AI and the Rule of Law ft. David Mattin & Mikhail Voloshin

0:00Hi, everyone. Today's Real Vision Daily Briefing is sponsored by Crane Shares. Learn about their KCCAETF at CraneShares.com. forward slash KCCA forward slash Real Vision. Now to the top analysis of today's markets.

0:22What's going on, guys? Welcome to AI Firehose. I'm Ash Bennington. Great panel of guests today. Today, I'm joined, as always, by Mikhail Voloshan and also David Matten. David, welcome. Hi there, Ash. Hi there, Mikhail. How are you? Always a pleasure to be here. Doing great, man. Listen, I should say we had a Twitter Spaces conversation a couple of weeks ago that was just absolutely outstanding. I was just blown away by the insight and the observations that you guys brought together, interacting, talking about technology in general and AI specifically. So I'm thrilled to have both of you back here for this conversation.

0:56Let's just jump right in. Mikhail, it's been a couple of weeks since we've done this. What's going on in the AI space? There's been a flurry of activity. There's plenty of stuff coming out every single week, and it's been a couple of weeks since we talked. So it's kind of hard to narrow down a subject to specifically focus on. But I think that what we're going to be talking about this week is face recognition, which is a different class of technology, or at least a different use case of technology than what we've been talking about, which is generative models, primarily LLMs. In particular, there was a feature story a little earlier this month about, or I guess last month, in the New York Times about a company called Clearview AI.

1:45They have some pioneering software that scans just general pictures that are out there, photographs, and claim to be able to identify any face in any photo. The performance of the software is pretty impressive. They're not the only company that's doing this, but the reason that they're kind of a big deal is because they are prominently integrated with law enforcement. Turns out there's a lot of police agencies out there that are using face recognition technology to try and find perpetrators, sometimes find witnesses. and Clearview isn't the only one that produces this software, but it's been the most vocal about how it's used and just sheer numbers and, you know, just like the claims of capability.

2:37Yeah, there's a tremendous amount to talk about here. This particular feature seems to just scare the hell out of people in general for a whole variety of reasons that we're going to get into in just a minute. David, thoughts on this big picture, facial recognition, AI and law enforcement? Yeah, it's a huge one. And I've been following the story of Clearview for years through my newsletter. And it is an absolutely fascinating story. I mean, just as Mikhail said, this is a company that essentially has gone to the internet and just scraped every single human face it can find from the net, billions of faces from social media profiles and so on, and just built this huge database of humanity.

3:20And then it stepped out to clients. And the core clients it's found have been law enforcement departments in the United States. And civil rights organizations in the US have essentially been at war with Clearview about this, that this is a breach of privacy and a kind of appropriation of data on a scale, the like of which we've never seen before. The privacy implications, the human rights implications are huge. Some US police departments are really using Clearview services pretty heavily. I think in other states, they've said, look, we're not going to use it. The EU is taking quite a different position, I think.

4:05And don't quote me on this, but I mean, there's certainly been conversations around it being banned. This is a huge rolling story. It's not going to go away. And it poses huge questions to us when it comes to our response, because just as the idea of a private organization ripping through the internet, scraping billions of faces, building a database, and then selling that back to our governments. Just as that idea is incredibly scary, there's no denying some of the real potent benefits it's going to deliver. It's going to make it easier to catch bad guys. There's all kinds of legitimate questions about what it's going to do that's wrong.

4:51It's going to do some great stuff as well. And when you push it to the logical extension, I mean, when I write about Clearview, I so often also end up mentioning what is happening in China right now. What's emerging in China right now is just an unprecedented form of techno-authoritarianism underpinned by AI. And a big part of the way it's underpinned by AI is that it's underpinned by facial recognition. The Chinese government is building an unparalleled centralized system of control over 1.2 billion people underpinned by facial recognition. Here's one tiny example. If you're playing video games in China, and you're a teenager, there are now laws that have been enacted that the video game companies have to use facial recognition technology to shut the video game off after like one hours, two hours.

