:27:54.720 --> 0:27:57.000 tried to market it in all kinds of ways, including 0:27:57.400 --> 0:28:01.359 you know, for like personal use to like private security things, 0:28:01.440 --> 0:28:04.520 and for like commercial use and even then like political 0:28:04.520 --> 0:28:08.600 opposition research and stuff. Yeah, but this, uh, but you 0:28:08.600 --> 0:28:11.040 can definitely see the advantage to law enforcement here, because 0:28:11.200 --> 0:28:14.399 a detective, for instance, has their disposal a number a 0:28:14.440 --> 0:28:17.280 limited number of talents and tools that are useful and 0:28:17.320 --> 0:28:20.560 attempting to solve a case, and adding this to the 0:28:21.080 --> 0:28:23.760 to the toolkit is no brainer. Because you know, larger 0:28:23.760 --> 0:28:26.119 issues of stability of the platform aside, you know some 0:28:26.119 --> 0:28:30.520 of these legal issues we potential legal issues we're discussing earlier. Um, 0:28:30.560 --> 0:28:33.119 you know, this would be something you could use in 0:28:33.200 --> 0:28:36.360 congress with other techniques. You know, you could say, all right, 0:28:36.359 --> 0:28:39.560 this face seems to match up with this individual. We 0:28:39.600 --> 0:28:42.520 also know that this individual was in the correct you know, 0:28:42.600 --> 0:28:44.920 vicinity at the time. You know, you could lean on 0:28:44.960 --> 0:28:49.000 your other detective tools then to to actually make the case. 0:28:49.320 --> 0:28:51.480 That's not to say there's not potential for misuse here, 0:28:51.880 --> 0:28:54.560 but uh, I'm just saying you can. You can definitely 0:28:54.640 --> 0:28:58.000 see the appeal and how if everything is working perfectly, 0:28:58.360 --> 0:29:01.040 it would be an effective law and forcement tool. Yeah, 0:29:01.080 --> 0:29:03.920 and however effective it actually is, it's clear that this 0:29:04.000 --> 0:29:07.240 and similar tools are increasingly popular with law enforcement in 0:29:07.360 --> 0:29:09.880 countries all over the world. If you're one of those 0:29:09.880 --> 0:29:11.960 people who feels like pumping the brakes on this kind 0:29:11.960 --> 0:29:15.880 of technology, what could actually be done about it? Well, 0:29:15.960 --> 0:29:19.360 Hill quote somebody named al Ghadari, a privacy professor at 0:29:19.360 --> 0:29:23.240 Stanford Law School who says, bluntly quote, absent of very 0:29:23.280 --> 0:29:27.960 strong federal privacy law, we're all screwed. Uh. And He's 0:29:28.000 --> 0:29:31.280 not alone. There are plenty of privacy experts today advocating 0:29:31.280 --> 0:29:34.600 the point of view that facial recognition technology, or at 0:29:34.680 --> 0:29:37.720 least some specific uses of it, aren't just something that 0:29:37.800 --> 0:29:40.080 maybe we should be a little concerned about, there's something 0:29:40.120 --> 0:29:44.240 that needs to be banned outright. Um. For example, Hill 0:29:44.320 --> 0:29:47.640 also quotes somebody named Woodrow Hertzog, who is a professor 0:29:47.680 --> 0:29:51.959 of law and computer science at Northeastern University. Uh and 0:29:52.080 --> 0:29:55.400 hart Sog says, quote, we've relied on industry efforts to 0:29:55.560 --> 0:29:59.800 self police and not embrace such a risky technology, but 0:30:00.000 --> 0:30:02.920 now those damns are breaking because there's so much money 0:30:02.920 --> 0:30:05.520 on the table. I don't see a future where we 0:30:05.600 --> 0:30:10.240 harness the benefits of face recognition technology without the crippling 0:30:10.280 --> 0:30:13.200 abuse of the surveillance that comes with it. The only 0:30:13.240 --> 0:30:16.120 way to stop it is to ban it. So whether 0:30:16.200 --> 0:30:18.800 we should do that, or if so, what form that 0:30:18.880 --> 0:30:21.080 band should take, what whatever, is the best way to 0:30:21.120 --> 0:30:24.680 address it. I think it is at least clear that 0:30:24.760 --> 0:30:29.280 this is a very pressing and uh like time sensitive issue, 0:30:29.440 --> 0:30:33.760 that that is of urgent public concern right now. Yeah, because, again, 0:30:33.880 --> 0:30:35.520 as the author points out, I don't think any of 0:30:35.560 --> 0:30:37.400 us want to live in a world where any stranger 0:30:37.520 --> 0:30:40.080 can you know, surreptitiously take a photo of our