[MUSIC PLAYING] SPEAKER 1: This is the [INAUDIBLE]. [MUSIC PLAYING] DAVID MALAN: Hello, world. This is the CS50 Podcast. My name is David Malan. And I'm here again, with CS50s own, Brian Yu. BRIAN YU: Good to be back. DAVID MALAN: So in the news, of late, has been this app for iOS and Android called FaceApp, as some of you might have heard. Brian, have you used this app before? BRIAN YU: So people keep talking about this. I don't have it myself, on my phone. But one of our teaching fellows for CS50 does have it and was actually showing it to me earlier, today. It's kind of scary, what it can do. So the way it seems to work is that you open up the app, on your phone. And you can choose a photo of yourself, from your photo library, a photo of you or a friend. And you submit it. And then you can apply any number of these different image filters to it, effectively. But they can do a variety of different things. So they can show you what they think you would look like when you are older, and show you an elderly version of yourself, or what you looked like when you were younger. They can change your hairstyle. They can do all sorts of different things to the background of the image, for instance. So it's a pretty powerful image tool. And it does a pretty good job of trying to create a realistic looking photo. DAVID MALAN: Yeah. It's striking. And I discovered this app two years after everyone else did, it seems. Because someone sent me a modified photo of myself recently, whereby I was aged in the photo. And it was amazing. The realism of the skin, I thought was compelling. And what was most striking to me, so much so, that I forwarded it to a couple of relatives afterward, is that I look like a couple of my older relatives do in reality. And it was fascinating to see that the app was seeing these familial traits in myself, that even I don't see when I look in the mirror, right now. But apparently, if you age me and you make my skin a different texture, over time, oh my god, I'm going to actually look like some of my relatives, it seems. BRIAN YU: Yeah. It's incredible what the app can do. A human trying to do this type of thing on their own, might not be able to do at all. And I think that really just speaks to how powerful machine learning has gotten, at this point. That these machines have been trained to look at these huge data sets of analyzing younger and older pictures probably, and trying to understand, fundamentally, how you translate one younger photo to an older photo. And now, they've just gotten really good at being able to do that in a way that humans, on their own, never would've been able to. DAVID MALAN: So this, is then, related to our recent chat about machine learning, more generally. Where, I assume, the training data in this case is just lots, and lots, and lots of photos of people, young and old, and of all sorts. BRIAN YU: Yeah. That would be my guess. So FaceApp doesn't publicly announce exactly how their algorithm is working. But I would imagine that it's probably just a lot of training data, where you give the algorithm a whole bunch of younger photos and older photos. And you try and train the algorithm to be able to figure out how to turn the younger photo into the older photo. Such that you can give it a new younger photo as input and have it predict what the older photo is going to look like. DAVID MALAN: It's amazing. It's really quite fascinating too, to allow people to imagine what they might look like in different clothes, or I suppose, with different makeup on, or so forth. Computers can do so much of this. But it's actually quite scary too, because a corollary of being able to mutate people's faces in this way, digitally, is that you can surely identify people, as well. And I think that's one of the topics that's been getting a lot of attention here, certainly in the US, whereby a few cities, most recently, have actually outlawed, outright, the police's use of facial recognition to bring in suspects. For instance. Somerville, Massachusetts, which is right around the corner from Cambridge, Massachusetts, here, did this. And I mean, that's actually the flip side of the cool factor. I mean, honestly, I was pretty caught up in it, when I received this photo of myself some 20, 30, 40 years down the road. Sent it along happily to some other people. And then didn't really stop to think, until a few days later, when I started reading about FaceApp and the implications thereof. That actually, this really does forebode a scary future, where all too easily can computers, and whatever humans own them, pick us out in the crowd or track, really in the extreme, your every movement. I mean, do you think that policy is really the only solution to this? BRIAN YU: So I think that certainly, technology is going to get good enough that facial recognition is going to keep getting better. Because it's already really, really good. And I know this from whenever photos get posted on Facebook. And I'm in the background corner of a very small part of the image, Facebook, pretty immediately, is able to tell me, oh, that's me in the photo. When I don't even know if I would have noticed myself in the photo. DAVID MALAN: I know. Even when it just seems to be like a few pixels, off to the side. BRIAN YU: Yeah. So technology is certainly not going to be the factor that holds anyone back, when it comes to facial recognition. So if a city wants to protect