5:51If you have young children, then you know how difficult it is to get them off video games sometimes and the war you wage with them over screen time. In China, central government has solved that problem with facial recognition. It's done it in a way that we would view as an extreme breach of privacy rights, but at the same time, it's solved a huge problem. So we're going to face this ethical conundrum? Do we want to buy into the enormous benefits here, or do we want to remain true to a set of norms and sort of institutional norms and rights and laws that we've built up over centuries? And Clearview is where that ethical conundrum is really kind of where that rubber is really meeting the road.

6:37And that's what makes it such a fascinating company at the moment. You know, Clearview's innovation isn't necessarily technical. In fact, I believe that the Times article specifically says that its innovation is ethical. The Times article claims that this technology was forbidden by previous large stakeholders such as Google or Facebook. The idea is that these guys already had face rec technology of equivalent power way back in 2011, but they decided to not release it to the public because it was too powerful. I've got opinions about that. But the bottom line is that Clearview is innovating specifically because instead of withholding this technology, they specifically said, yeah, let's do it and let's put it out there.

7:29And their CEO, Hoan Ton Thot, I'm so sorry for what I'm doing to that name, but he specifically said that upon building this thing, he wanted to release it in a pro-social manner that benefits the world in some way. And the avenue that he chose to pursue is by releasing it in a way that benefits law enforcement. So whether or not that is pro-social or literally the exact opposite thereof is kind of what we're talking about. Now, you're absolutely right about what's going on in China. And in fact, I didn't know that about video games. I would be so arrested in China for so many reasons, but that's just one more.

8:13That's good to know. Um, there's a slightly more amusing anecdote that I have from China that dates back to 2018, where a woman was arrested for jaywalking because a traffic cam picked her up in the middle of the street. The woman's name was Dongming Zhu. Again, sorry what I'm doing to that name. And she has a very large air conditioning business. And it turned out that she advertises on the sides of buses. And what the traffic cam picked up was her face in an ad on the side of a bus. And for that, she got a citation. She was able to plead that down. But that's just a great example of a way that this technology can be used where it gives too much confidence to law enforcement agencies that they've got their guy when in actuality we're dealing with false positives and other glitches that just seem, in this case, silly, but in other cases could be pretty horrifying.

9:20You know, it's interesting when we talk about this case about sort of scraping the web for vast amounts of information. I feel like we all have that one buddy from high school who refuses to get a Facebook account because he's totally paranoid about the fact that it's going to somehow be scraped. And now here we are. Here's your example. Here's your use case. Our paranoid friend from high school seems to be right in terms of the ability of large databases to be gathered based on social media data? Well, not having a Facebook account is not going to save you from having your photo on the internet.

9:52There was a BBC report where they, there was a reporter who interviewed the Clearview CEO, whose name I will not try to pronounce again. And he tried to see, he did a certain, the reporter did a search for himself. And sure, he got all of his like LinkedIn pics and all of his Facebook pics and stuff like that. But he also got pictures of himself in the background of like concerts and music festivals and just on the street from like people whom he doesn't know who were happening to take a selfie while he coincidentally was in the background in that exact moment. And there he is. He's like, wow, that's me.

10:30I'm, you know, I'm in this photo. I, you know, had no idea. You know, it reminds me of this one scene from Parks and Rec where Ron Swanson Googles his own home address and immediately throws his computer into the trash. And for me, what was funny about that was like, throwing your computer into the trash is not going to get your house off of Google. Right. Yeah, David, go ahead, jump in. I see you're not. I mean, yeah, it's a fascinating place we're in. And I can't help but be discomforted by the idea of a private organization, a startup, essentially, just ripping through the internet, scraping billions of faces, compiling a database and selling that to government organisations.

11:15You know, we have huge conversations about what government can and can't and should and shouldn't be able to do and to do to us. And, you know, nowhere are those conversations more energetic than in the United States. And that is in many ways a wonderful thing. There at least needs to be. In my view, I think a conversation, a big shared public conversation about whether this should happen or not. And, you know, in the end, I think this sort of taps into bigger positions I take on sort of AI for the people, essentially. I think we need to move to a place where we as a collective kind of come together and do some of this together in ways that ensure it benefits all of us.