face 0:30:40.160 --> 0:30:43.200 and then face searches and get all this data on us. 0:30:43.320 --> 0:30:45.239 You know, I don't want for us to build that 0:30:45.320 --> 0:30:47.880 kind of world for our children, who, more than any 0:30:47.960 --> 0:30:50.920 of us, never had a chance to opt out of 0:30:50.960 --> 0:30:53.520 this uh, this face trade, you know. And that's the 0:30:53.720 --> 0:30:55.680 that's the pressing to think about, you know, because because 0:30:55.680 --> 0:30:58.240 you're not thinking about that when you when you share 0:30:58.400 --> 0:31:02.000 images of your child on on Facebook or Instagram or 0:31:02.000 --> 0:31:04.600 whatever it happens to be. Um, you know, you you 0:31:04.720 --> 0:31:08.080 just you're wanting to celebrate that this person exists at all, 0:31:08.120 --> 0:31:10.800 but you're you're laying the groundwork from like you know, 0:31:11.720 --> 0:31:15.360 age zero Hola onward right, Uh, like this is their 0:31:15.480 --> 0:31:18.320 their digital history. We were talking before we came into 0:31:18.320 --> 0:31:20.680 the studio about like if a person wanted to do 0:31:20.760 --> 0:31:23.880 something about this, what could you do? You can't unpost 0:31:23.960 --> 0:31:26.920 photos of yourself and data that has already been scraped. 0:31:27.160 --> 0:31:29.640 But I wonder if maybe you could try to come 0:31:29.720 --> 0:31:34.080 up the works by constantly just polluting the Internet with 0:31:34.280 --> 0:31:37.240 false pictures of you that are not you. So you're 0:31:37.280 --> 0:31:41.000 like sort of like deep facing enough to where you um, 0:31:41.040 --> 0:31:44.720 you've obscured the visual record of yourself. Maybe yeah, like 0:31:44.840 --> 0:31:48.040 if you, I guess it would then depend on the 0:31:48.160 --> 0:31:51.360 all those new images being taken in and causing enough 0:31:51.400 --> 0:31:54.479 confusion and the identification of you. But I don't know, 0:31:55.520 --> 0:32:00.640 or perhaps like altering your facial appearance with enough regularity 0:32:00.720 --> 0:32:04.120 that there is no concise version of you, or or 0:32:04.160 --> 0:32:06.360 at least making it to wear. Then the the AI 0:32:06.440 --> 0:32:08.320 would have to work a lot harder. It would have 0:32:08.360 --> 0:32:11.000 to have like a broader definition of what you look like, 0:32:11.080 --> 0:32:14.920 to the point that maybe it excuse with maybe it 0:32:15.000 --> 0:32:18.040 enhances the doppelgang or effect, like I'm thinking about you know, 0:32:18.040 --> 0:32:20.520 like you just you know, each day you inject a 0:32:20.520 --> 0:32:23.840 different portion of your face with collagen or something, or 0:32:23.840 --> 0:32:27.400 maybe not collagen, but maybe just saying lead, well, I mean, 0:32:27.440 --> 0:32:30.320 does this lead to if you, I mean, this sounds ridiculous, 0:32:30.360 --> 0:32:32.600 but does this lead to a future where everybody starts 0:32:32.640 --> 0:32:36.480 walking around with a broad array of interchanging masks? Yeah, 0:32:36.640 --> 0:32:40.040 and will and then you have an enhanced um laws 0:32:40.120 --> 0:32:43.320 against the wearing of masks. I mean, masks are outlat 0:32:43.400 --> 0:32:45.440 in a lot of places and and a lot of 0:32:45.480 --> 0:32:47.720 events for for a reason. Yeah, you're not supposed to 0:32:47.760 --> 0:32:52.120 drive a car wearing a mask. That's um, this does 0:32:52.240 --> 0:32:56.000 remind me. There's a there's an excellent show on Hulu, 0:32:56.080 --> 0:32:59.080 a titled Future Man that is a it's it's a 0:32:59.120 --> 0:33:01.920 it's a comedy. It's a satire with a lot of 0:33:01.960 --> 0:33:06.160 nostalgia for various you know, sci fi franchises. But there 0:33:06.240 --> 0:33:08.239 is a scene in one of the episodes where an 0:33:08.280 --> 0:33:12.600 individual knows that a facial recognition system is looking for him, 0:33:12.640 --> 0:33:15.120 so he gets himself beat up first so that his 0:33:15.200 --> 0:33:17.520 face is then all swelled and distorted and it and 0:33:17.560 --> 0:33:19.400 then it cannot make a match for him, and he's 0:33:19.400 --> 0:33:21.840 able to sneak past the guards. I don't think that's 0:33:21.880 --> 0:33:25.280 a sustainable strategy. It's not. And more to the point, yeah, 0:33:25.280 --> 0:33:28.160 we should not even we should not have to even 0:33:28.240 --> 0:33:32.080 entertain that possibility to to hold on to, uh, you know, 0:33:32.120 --> 0:33:34.920 our sense of privacy. Yeah. Now, the fact that we're 0:33:34.920 --> 0:33:36.720 talking about this story in the New York Times about 0:33:36.720 --> 0:33:39.960 clear View is it's just a result of timing. It's 0:33:40.040 --> 0:33:43.760 like this one specific company is not the entirety of 0:33:43.880 --> 0:33:47.280 the end of privacy problem, nor of the facial recognition 0:33:47.280 --> 0:33:51.480 technology landscape in particular, Another company could do the same thing. 