WEBVTT - Will Deepfakes Ruin the World? 0:00:01.120 --> 0:00:03.400 Welcome to Stuff You should know, a production of My 0:00:03.600 --> 0:00:12.560 Heart Radios How Stuff Works. Hey, and welcome to the podcast. 0:00:12.600 --> 0:00:15.920 I'm Josh Clark. There's Charles W. Chuck Bryant. Here's our 0:00:15.960 --> 0:00:20.079 brand new producer from now on, Josh Brother, Josh, Have 0:00:20.200 --> 0:00:22.920 you said your last name? Josh? We don't do that, Okay, 0:00:23.440 --> 0:00:27.840 sometimes we say Jerry's Jerry quit? She really did, but 0:00:27.960 --> 0:00:30.560 she did it like that, kind of like quiet, silent 0:00:30.560 --> 0:00:34.320 wait for her. She just stopped showing up. Yeah, Jerry 0:00:34.320 --> 0:00:37.000 didn't quit. Everyone. We don't think we're not. It's not 0:00:37.280 --> 0:00:40.040 entirely certain until I see her sitting in that chair. Then, 0:00:41.320 --> 0:00:44.680 assuming she's quit, what if, Chuck, she sent us a 0:00:44.800 --> 0:00:48.239 video of herself saying, I quit. I'm so sick of 0:00:48.280 --> 0:00:51.800 you guys. I'm done with this forever. Would you believe 0:00:51.800 --> 0:00:56.319 it then, but the lips didn't quite match up, then 0:00:56.320 --> 0:00:58.120 it would be a deep fake. It would be a 0:00:58.120 --> 0:01:00.680 deep fake. I saw a tweet. I don't remember who 0:01:00.720 --> 0:01:03.920 it was, but they were maybe Ryan Liza or somebody 0:01:04.280 --> 0:01:07.440 was complaining said why do we call these things deep fakes? 0:01:07.440 --> 0:01:10.120 And somebody school them on it. It was kind of 0:01:10.240 --> 0:01:13.640 nice to wat who he said, Ryan Liza I think 0:01:14.360 --> 0:01:19.480 he's like a CNN correspondent journalist. Uh So, first of all, 0:01:19.520 --> 0:01:22.080 we want to issue a c o A here, uh 0:01:23.240 --> 0:01:27.319 that maybe any kids shouldn't listen to this one. We're 0:01:27.360 --> 0:01:33.240 talking about some really every day dark, harmful stuff. Yeah, 0:01:33.280 --> 0:01:36.160 really despicable, gross stuff. The only thing I can think 0:01:36.200 --> 0:01:38.000 of that would be worse than covering this would be 0:01:38.040 --> 0:01:43.200 to to one on like snuff films. I kept thinking 0:01:43.240 --> 0:01:45.679 of that while I was reading this. I don't know, man, 0:01:45.840 --> 0:01:49.480 I don't want to. I don't want to pit types 0:01:49.520 --> 0:01:52.760 of despicable media against one another. But um, I think 0:01:52.800 --> 0:01:55.360 revenge porn might have a leg up on this, which 0:01:55.400 --> 0:01:58.080 this sort of is as well. Right, it's definitely a 0:01:58.120 --> 0:02:00.240 close cousin of it at least. Yeah, but this one 0:02:00.360 --> 0:02:04.280 in for kids. Uh. And I was shocked and dismayed 0:02:04.320 --> 0:02:06.640 because I didn't know about this, And when I saw 0:02:07.440 --> 0:02:11.320 it's a podcast on fake videos, I thought, well, how fine, 0:02:11.880 --> 0:02:16.040 because I love those videos of David Beckham kicking soccer 0:02:16.080 --> 0:02:18.040 balls in the cans from a hundred yards out on 0:02:18.080 --> 0:02:22.480 the beach. That's for real. Uh No, yeah, and it's 0:02:22.520 --> 0:02:25.080 just coincidence he's holding it PEPSI can up. I saw 0:02:25.080 --> 0:02:27.720 it with my own eyes. I thought that's what this 0:02:27.760 --> 0:02:29.200 is about. I was like, all those are fun. It 0:02:29.360 --> 0:02:31.160 is kind of but then I wanted to take a 0:02:31.360 --> 0:02:35.880 bath after this. I can I can understand, So we 0:02:36.280 --> 0:02:39.480 should probably start out after the CEA by saying what 0:02:39.560 --> 0:02:41.880 a deep fake is? A deep fake d E E 0:02:42.080 --> 0:02:46.080 p F A k E all one word is a 0:02:46.160 --> 0:02:50.600 type of video where somebody is saying or doing something 0:02:50.639 --> 0:02:54.639 that they never actually said or did. What you say, Okay, 0:02:54.639 --> 0:02:56.560 this is nothing new. This has been around for a while, 0:02:56.720 --> 0:02:59.200 Like people have doctored photos and videos and stuff like 0:02:59.240 --> 0:03:02.240 that for base sically as long as there's been videos. 