12:04That's partly why I'm so interested in stability AI and stable diffusion and the open source attitude they're taking to large language models and their fundamental mission, which is about empowering everyone with AI and allowing people to build and fine-tune AI models according to their own values and their own beliefs and their own worldview. I think that that sort of AI for the people movement is really powerful. I think it can be a part of a better future or it can lead us to the version of the future that I think most of us want. And perhaps something similar can happen when it comes to billion strong, multi-billion strong databases of human faces.

12:49There's something analogous about the way they've built it. I mean, with large language models, essentially, you've gone to the net, you've scraped billions and billions of words. You've done statistics and you've built this model. They've gone to the net, scraped billions of faces and built this database. We're talking about private organizations raiding the commons, like raiding what is held collectively and what belongs to all of us. And look, mixing that with their labor, doing work on it, and then selling that product back to us. We need to be, and, you know, incredible innovations, incredible creativity is at work.

13:30I don't want to downplay the achievements of these organizations, some of them, but we need to be very careful about how we let that happen and where it's leading us. Hey, everyone, we're going to take a quick break right now to hear a word from our partners. We'll be right back with more of the day's top analysis on the Real Vision Daily Briefing. Have you ever wanted to trade Bitcoin but haven't dared try? With Plus500 Futures, you can trade crypto without the hassle of opening a wallet. With just a few clicks, you can register and start practicing with their free and unlimited demo. See a trading opportunity?

14:05You'll be able to trade it in just two clicks. Feel ready? You can move to real money with as little as$100 once your account is approved. And the great thing is that in addition to crypto, Plus500 gives you access to a wide range of instruments. S &P 500, NASDAQ, gas, and much more. Explore equity indices, energy, metals, forex, and beyond. With a simple and intuitive platform, you could trade anytime, anywhere. Experience the fast, accessible futures trading you've been waiting for with Plus500. With over 20 years of experience, Plus500 is your gateway to the markets. Visit us.plus500.com to learn more.

14:40Trading in futures involves the risk of loss and is not suitable for everyone. Not all applicants will qualify. Plus500. It's trading with a plus.

14:53Let me just jump in real quick here. I want to ask this question of you, Mikhail. We're talking about the public policy implications. We're talking about some of the philosophical aspects of this. Let's zoom the camera out a little bit for people who are just trying to get their heads around the current state of play with this technology. Talk a little bit, Mikhail, about where we are in the state of facial recognition. How good is this technology? How pervasive is this technology? What's the current state of play as we find ourselves here at the end of 2023? Oh, man, that's a question. So first and foremost, face recognition technology is still generally done with neural networks, but it's done with a slightly different wiring pattern of neural networks than a lot of these language models.

15:39Nowadays, most facial recognition technology is done with an image recognition system called a convolution neural network whose wiring patterns are designed to be reminiscent of certain ways that the brain is wired in the visual cortex. And the results have started to get really good since about 20, I want to say 2017 or so, maybe a little bit earlier. There are some underlying technical factors that have allowed for deep neural networks to be a lot more trainable nowadays than they used to be. Now, in terms of how good they are, it's a great question because a lot of the articles that you read on this topic will give you a lot of hype.

16:29And I will tell you that at scale, a lot of that hype is unwarranted. And so here's what I mean. um the uh so clear view itself uh boasts a very very high uh recognition rate uh they claim a hundred percent nearly a hundred percent recognition but that's only on mug shots in other words if you have two mug shots of the same person they claim that you're able to identify that same person in you know that the ai can identify that it's the same person that is a very, very far cry from having some random photo of, you know, by some random camera with random lens properties, and then have a completely different photo of that same person at a different time of day with different facial hair, and, you know, or different styles, wearing different glasses and so on, and then say that that's still the same.