0:33:51.560 --> 0:33:55.120 Other copycats I'm sure are already getting in there. Uh, 0:33:55.120 --> 0:33:58.360 it's just one high profile example of of the potential 0:33:58.480 --> 0:34:01.280 already being put to use that that's getting a lot 0:34:01.280 --> 0:34:03.600 of attention in the past couple of weeks. Yeah, in 0:34:03.680 --> 0:34:05.400 part two. You have to read the full article for 0:34:05.400 --> 0:34:07.480 the details. But it's also like their key individuals that 0:34:07.520 --> 0:34:10.799 are notable that are tied into its funding. Yes, there's that, 0:34:10.920 --> 0:34:13.600 and and of course there's the ominous way it ends, 0:34:13.640 --> 0:34:16.319 which is the idea that it will soon probably be 0:34:16.440 --> 0:34:18.719 rolled out not just to law enforcement, but to be 0:34:18.800 --> 0:34:21.759 a publicly usable app, you know, which I guess is 0:34:21.800 --> 0:34:23.800 the sort of the scenario we were describing at the 0:34:23.800 --> 0:34:27.360 beginning of the episode, just having a publicly available personal 0:34:27.400 --> 0:34:30.080 search bar tool. Yeah. Well, well, markin me down for 0:34:30.120 --> 0:34:33.040 being against that. Yes, alright, time to take a quick break, 0:34:33.080 --> 0:34:38.560 but we'll be right back with more than thank alright, 0:34:38.560 --> 0:34:41.080 we're back. So, of course, I also want to make 0:34:41.120 --> 0:34:46.000 the case that facial recognition technology is not necessarily always 0:34:46.080 --> 0:34:48.960 a dangerous or scary or ominous thing. I mean, I 0:34:49.040 --> 0:34:50.919 think there are some uses of it that one could 0:34:50.960 --> 0:34:54.680 quite easily find benevolent or even delightful. And one of 0:34:54.680 --> 0:34:57.600 the ways I was thinking about this was many of 0:34:57.640 --> 0:35:01.120 its developing uses in non human animals, not all because 0:35:01.160 --> 0:35:03.360 some of its uses in non human animals are also 0:35:03.600 --> 0:35:06.040 like kind of horrifying, but some of the ones that 0:35:06.120 --> 0:35:08.440 non human animals are pretty great. I was reading an 0:35:08.520 --> 0:35:12.880 article in New York Magazine from October called Here's a 0:35:12.920 --> 0:35:16.560 list of every animal humans currently monitor using facial recognition 0:35:16.600 --> 0:35:21.480 technology by Mac dagaron As A complete list is probably 0:35:21.520 --> 0:35:23.440 wildly out of date at this point because this was, 0:35:24.320 --> 0:35:27.479 but a few of the entries include things like there's 0:35:27.520 --> 0:35:31.960 a Norwegian fish farming company called Sermac Group as that 0:35:32.000 --> 0:35:36.560 commissioned a system for facial recognition of salmon which would 0:35:36.719 --> 0:35:40.080 use distinctive patterns of spots around the eyes, mouth, and 0:35:40.160 --> 0:35:46.040 gills of individual salmon to build individual digital medical records 0:35:46.080 --> 0:35:49.120 associated with each fish. And this would be for the 0:35:49.160 --> 0:35:53.280 purpose of fighting epidemics of parasitic sea lice primarily. Okay, 0:35:53.280 --> 0:35:56.759 so not merely just presenting it with with the dish 0:35:56.760 --> 0:35:59.399 when you ordered it a restaurant, but I will take 0:35:59.480 --> 0:36:03.600 the baked salmon and please include it's complete medical history. Yes, 0:36:03.840 --> 0:36:07.279 this salmon was named Jeffrey. Here's his Facebook profile, you 0:36:07.320 --> 0:36:10.879 know on on salmon Facebook. Oh, but that comes back. 0:36:11.360 --> 0:36:14.600 Of course, facial recognition technology is being deployed to keep 0:36:14.600 --> 0:36:17.480 individual track of all kinds of livestock