itself against the potential implications of this, then I think policy is probably the only way to do it. Though it seems like the third city that most recently banned facial recognition in the city is Oakland. And it looks like their main concern is the misidentification of individuals, and how that might lead to the misuse of force, for example. And certainly, facial recognition technology is not perfect, right now. But it is getting better and better. So I can understand why more and more people might feel like they could begin to rely on it, even though it's not 100% accurate and may never be 100% accurate. DAVID MALAN: But that too, in and of itself, seems worrisome. Because if towns, or cities, are starting to ban it on the basis of the chance of misidentification, surely the technology, as you say, is only going to get better, and better, and better. And so that argument, you would think, is going to get weaker, and weaker, and weaker. Because I mean, even just a few years ago, was Facebook, you noted, claiming that they could identify humans in photos with an accuracy-- correct me if I'm wrong-- of 97.25%. Whereas, humans, when trying to identify other humans and photos, had an accuracy level of 97.5%. So almost exactly the same statistic. So at that point, if the software is just as good, if not better, than humans' own identification, it seems like a weak foundation on which to ban the technology. And really, our statement should be stronger than just, oh, there's this risk of misidentification. But rather, this is not something we want societally, no? BRIAN YU: Yeah. I think that, especially now that facial recognition technology has gotten better. But when the Facebook did that study, I think that was back in 2014, or so. So I would guess that Facebook's facial recognition abilities have gotten even better than that, over the course of the past five years, or so. So facial recognition is probably better when a computer is doing it than when humans are doing it, by now, or at least close to as good. And so given that, I do think that when it comes to trying to decide on how we want to shape the policies in our society, that we should not just be looking at how accurate these things are. But also, looking at what kind of technologies do we want to be playing a role in our policing system, and in the way that the society runs, and the rules there. DAVID MALAN: And I imagine this is going to play out differently in different countries. And I feel like you've already seen evidence of this, if you travel internationally. Because customs agencies, in a lot of countries, are already photographing, even with those silly little webcams, when you swipe your passport and sign into a country. They've been logging people's comings and going, for some time. So really the technology is just facilitating all the more of that and tracking. I mean, in the UK, for years, they've been known as having hundreds, thousands of CCTVs, closed-circuit televisions. Which, I believe, historically were used really for monitoring, either in real time or after the fact, based on recordings. But now, you can imagine software just scouring a city, almost like a Batman. I was just watching, I think, The Dark Knight, the other day, where Bruce Wayne is able to oversee everything going on in Gotham, or listen in, in that case, what bias people's cell phones. It just feels like we're all too close to the point where you could do a Google search for someone, essentially on Google Maps, and find where they are. Because there are so many cameras watching. BRIAN YU: Yeah. And so, those privacy concerns, I think, are part of what this whole recent controversy has been with facial recognition and FaceApp. And in particular, with FaceApp, the worry has been that when FaceApp is running these filters to take your face and modify it to be some different face. It's not just a program that's running on your phone to be able to do that sort of thing. It's that you've taken a photo, and that photo is being uploaded to FaceApp servers. And now your photo is on the internet, somewhere. And potentially, it could stay there and be used for other purposes. And who knows what might happen to it. DAVID MALAN: Yeah. I mean, you, and some other people on the internet, dug into the privacy policy that FaceApp has. And if we read just a few sentences here, one of the sections in the "Terms of Service" are that, "You grant FaceApp consent to use the user content, regardless of whether it includes an individual's name, likeness, voice, or persona sufficient to indicate the individual's identity. By using the services, you agree that the user consent may be used for commercial purposes. You further acknowledge that FaceApp's use of the user content for commercial purposes will not result in any injury to you or any other person you authorized to act on your behalf." And so forth. So you essentially are turning over your facial property, and any photos thereof, to other people. And in my case, it wasn't even me who opted into this. It was someone else who uploaded my photo. And, at the time, I perhaps didn't take enough offense or concern. But that too is an issue, ever more so, when folks are using services like this, not to mention Facebook and other social media apps, and are actually providing, not only their photos, but here is my name, here is my birthday, here are photos from what I did yesterday, and