17:20The reason that, in my opinion, the reason that this technology was taken off the board by giants such as Facebook and Google is precisely because they had giant data sets to work with and had a lot to lose, whereas this startup Clearview AI did not. So to illustrate what usually goes wrong with this, do normal people know about this thing called the birthday paradox. It's a mathematical phenomenon that comes up a lot in statistics. I think it's fair to assume that most people probably do not remember birthday paradox from their high school statistics class. So the gist is, if you have random people in a room, what's the probability that any two of those people happen to have the same birthday?

18:13And there's 365 days in a year, so people usually think that that number has to be very high. But in order to reach a 50 % level requires a shockingly low number of people. Something like if memory serves like 23 people, something like that. If you have 23 people in a room, and I'll need to Google this later, then there's a 50 % chance that two of them happen to already have the same birthday. And just as background, the reason that that's the case is that every single time you introduce a new person into the room, you are performing a number of comparisons that is equal to the number of people that are already in that room.

18:52It's an N-squared scaling level. Anyway, the point is the probability of two people having faces that happen to look the same within the tolerances of the software and of the resolutions of the cameras of the pictures that are being compared gets much, much higher the more people you have in your database. Yeah, and the difference is between sort of physiognomy, facial structure, and birthdays is that they change over time. So when you talk about this idea of tolerance, you have to have some degree. I mean, for example, you change your glasses, you grow a beard, you cut your hair, you gain 20 pounds, you lose 20 pounds.

19:32You need to be able to have this variability built into the system. Here's the problem with that. That variability is exactly, as you point out, means that you may look more or less like someone else on any given day. And obviously, when what we're talking about here is guilt and innocence, the ability of the police to show up behind your door. I know one of the stories that we're going to talk about today, Mikhail, is about a guy who was accused of murder, vehicular homicide, homicides specifically. But these are extremely, extremely sensitive contacts. I know the story is a little bit different.

20:04It wasn't him on camera. But we're going to talk a little bit about that. But obviously, this is of the highest priority and the highest sensitivity. So, you know, I myself just had an encounter with failure of visual, of face recognition. I have a financial institution that has a mobile app, and I was trying to prove my identity on this app in order to unlock some features. And it makes me take a photo of my driver's license, and it checks the face on the driver's license, and then it makes me take a live photo in front of the camera itself. right and the um like i could not get this thing to understand that the photo of me in person is the same photo is the same guy as who's in the driver's license no amount of like angles or like changing of lighting or anything could could get this thing to believe that i'm actually me like i don't know have i like look the way that this uh technology works under the hood is supposed to be based on physiological invariance.

21:08It's supposed to be based on things like the spacing between your eyes and the distance from your pupil to the corner of your mouth, that kind of thing. When we did this back in the 90s, these systems were actually hard-coded. There was a really prominent roboticist named Rodney Brooks who ran a very widely acclaimed lab in MIT that had a sort of 17-point checklist for measurements that would stay invariant, and that was the gold standard of face recognition at the time. Nowadays, we just let neural networks figure that out for us, but the point still stands. These are supposed to be based on things that don't change over time.

21:58And yet this thing could not recognize me to save its life. So that's a type one error, a false positive. Sorry, sorry, sorry, sorry. That's a type two error, a false negative. I am me and it doesn't recognize it. There's also, of course, problems with false positives, which is where somebody who isn't me gets recognized as me. And that, like, especially when it comes to law enforcement, leads to a world of hurt. real quick, there was a case in 2018 with, I don't think it was Clearview, I think it was a different face recognition system where a man in the Bronx was pinged for stealing a pair of socks, or a six-pack of socks, and he was pinged because of face recognition technology.

22:41He was arrested, and for whatever reason, the cops believed that the charge stuck. And the reason why he was arrested for it was because he apparently was armed while stealing these socks. So something that would have been a like just a minor shoplifting turns into armed robbery and blah, blah, blah. Um, the point is this man was provable. This man's son was born that day and he was provably on his way to the hospital at the time at which this robbery occurred. So a lot of people believe that it's – I'm not going to – I'm going to take sides on this, but I don't want to like – I believe that it's very, very, very unlikely that the man who was seen in the shoplift cam was also the man who happened to be on his way to his son's birth and happened to decide to commit armed robbery just going to the hospital.