like cows and 0:36:17.560 --> 0:36:21.120 chickens for maybe medical reasons or reasons having to do 0:36:21.200 --> 0:36:24.280 in the case of cows, reasons having to do with tracking, 0:36:24.360 --> 0:36:27.160 like periods of peak milk output and stuff like that. 0:36:27.680 --> 0:36:30.520 But there are also stories about conservation efforts to non 0:36:30.560 --> 0:36:35.799 invasively monitor wild populations of vulnerable animals by way of 0:36:35.840 --> 0:36:38.920 facial recognition, which if that works, that sounds awesome. Like 0:36:39.200 --> 0:36:43.120 I was looking at article in Scientific American that described 0:36:43.160 --> 0:36:47.000 efforts to use facial recognition to track wild lions through 0:36:47.040 --> 0:36:51.840 a platform called the Lion Identification Network of Collaborators or link. 0:36:52.400 --> 0:36:54.560 That's interesting. It reminds me of I want to say, 0:36:54.600 --> 0:36:57.160 like a decade ago, maybe a little a little further 0:36:57.200 --> 0:37:00.959 back in time. Uh, there was a piece I read 0:37:01.040 --> 0:37:05.279 about tracking whale sharks, and whale sharks all have you know, 0:37:05.320 --> 0:37:08.239 distinctive patterns on their you know that's sort of the 0:37:08.239 --> 0:37:11.560 top of their heads that area, and uh, you know, 0:37:11.600 --> 0:37:15.440 it doesn't mean anything, you know, much to humanize. But 0:37:15.960 --> 0:37:18.840 I think at the time they were utilizing NASA technology 0:37:18.840 --> 0:37:21.680 that was aimed at making sense of of the stars 0:37:22.320 --> 0:37:26.320 like astronomical um computation systems to then to make sense 0:37:26.360 --> 0:37:30.319 and sort of track um identify at any rate these 0:37:30.400 --> 0:37:33.000 various whale sharks. So this would but this sort of 0:37:33.000 --> 0:37:34.960 thing would be I think an even better method of 0:37:35.040 --> 0:37:37.400 doing that because because you're you're probably dealing with it 0:37:37.440 --> 0:37:39.480 with creatures in all these cases that uh, you know, 0:37:39.520 --> 0:37:43.120 there's a they definitely are not all identical. There are differences, 0:37:43.120 --> 0:37:44.839 but we just may not have the eye for it, 0:37:45.040 --> 0:37:48.279 whereas technology can be can be used to say the 0:37:48.320 --> 0:37:51.880 Sharper eye for chicken identity exactly, and you know, it 0:37:51.960 --> 0:37:55.440 has the advantage of not having to physically tag the 0:37:55.480 --> 0:37:58.279 animal in some way, which can be difficult to do 0:37:58.600 --> 0:38:01.000 or it can be harmful to the animal right or 0:38:01.040 --> 0:38:04.480 day are dangerous to the individuals doing the tagging. Yeah, 0:38:04.840 --> 0:38:07.760 um and so apparently so you've got this one with lions, 0:38:07.800 --> 0:38:10.000 the link project, But there are similar things that have 0:38:10.080 --> 0:38:14.200 been attempted with tigers, elephants, even whales. You mentioned whale sharks, 0:38:14.239 --> 0:38:17.360 but with like actual mammal whales, with a project that 0:38:17.480 --> 0:38:21.120 an article in the Atlantic called Facebook for whales. I 0:38:21.160 --> 0:38:23.520 hope it's not as addictive for whales as it is 0:38:23.520 --> 0:38:26.479 for humans. That was a joke, but you didn't laugh. 0:38:26.600 --> 0:38:33.360 That's okay. But similar technologies have been proposed and tested 0:38:33.400 --> 0:38:36.520 to help link people with lost pets, including cats and dogs. 0:38:36.560 --> 0:38:38.520 That seems like a great use of this. Yeah, like, 0:38:38.640 --> 0:38:42.440 you know, certainly, I'm on enough social media boards where 0:38:42.520 --> 0:38:45.839 their pet owners and occasionally pedicle missing. And then there's 0:38:45.880 --> 0:38:48.439 this whole back and forth where like like, oh my 0:38:48.440 --> 0:38:51.439 my orange cat is missing that it looks like this, 0:38:51.560 --> 0:38:53.239 and then somebody would be like, well does he look 0:38:53.280 --> 0:38:55.800 like this? And I think I saw him in the backyard, 0:38:55.840 --> 0:38:57.120 and someone else is like, oh, I think I saw 0:38:57.200 --> 0:39:00.880 him over here across town, and nobody can be for sure, right, 0:39:00.960 --> 0:39:03.640 because it's hard to get close to a stray cat 0:39:03.680 --> 0:39:06.440 in some