0:03:02.280 --> 0:03:06.200 C G. I sure this is different. This is the 0:03:06.400 --> 0:03:10.359 in the same ballpark, but this is an entirely different league. 0:03:10.720 --> 0:03:13.320 Like this league plays on Saturday and Sunday afternoons, not 0:03:13.600 --> 0:03:15.680 Tuesday night, you know what I mean. Like this is 0:03:15.720 --> 0:03:18.399 something totally Just let it simmer for a little while 0:03:18.400 --> 0:03:20.919 and you'll be like, wow, it's a really good analogy. Uh, 0:03:20.919 --> 0:03:23.120 this is just it's different. It's it has a lot 0:03:23.160 --> 0:03:25.919 of the same in principle, but because they are so 0:03:26.000 --> 0:03:29.480 realistic and they're getting more and more realistic by the day, 0:03:29.600 --> 0:03:33.280 they they actually in a lot of people's minds pose 0:03:33.360 --> 0:03:36.400 a threat not just two individual people, as we'll see, 0:03:36.720 --> 0:03:39.680 but possibly to society at large, say a lot of 0:03:39.720 --> 0:03:42.320 people who are really worried about this kind of stuff. Yeah, 0:03:42.360 --> 0:03:46.960 and we're not talking about I'm assuming the fake lip 0:03:47.160 --> 0:03:51.080 reading thing that's deep fake, right or is that just 0:03:52.320 --> 0:03:55.360 no manipulation of video whatsoever? And that's just people using 0:03:55.360 --> 0:03:58.960 their voice So what that is is the Yeah, it's 0:03:59.000 --> 0:04:02.080 just somebody pretend like they're just fake lip reading and 0:04:02.120 --> 0:04:05.360 then doing a voiceover. They're not. They're not manipulating the 0:04:05.440 --> 0:04:08.600 video at all. Okay, No, they're just doing a really 0:04:08.600 --> 0:04:12.240 bad job of lip reading, which is are hysterics. They 0:04:12.240 --> 0:04:15.680 are hilarious. Those I would put those up with the 0:04:15.680 --> 0:04:18.599 the G I. Joe p S A S. Pork chop 0:04:18.640 --> 0:04:21.520 sandwiches like those are just all time classic. Can watch 0:04:21.600 --> 0:04:23.839 them anytime and still laugh. Have you ever seen the 0:04:23.880 --> 0:04:27.520 G I. Joe's action figures on like the dead road Kills. 0:04:30.320 --> 0:04:32.840 It's sad because it's a dead animal, but they'll be like, 0:04:33.120 --> 0:04:35.039 you know, a dead squirrel in the road, and someone 0:04:35.080 --> 0:04:37.200 will pose like a G. I. Joe action figure with 0:04:37.240 --> 0:04:40.719 his like foot on his head. It's like it's a trophy. Yeah, game, 0:04:40.839 --> 0:04:42.960 it's kind of funny. Yeah, I can see that. It's 0:04:43.040 --> 0:04:44.520 let's put it this way, it's as funny as you 0:04:44.520 --> 0:04:48.120 can make a picture of a dead animal and that 0:04:48.200 --> 0:04:51.520 I assume got hit by a car. You hope, But 0:04:51.760 --> 0:04:55.080 maybe they are like killing squirrels. Just after reading this, 0:04:55.279 --> 0:04:57.359 I don't doubt anything. And it makes me hate the 0:04:57.400 --> 0:04:59.760 Internet even more. All right, so let's let's get into 0:04:59.800 --> 0:05:02.640 this a little bit. Okay, Chuck, calm down. We're not 0:05:02.680 --> 0:05:05.919 allowed to have personal positions on these things, so this 0:05:06.000 --> 0:05:09.680 is totally a neutral thing. Okay. So, um, there's this 0:05:09.720 --> 0:05:13.960 really interesting Gizmoto article