God knows how much more information about you. I mean, we've all tragically opted into this, under the guise of, oh, this is great. We're being social with other people online. When really, we're providing a lot of companies with treasure troves of information about us. And now, governmental agencies seem to be hopping on board, as well. BRIAN YU: Yeah. Facebook, especially. It's just scary how much they know about exactly who you are and what your internet browsing habits are like. It's all too often that I'll be reading about something, on the internet, that I might be interested in purchasing. And all of a sudden, I go and check Facebook, and there's an advertisement for the very thing that I was just thinking about purchasing. Because Facebook has their cookies installed on so many websites that are just tracking every website you visit. And they can link that back to you and know exactly what you've been doing. DAVID MALAN: Yeah. I know. And I was thinking that the other day, because I was seeing ads for something, that I actually went ahead and bought from some website. I don't even remember what it was. But I was actually annoyed that the technology wasn't smart enough to opt me out of those same adverts, once I had actually completed the transaction. But you know, I was thinking too, because just yesterday, I was walking back to the office. And I passed someone who I was, for a moment, super sure that I knew. But I wasn't 100% confident. So I kept walking. And then I felt guilty. And so I turned around, because I didn't want to just walk past someone without saying hello. But then when I saw them a second time, nope, it still wasn't the person I thought it was. But I had that hesitation. And I couldn't help but think now, in hearing these statistics, that Facebook and real humans are on statistically 97% good at detecting faces. That was my 3%, yesterday. Out of 100 people, he was one of the three people in this week that I'm going to fail to recognize. And it really put this into perspective. Because while you might think that humans are perfect, and it's the machines that are trying to catch up, it feels like sometimes it's the machines that are already catching up. And, case in point, there was my own mistake. BRIAN YU: Yeah. And when machine learning algorithms, like facial recognition, but machine learning more generally, are trained, humans are often the baseline that computers are striving to match, in terms of performance. Where, you try and have a human perform some task of trying to label images, or documents, or the like. And then you give the same task to the computer and see, how accurately does the computer match up with the human's task. With the goal of being, how human can we get the computer to be. But there are certain tasks where, you could actually imagine cases, where the computer can get better. And facial recognition is one of those cases, where I feel like, eventually, if not already, it could be better than humans. I think, self-driving cars is another example, which we've talked about before, where there's a lot of potential for cars to be better when they're being driven by computers than when they're being driven by people. DAVID MALAN: But I think that's an interesting one, because it's hard for people I think to rationally acknowledge that, right? Because I feel like, you read all the time about a self-driving car that's been involved in an accident. Because this seems to be evidence, among some minds, of this is why we shouldn't have self-driving cars. Yet, I'm guessing we're nearing the point if we're not there already where it is humans who are crashing their cars far more frequently than these computers. And so, we need to appreciate that, yes, the machines are going to make mistakes. And in the worst extreme case, God forbid, a computer, a machine, might actually hit, and hurt someone, or kill someone. But that's the same reality in our human world. And it's perhaps a net positive, if machines get to the point of being at least better than we humans. Of course, in facial recognition that could actually mean adversarily, for humans, that they're being detected. They're being monitored far more commonly. So it almost seems these trends in machine learning are both for good and for bad. I mean even FaceApp, a couple of years ago, apparently-- and I only realized this by reading up on some of the recent press it's now gotten again-- I mean, even they got themselves into some touchy social waters, when it came to some of the filters they rolled out. Apparently, a couple of years ago, they had a hot filter, which was supposed to make you look prettier in your photos. The catch is, that for many people, this was apparently exhibiting patterns like lightening skin tone, thereby invoking some racial undertones, as to what defines beauty. And they even had in more explicit filters, I gather, a couple of years ago, where you could actually change your own ethnicity, which did not go over well, either. And so those features have since been removed. But that doesn't change the fact that, we are at the point technologically, where computers can do this and are probably poised to do it even better, for better or for worse. And so again, it seems to boil down to then, how we humans decide proactively, or worse, reactively, to put limits on these technologies or restrain ourselves from actually using them. BRIAN YU: Yeah. I think that's one of the big challenges for societies and governments, especially right at this point in time, is catching up with technology, where technology is really moving fast, and, every year, is capable of more things than the year before. And that's expanding the horizon on what computers can do. And I think it's really incumbent upon society to be able to figure out, OK, what things should this computers be able to do, and placing those appropriate limits earlier rather than later. DAVID MALAN: Yeah, absolutely. And I think it's not just photos, right? Because there's been in the press, over the past year or two, this notion of deepfake videos, as well. Whereby, using machine learning and algorithms, you feed these algorithms lots of training data, like lots of videos of you teaching, or talking, or walking, and moving, and so forth. And out of that learning process can come a synthesized video of you saying something, moving something, doing something, that you never actually said, or did, or moved. A couple of clips gained a decent amount of notoriety some months ago, because someone did this, for instance, for President Obama in the US. In fact, do you want to go ahead and play the clip of this deepfake? So there is a video component, too. But what you're about to hear is not Obama, much as it sounds like him. BRIAN YU: Yeah, sure. [AUDIO PLAYBACK] - We're entering an era in which our enemies can make it look like anyone is saying anything, at any point in time. Even if they would never say those things. So for instance, they could have me say things like, I don't know, Killmonger was right, or Ben Carson is in the sunken place. Or, how about this, simply, President Trump is a total and complete dipshit. Now, you see, I would never say these things, at least not in a public address. But someone else would, someone like Jordan Peele. This is a dangerous time. Moving forward, we need to be more vigilant with what we trust from the internet. That's a time when we need to rely on trusted news sources. It may sound basic, but how we move forward in the age of information is going to be the difference between whether we survive or whether we become some kind of fucked-up dystopia. Thank you. And stay woke bitches. [END PLAYBACK] DAVID MALAN: So if you're familiar with Obama's voice, this probably sounds quite like him, but maybe not exactly. And it might sound a bit more like an Obama impersonator. But honestly, if we just wait a year, or two, or more, I bet these deepfake impressions of actual humans are going to become indistinguishable from the actual humans themselves. And in fact, it's perhaps all too appropriate that this just happened on Facebook, or more specifically, on Instagram recently, where Facebook's own Mark Zuckerberg was deepfaked via video. Should we go ahead and have a listen to that, too? [AUDIO PLAYBACK] - Imagine this for a second, one man with total control of billions of people's stolen data, all their secrets, their lives, their futures. I owe it all to Spectre. Spectre showed me that, whoever controls the data, controls the future. [END PLAYBACK] DAVID MALAN: So there too, it doesn't sound perfectly like Mark Zuckerberg. But if you were to watch the video online-- and if you go ahead Indeed and google President Obama deepfake and Mark Zuckerberg deepfake, odds are, you'll find your way to these very same videos, and actually see the mouth movements and the facial movements that are synthesized by the computer, as well. That too, is only going to get better. And I wonder, you can certainly use this technology all too obviously for evil, to literally put words in someone's mouth that they never said, but they seem to be saying, in a way, that's far more persuasive than just misquoting someone in the world of text or synthesizing someone's voice, as seems to happen often in TV and movie shows. But doing even more compellingly because people are all the more inclined, I would think, to believe, not only what they hear or read, but what they see, as well. But you could imagine, maybe even using this technology for good. You and I, for instance, spend a lot of time preparing to teach classes on video, for instance, that don't necessarily have students there physically. Because we do it in a studio environment. So I wonder, to be honest, if you give us a couple of years time and feed enough recordings of us, now in the present, to computers of the future, could they actually synthesize you teaching a class, or me teaching a class, and have the voice sound right, have the words sound right, have the facial and the physical movements look right? So much so, that you and I, down the road, could just write a script for what it is that we want to say, or what it is we want to teach, and just let the computer take it the final mile? BRIAN YU: That's a scary thought. We'd be out of a job. DAVID MALAN: Well, someone's got to write the content. Although, surely if we just feed the algorithms enough words that we've previously said. You could imagine, oh, just go synthesize what it is my thoughts would be on this topic. I don't know. I mean, there's some actually interesting applications of this, at least if you disclaim to the audience, for instance, that this is indeed synthesized and not the actual Brian or the actual David. But if you're a fan of Black Mirror, the TV show that's been popular for a few years now, on Netflix, there's