cases, or or an escaped cat or a 0:39:06.480 --> 0:39:09.319 feral cat that is then misidentified as a lost cat. 0:39:10.080 --> 0:39:11.799 But if you had the ability, if you had some 0:39:11.840 --> 0:39:15.719 sort of app infrastructure where your your cats missing, fine, 0:39:15.760 --> 0:39:18.560 you upload them to this database and then when someone 0:39:18.640 --> 0:39:20.480 finds a cat, they just take a picture of it 0:39:20.600 --> 0:39:22.520 and it tells you if that cat is missing. Like 0:39:22.560 --> 0:39:24.319 that would be that would be great. That would cut 0:39:24.320 --> 0:39:26.560 out a lot of the the the anguish and the 0:39:26.920 --> 0:39:29.880 work that goes with having a runaway pet. I agree 0:39:30.080 --> 0:39:32.600 that sounds great, and maybe I'm suffering from a lack 0:39:32.640 --> 0:39:36.000 of imagination, but I'm I'm thinking in cases like that, 0:39:36.000 --> 0:39:39.040 that's a case where I think the the risks to 0:39:39.080 --> 0:39:42.359 the cats privacy would be far outweighed by the benefits 0:39:42.400 --> 0:39:45.440 of people finding their lost animals, because we know how 0:39:45.480 --> 0:39:47.640 cats are. They don't give a damn about privacy. Now 0:39:47.719 --> 0:39:49.759 they have, Yeah, they have a whole different set of 0:39:50.640 --> 0:39:55.440 set of values. Now. Another sort of quirk of timing 0:39:55.480 --> 0:39:58.520 here as we were putting this episode together, in fact, 0:39:58.520 --> 0:40:00.880 is we're sort of finishing our no for this episode. 0:40:01.040 --> 0:40:04.719 This morning, I actually read a new blog post, uh 0:40:05.200 --> 0:40:09.240 titled depth of Field Fails by Janelle Shane at ai 0:40:09.320 --> 0:40:12.279 weirdness dot Com. Oh, we've talked about her blog on 0:40:12.320 --> 0:40:14.799 the show before because it came up in the pair 0:40:14.800 --> 0:40:17.919 of episodes we did called flat a Sex Mackinah, which 0:40:17.960 --> 0:40:21.440 was about why it's so funny when machines fail. Yes, yeah, yeah, 0:40:21.560 --> 0:40:24.080 this is and I imagine a lot of you have encountered, uh, 0:40:24.360 --> 0:40:27.479 you know, the various scenarios she's run with, like AI 0:40:27.600 --> 0:40:30.960 s coming up with names for Halloween costumes or names 0:40:30.960 --> 0:40:33.520 for I think at one point evenel like She's they 0:40:33.520 --> 0:40:35.479 were using it to come up with not only names, 0:40:35.480 --> 0:40:40.439 but actual recipes for cocktails, yes, recipes for foods, also 0:40:40.840 --> 0:40:43.760 names for dishes. We talked about D and D. Character 0:40:43.840 --> 0:40:48.040 biography is generated, and and spells names for spells. Remember 0:40:48.280 --> 0:40:52.640 remember song of the Darn Daving Fire. Yeah, there's even 0:40:52.640 --> 0:40:56.480 one about cat names. But in this particular post, um 0:40:57.200 --> 0:41:00.839 Shane the research scientists. She tests out the facial recognition 0:41:00.920 --> 0:41:04.359 AI that is employed by Skype for its blur My 0:41:04.440 --> 0:41:08.640 Background for All Calls feature. Now, the curious thing about 0:41:08.640 --> 0:41:11.360 this is I've been using Skype. We've been using Skype 0:41:11.360 --> 0:41:14.239 here on the show for years for interviews, and I 0:41:14.239 --> 0:41:16.480 guess I just I don't dig in deeply enough or 0:41:16.520 --> 0:41:19.400 read emails because I didn't realize this was a feature 0:41:19.480 --> 0:41:22.720 at all until today. I think we always let somebody 0:41:22.760 --> 0:41:24.920 smarter than us figure out how to use Skype and 0:41:24.960 --> 0:41:27.279 there we'll just do the talking. But but it, but 0:41:27.360 --> 0:41:30.279 it totally makes sense as a feature because perhaps you 0:41:30.400 --> 0:41:33.080 do have an ideal environment set up for a call 0:41:33.080 --> 0:41:36.680 with a business friendly background behind you, but maybe you don't. 0:41:36.719 --> 0:41:39.160 Maybe you have a fridge with a bunch of notes 0:41:39.160 --> 0:41:41.239 stuck to it with magnets, and maybe some of those 0:41:41.239 --> 0:41:44.120 are like bills or you know, or they have some 0:41:44.239 --> 0:41:47.080 data on there that you you wouldn't even want there 0:41:47.120 --> 0:41:49.799 to be a chance somebody might be able to decipher it, 0:41:50.000 --> 0:41:52.399 or