23:34Well, let's pull David into the conversation because obviously there's a lot there. David, clearly a lot of technical questions, a lot of ethical questions, a lot of questions about the nature structure of society, how we negotiate these changes with ourselves. Many points. I saw you smiling and nodding. Thoughts? yeah i mean the first first off the bat you know listening to these stories of course makes you think we can't accept these forms of facial recognition technologies yet as as evidence in in you know in the way that we accept many other things as evidence you know they're not reliable enough yet that they're not kind of secure enough yet to really to really work in that way um we can continue to develop them, they'll continue to get better.

24:20What we're going to hit up against then are different questions, and they will be questions about trade-offs. And that's really what I'm driving at, too, with this comparison with China. There's an incredible and profoundly different technology-fuelled social system emerging across the seas in China that is going to be potent in all kinds of new and incredible ways, what you incur for that potency is you trade off things like privacy and freedoms and individual freedom against centralised control and centralised government. And we in the global north are going to have to have a collective conversation about the extent to which we are happy to make those trade-offs.

25:16And it's going to be very difficult. And these technologies are in all kinds of ways going to be extremely seductive. If you tell millions of parents, like Mikhail, you'd be fine in China on the video games because this only applies to children. But if you tell millions of parents, we've got this tech-fuelled way that will just squash your arguments about screen time and video game time instantly, many instinctively will embrace that. It will appeal to them. But there's a trade-off going on underneath, and it's a trade-off around privacy and individual liberty. We've already demonstrated that almost everyone in the global north is willing to walk down that road and make significant trade-offs for convenience and the kind of superpowers that online life kind of allow you.

26:05We've given our data and we've given so much of ourselves to big technology platforms and big companies in return for the convenience and the superpowers and the entertainment and everything that we get back. We're coming to a kind of terminal station in that journey. We're going to have to ask ourselves really big, really hard questions about who we are and how we want to live as a society and how we want the collective to be sort of administered when it comes to technologies like, you know, like the kind that Clearview is building, really powerful kind of huge facial recognition databases. I mean, and I do take the point, you know, it's right.

26:44Yes, completely. The innovation is not technical. You know, this could have been done. Google or Facebook could have done this a decade ago. The innovation is more sort of, yeah, social ethical innovation, you know, they dared to do it. And exactly as Mikhail said, they had nothing to lose. You know, if Facebook as was, you know, had tried to do this 10 years ago, there would have been huge blowback. That wasn't an issue for Clearview. But we need to get a handle on Clearview in particular, facial recognition in particular, but use this as a lesson about the broader journey and the trade-offs that are coming.

27:21We haven't had the big conversation about trade-offs yet properly, and we need to have it really soon. And how can we do that in a way that allows us to retain the incredible benefits that technology is bringing without selling centuries of social development and the evolution of norms and rights and individual liberty without selling all of that down the river? Because I think that that is a legacy deeply worth protecting. We have to be really careful about that. We're going to take another quick break to hear a word from our partners. We'll be right back with more of the day's top analysis on the Real Vision Daily Briefing.

28:01Yeah, you know, and as we begin to negotiate these trade-offs with ourselves, have this conversation, particularly here in the global north and democratic countries, where there is this dialogue, I guess one of the fears that people have around this technology is it becomes established in the system before that conversation can happen. And we all know with technology that it's very difficult to get the toothpaste back into the tube once it gets out. Yeah, I think that's absolutely right. And that's fundamental to the challenge is that in this exponential age, the pace of technological change is far outpacing our ability to process it or respond to it and certainly to respond to it collectively.

28:42You know, and I think that's a that's a challenge. No one has a good answer to yet. I mean, our institutions, I've said this before. I've said it before on Real Vision. Our institutions are still grappling with the technologies of Web 2.0, you know, like a 25 year old kind of phenomenon that they're grappling with that now. As though that is the live issue, that is the big thing they need to get their head around. and it's proving really difficult for them, they're nowhere near really, truly getting to grips with this AI moment, technologies like Clearview, technologies like ChatGPT. And what do we do about that?