actually in the most recent season-- no spoilers here-- but in most recent season, starring Miley Cyrus, and the rest of the cast, actually touch on this very subject, and use, although they don't identify it by name, this notion of deepfaking, when it comes to videos. BRIAN YU: Yeah. It's a very interesting technology, for sure. And these videos of Mark Zuckerberg and Obama, certainly you can tell, if you're watching and paying attention closely, that there's certain things that don't look or don't feel quite right. But I would be very curious to see a Turing test, of sorts, on this type of thing. Where you ask someone to be able to look at two videos and figure out, which one is the actual Obama, and which one is the actual Mark Zuckerberg. I guess that, on these videos, most people would probably do a pretty good job. But I don't think it'd be 100%. But I would be very curious to see, year after year, how that rate would change. And as these technologies get better, as people become less able to be able to distinguish, to the point where it just be a 50/50 shot as to which one of the fake. DAVID MALAN: Yeah. Especially when it's not just celebrities, but it's a person you've never met and you are seeing them, or quote-unquote them for the first time on video. I bet it would be even harder for a lot of folks to distinguish someone for whom they don't have just ample press clippings in their memory, of having seen them or heard them before. So what do you think, in the short term-- because this problem only seems to get scarier and worse down the road-- is there anything people like you, and I, and anyone else out there can actually do to protect themselves against this trending, if you will? BRIAN YU: So I think one of the important things is just being aware of it, and being mindful of it, and being on the lookout for it as it comes up. Because certainly, there is nothing we can really do to stop people from generating content like this, and generating fake audio recordings, or video recordings. But I think that, if people look at something, and it's potentially a fake video, and you just take it at face value as accurate. Then that's a potentially dangerous thing. But encouraging people to take a second look at things, to be able to look a little more deeply to try and find the primary sources, that's probably a way to mitigate it. But even then, the ultimate primary source is the actual person doing the speaking. So if you can simulate that, then even that's not a perfect solution. DAVID MALAN: So is it fair to say maybe, that the biggest takeaway here, certainly educationally, would be just critical thinking and seeing, hearing something and deciding for yourself evaluatively if this is some source I should believe. BRIAN YU: Yeah, I'd say so. DAVID MALAN: And you should probably stop uploading photos of yourself to Facebook, and Instagram, and Snapchat, and the like. BRIAN YU: Well, that's a good question. Should you stop uploading photos to Facebook, and Instagram, and Snapchat. I mean, certainly there's a lot of positive value for that. Like my family always loves it when they see photos of me on Facebook. And maybe is that worth the trade-off of my photos being online? DAVID MALAN: Living in a police state. [LAUGHTER] I don't know. I mean, I think that to some extent, the cat is out of the bag. There's already hundreds of photos, I'm guessing, of me out there online, whether it's in social media or other people's accounts that I even know about, for instance, because I just wasn't tagged in there. But I would think that that's really the only way to stay off the grid is not to, at least, participate in this media. But again, especially in the UK, and other cities, and surely other locations here in the US, you can't even go outside anymore without being picked up by one or more cameras, whether it's in an ATM, at a bank, or whether it's street view camera up above, or literally Street View. I mean, there are cars driving around taking pictures of everything they see. And at least companies, like Google, have tended to be in the habit of blurring out faces. They still have those faces somewhere in their archives [INAUDIBLE]. BRIAN YU: Yeah. I was actually just grocery shopping at Trader Joe's the other day. And I was walking outside. An Apple Maps car drove by with all their cameras that we're looking around and taking photos of the street. DAVID MALAN: I saw one recently, too. But their cars are not nearly as cool as Google's. BRIAN YU: I've never seen a Google car, in person. DAVID MALAN: Oh, yeah. No, I've seen them from time to time. They're much better painted and branded. Apple's looked like someone had just set it up on top of their own car. [LAUGHTER] Well, and on that note, please do keep the topics of interest coming. Feel free to drop me and Brian a note at podcast@cs50.harvard.edu, if you have any questions or ideas for next episodes. But in the meantime, if you haven't yourself seen or tried out FaceApp, don't necessarily go, and rush, and download, and install this app. That was not intended to be our takeaway. But be mindful of it. And certainly, if you just google FaceApp on Google Images, or the like, you can actually see some examples of just how compelling, or how frightening, the technology is. So it's out there. This then was the CS50 Podcast. My name is David Malan. BRIAN YU: I'm Brian Yu. See you all next time.