a bookshelf full of occult tones that you don't 0:41:52.400 --> 0:41:55.120 want people to know you've been researching that that's a 0:41:55.160 --> 0:41:57.920 good one. Or perhaps you're at work and there's a 0:41:57.960 --> 0:42:01.279 marker board full of you know, of data back there 0:42:01.320 --> 0:42:03.600 might be something you don't want out, Or perhaps you 0:42:03.640 --> 0:42:05.960 have some distracting art up there in the wall and 0:42:06.000 --> 0:42:08.919 you don't want to compromise the interior of your own 0:42:08.960 --> 0:42:11.840 home so that you can do a skype call. You 0:42:11.840 --> 0:42:13.920 don't want to have to like take things off the 0:42:13.920 --> 0:42:15.839 wall in order to do this, or you're in one 0:42:15.840 --> 0:42:21.000 of those weird Roman mass toilets, whatever the case may be. 0:42:21.520 --> 0:42:24.520 The AI then can auto blur all of that out 0:42:24.800 --> 0:42:27.360 for you, But to do so, it has to be 0:42:27.400 --> 0:42:29.560 able to tell the difference between the face of the 0:42:29.640 --> 0:42:33.520 collar and mere objects in the background. Uh. And it's 0:42:33.560 --> 0:42:35.880 in this case the AIS definition of a face is 0:42:35.920 --> 0:42:40.320 pretty broad. As Shane discovers, it will allow ancient Egyptian 0:42:40.360 --> 0:42:44.600 illustrations to to come through unblurred, to be the face, 0:42:44.880 --> 0:42:46.600 to be the face. So it's like, okay, this call 0:42:46.680 --> 0:42:51.680 is being made by this individual from ancient Egypt isis 0:42:52.320 --> 0:42:57.640 but also increasingly abstract depictions of a human face. Uh. 0:42:57.680 --> 0:43:00.080 So it showed various works of art and it's some 0:43:00.120 --> 0:43:02.080 of them were really abstract, and it was like, all right, 0:43:02.120 --> 0:43:04.320 that's a face. Sure that will work, just like monk 0:43:04.440 --> 0:43:10.000 the scream, Yes, yeah exactly. Um. Also stuffed giraffes. It 0:43:10.080 --> 0:43:14.000 had a problem with the like the horns of the draffe, 0:43:14.280 --> 0:43:17.640 but but not so much the face of the stuffed giraffe. Uh. 0:43:17.680 --> 0:43:22.320 It gets a little confused with life size plastic skeletons, however, 0:43:22.440 --> 0:43:25.560 and also cats can throw it off as well. But 0:43:26.320 --> 0:43:28.840 so do check out that that blog post. It's it's 0:43:28.880 --> 0:43:32.160 amusing and also insightful. But all this is certainly I 0:43:32.160 --> 0:43:35.399 think an example of facial recognition AI doing something that's 0:43:35.440 --> 0:43:39.719 not only helpful but could actually help you with your privacy. Now, 0:43:39.760 --> 0:43:41.839 one thing I think that would be different there is 0:43:41.880 --> 0:43:45.120 that that's facial recognition in the sense of recognizing a 0:43:45.440 --> 0:43:49.600 face as opposed to a background, versus recognizing whose face 0:43:49.680 --> 0:43:52.200 a picture is of right. But then again you could 0:43:52.200 --> 0:43:55.080 easily imagine like that being an upgrade or being a 0:43:55.080 --> 0:43:58.160 situation where if you had a more robust um facial 0:43:58.200 --> 0:44:01.920 recognition um AI that was then used on some sort 0:44:01.960 --> 0:44:04.560 of skypelike system, you might actually go ahead and have 0:44:04.600 --> 0:44:08.240 a feature where the caller's face was logged and therefore 0:44:08.400 --> 0:44:10.759 it would blur out any face that was not the 0:44:10.800 --> 0:44:14.520 authorized users face. So that way of you know, there 0:44:14.560 --> 0:44:17.320 are other employees walking by in the background getting coffee, 0:44:17.440 --> 0:44:19.279 they're not going to show up on your call. If 0:44:19.320 --> 0:44:21.880 you're you know, significant other walks by in the background, 0:44:21.960 --> 0:44:23.920 they're not going to show up on the call, etcetera. 0:44:24.000 --> 0:44:26.040 So even with this, I mean, when I'm in these 0:44:26.160 --> 0:44:29.200 kind of scenarios in my head, I'm always wondering if 0:44:29.280 --> 0:44:32.960 there are freaky applications that I'm just not being imaginative 0:44:33.080 --> 0:44:35.640 enough to to get to. Yet, We'll let me put 0:44:35.640 --> 0:44:38.880 on my black mirror neural lace cap for a second 0:44:39.200 --> 0:44:42.839 and think, um how about a simple case where law 0:44:42.920 --> 0:44:46.960 enforcement wants to access an unblurred background. Now, I'm not 0:44:46.960 --> 0:44:49.640 sure to what extent that's even possible with this technology, 0:44:49.680 --> 0:44:52.239 but what if, say, you know, a government agency made 0:44:52.239 --> 0:44:55.440 a claim for a need to override the auto blurring 0:44:55.480 --> 0:44:58.440 features utilized by others, so they would just have blanket 0:44:58.920 --> 0:45:02.000 power to do this. So you think you're blurring your background, 0:45:02.040 --> 0:45:04.839 but actually you're Yeah, nobody can see, not if you're 0:45:04.880 --> 0:45:08.239 on the phone with you know, with with with someone 0:45:08.280 --> 0:45:11.040 who's actually a government employee or whoever happens to have 0:45:11.080 --> 0:45:13.480 the magic key in this scenario. Oh, I guess it's 0:45:13.520 --> 0:45:15.200 kind of like how you think you can have your 0:45:15.200 --> 0:45:17.600 phone turned off, or you think you can have GPS 0:45:17.680 --> 0:45:20.879 turned off, but in fact it is still location tech. Yeah, yeah, 0:45:20.880 --> 0:45:23.480 that's sort of thing. And again I don't have a 0:45:23.480 --> 0:45:25.879 detailed enough knowledge of this particular software. I'm not saying 0:45:26.600 --> 0:45:29.879 not not not applying this directly to the Skype scenario here, 0:45:29.880 --> 0:45:33.080 but just sort of thinking in general. Um, now, in 0:45:33.120 --> 0:45:36.279 this particular case, I'm also assuming that the broad definition 0:45:36.320 --> 0:45:38.480 of a face is in place, at least in part 0:45:38.560 --> 0:45:41.880 to avoid situations where a human beings face is blurred 0:45:41.880 --> 0:45:46.120 because the AI can't handle, say, facial disfiguration. Because I 0:45:46.120 --> 0:45:48.720 think we can understand why we wouldn't want an AI 0:45:48.880 --> 0:45:53.399 like this to lean heavily on norms to promote ideals 0:45:53.440 --> 0:45:57.040 about who and who doesn't have a face? Sure though 0:45:57.200 --> 0:46:00.319 this in considering this with face recognition, and we get 0:46:00.360 --> 0:46:04.080 into some some interesting and you know, at times disturbing territory. 0:46:04.719 --> 0:46:08.480 Matthew Galt had an article in Vice last February titled 0:46:08.480 --> 0:46:13.279 Facial recognition software regularly miss genders trans people, detailing how 0:46:13.320 --> 0:46:17.000 these systems were simply not built with trans or non 0:46:17.000 --> 0:46:20.839 binary people in mind, and can quote continue to reinforce 0:46:21.120 --> 0:46:24.000 existing biases? Oh yeah, I mean, as with a lot 0:46:24.000 --> 0:46:27.640 of things, I think sometimes there is an illusion that 0:46:28.320 --> 0:46:33.000 machines somehow will be free from applying biases to humans 0:46:33.000 --> 0:46:35.560 that that humans apply to each other. But I think 0:46:35.600 --> 0:46:38.239 we we've got ample evidence now that that is not 0:46:38.320 --> 0:46:42.480 the case. That human biases get quite easily mapped onto 0:46:42.719 --> 0:46:46.840 artificial intelligence through assumptions used in the in the creation 0:46:46.880 --> 0:46:49.640 of these algorithms, or through the data sets they're trained on. 0:46:50.080 --> 0:46:52.280 Right and then, and as Galt explores in the article, 0:46:52.320 --> 0:46:54.120 like part of it too is just like who's building 0:46:54.480 --> 0:46:57.040 these programs, You know, it's built being built by programmers 0:46:57.040 --> 0:47:00.520 and engineers, people that are that that may just not 0:47:00.680 --> 0:47:02.880 have have ever really given serious study to some of 0:47:03.000 --> 0:47:05.800 like say the gender issues that are you know inherent 0:47:05.840 --> 0:47:09.200 to the problem. He also mentions how past databases have, 0:47:09.320 --> 0:47:13.640 for instance, misidentified black people in criminal databases and even 0:47:13.680 --> 0:47:16.880 in some cases had they failed to see black people 0:47:16.960 --> 0:47:19.719 at all. Yeah, like say, if they are trained primarily 0:47:19.800 --> 0:47:24.160 on data sets with lighter skinned faces. In fact, just 0:47:24.320 --> 0:47:27.680 last December December of twenty nineteen, the National Institute of 0:47:27.719 --> 0:47:31.000 Standards and Technology in the US tested a hundred and 0:47:31.000 --> 0:47:35.120 eighty nine facial recognition algorithms from ninety nine developers, which 0:47:35.480 --> 0:47:39.000 includes like some big name developers, and found that they 0:47:39.000 --> 0:47:42.400 were far less accurate at identifying African American and Asian 0:47:42.440 --> 0:47:47.120 faces compared to Caucasian faces, and African American females were 0:47:47.120 --> 0:47:50.440 even more likely to be misidentified. Uh. Now this was 0:47:50.480 --> 0:47:54.200 reported in various places. Uh, but the article I was 0:47:54.239 --> 0:47:58.440 reading about it BBC News text article facial recognition fails 0:47:58.480 --> 0:48:01.760 on race Government study says Yeah, I've read about several 0:48:01.800 --> 0:48:04.480 cases like this. Uh. I mean, I think it's just 0:48:04.719 --> 0:48:08.280 so important for people, especially working in the technology space, 0:48:08.320 --> 0:48:10.680 to remember, you know, don't fall for the myth that 0:48:10.800 --> 0:48:13.720 it's unbiased just because it's a machine and not a person. 0:48:13.840 --> 0:48:17.919 People's biases end up in the machine. The rules come 0:48:18.000 --> 0:48:20.360 from us. Okay, well, I think we're gonna have to 0:48:20.400 --> 0:48:23.399 call the first episode right there. But when we come 0:48:23.400 --> 0:48:25.920 back in the in the next in the series of episodes, 0:48:25.920 --> 0:48:28.719 we're gonna be definitely talking about facial recognition in the 0:48:28.800 --> 0:48:32.239 organic brain, and we'll be moving on more to the 0:48:32.320 --> 0:48:36.080 history of technological facial recognition. We'll get to talk about Greebel's. 0:48:36.120 --> 0:48:39.439 We love Greebel's. Greebles were new to me, but there's 0:48:39.440 --> 0:48:41.799 a whole world. We could do a whole podcast on Greebel's. 0:48:41.840 --> 0:48:43.319 You might think they were new to you, they weren't 0:48:43.320 --> 0:48:45.080 new to you. We've talked about Greebeles, we have talked 0:48:45.080 --> 0:48:50.160 about Greebles have the most delectable spikes. Okay, if you're curious, 0:48:50.160 --> 0:48:51.879 you'll have to come back next time to find out. 0:48:52.000 --> 0:48:54.600 Come back for the Greebel's. In the meantime, if you 0:48:54.600 --> 0:48:56.359 want to check out other episodes of Stuff to Blow 0:48:56.360 --> 0:48:59.760 your Mind, you can find us anywhere you find podcasts. 0:49:00.200 --> 0:49:01.759 If you want to just a handy way to check 0:49:01.840 --> 0:49:03.799 us out, go to stuff to Blow your Mind dot 0:49:03.840 --> 0:49:06.080 com and that'll shoot you over to the I Heart 0:49:07.239 --> 0:49:10.000 listing for our program. But wherever you get the show, 0:49:10.120 --> 0:49:12.520 just make sure that you subscribe, make sure that you 0:49:12.640 --> 0:49:15.200 rate and review. These are great ways to help us 0:49:15.239 --> 0:49:17.680 out and just tell a friend about the show. That 0:49:17.800 --> 0:49:21.600 also helps. And don't forget our other podcast, Invention. Invention 0:49:21.800 --> 0:49:24.840 is a journey through human techno history. Oh yeah, I 0:49:24.840 --> 0:49:26.520 feel like we've been doing a lot most of our 0:49:26.560 --> 0:49:30.440 technology stuff on invention these days, and so uh so 0:49:30.680 --> 0:49:32.680 I'm glad to be getting back into the techno space 0:49:32.719 --> 0:49:35.359 a little bit on Stuff to Blow your Mind today. Yeah, absolutely, 0:49:35.480 --> 0:49:37.439 even if it is for a kind of dystopian sci 0:49:37.480 --> 0:49:41.160 fi topic like this. Anyway, huge things as always to 0:49:41.200 --> 0:49:44.359 our excellent audio producer Seth Nicholas Johnson. If you would 0:49:44.400 --> 0:49:46.160 like to get in touch with us with feedback on 0:49:46.200 --> 0:49:48.600 this episode or any other, to suggest a topic for 0:49:48.640 --> 0:49:50.920 the future, or just to say hello, you can email 0:49:51.000 --> 0:50:01.799 us at contact. That's Stuff to Blow your Mind dot com. 0:50:01.880 --> 0:50:03.720 Stuff to Blow Your Mind is a production of iHeart 0:50:03.760 --> 0:50:06.439 Radio's How Stuff Works. For more podcasts from my heart Radio, 0:50:06.600 --> 0:50:09.160 is that the iHeart Radio app, Apple Podcasts, or wherever 0:50:09.200 --> 0:50:15.680 you listen to your favorite shows.