that that talked about the history 0:05:13.960 --> 0:05:18.320 of kind of um not necessarily deep fixed but altering videos, 0:05:18.320 --> 0:05:23.000 like presenting a doctored video as as reality. And apparently 0:05:23.080 --> 0:05:25.120 there was a long tradition of it at the beginning 0:05:25.320 --> 0:05:28.320 of cinema where people got their news from newsreels, like 0:05:28.360 --> 0:05:30.480 you actually go to a movie theater to see the 0:05:30.520 --> 0:05:34.000 news because you were just a early twentieth century yokel 0:05:34.279 --> 0:05:37.960 living in Kansas or something like that. Yeah, And after 0:05:38.000 --> 0:05:40.720 reading this bit, I thought that was a very Gizmoto 0:05:40.800 --> 0:05:44.240 way to say, here's one not so interesting fact that 0:05:44.279 --> 0:05:46.000 really has not much to do with this. Oh I 0:05:46.040 --> 0:05:48.800 love it really. I personally selected it and put it 0:05:48.839 --> 0:05:50.559 in here, and I thought it was kind of funny. 0:05:50.600 --> 0:05:53.040 I think it's great. Uh yeah, they used to fake, 0:05:53.400 --> 0:05:57.400 uh real life events and recreate them. Don't try to backpedal. 0:05:57.560 --> 0:06:01.080 And that has nothing to do with deep it does 0:06:01.200 --> 0:06:03.640 because one of the big problems are threats from deep 0:06:03.880 --> 0:06:07.200 deep fakes is it's a way of seeing what you 0:06:07.240 --> 0:06:09.839 think is news, but it's not. It's it's a sham. 0:06:09.920 --> 0:06:13.760 It's recreated. Yeah. The difference I see is they were 0:06:13.880 --> 0:06:17.039 recreating real news events and just like here, you didn't 0:06:17.040 --> 0:06:18.520 see it. So this is what it may have looked like, 0:06:18.720 --> 0:06:21.880 but they were passing it off as real. Therein lies 0:06:21.960 --> 0:06:23.920 the tragedy of all times. I thought it was a 0:06:24.000 --> 0:06:29.080 very thin GIZMOTOI whatever, we'll edit this part out Webster's 0:06:29.120 --> 0:06:31.479 defined deep fake. I like it. I put it in 0:06:31.480 --> 0:06:34.760 there specifically because I thought it was good. That's all right, Okay, 0:06:34.800 --> 0:06:37.760 so we'll take another tech Okay, why don't we just 0:06:37.760 --> 0:06:41.560 talk about deep fakes, So Chuck, let's talk about deep fakes. 0:06:42.279 --> 0:06:44.640 We can just cut there. So deep fakes actually are 0:06:44.720 --> 0:06:48.480 super new. Yeah, they and the reason they're called deep fakes. 0:06:48.600 --> 0:06:52.159 Is because in late two thousand seventeen, I think November, 0:06:52.720 --> 0:06:55.559 this guy who was a reditor, a guy who posts 0:06:55.600 --> 0:06:59.200 on Reddit at your first sign warning sign not necessarily read. 0:06:59.200 --> 0:07:01.920 It's pretty sharp and smart and got some good ideas 0:07:01.920 --> 0:07:04.920 going on as as of all the social media platforms 0:07:04.920 --> 0:07:07.600 that throw my two cents in with Reddit, but there 0:07:07.640 --> 0:07:11.400 was a redditor called deep fake d E E P 0:07:11.680 --> 0:07:14.600 F A K E all one word and he said, hey, world, 0:07:14.920 --> 0:07:17.240 look at what I figured out how to do. And 0:07:17.280 --> 0:07:23.440 he started posting pornography but with celebrities faces transposed on it, 0:07:24.560 --> 0:07:26.400 and he said, this is just my hobby, but here's 0:07:26.440 --> 0:07:28.280 how I did it. And he said that he used, 0:07:28.520 --> 0:07:31.920 I'm assuming it's a him. I don't know if it's 0:07:32.200 --> 0:07:34.280 a man or a woman. I'm gonna go with a man. 0:07:35.520 --> 0:07:38.560 UM and he said, I just used Keras and TensorFlow. 0:07:38.680 --> 0:07:41.960 And these are a couple of UM basically open source 0:07:42.040 --> 0:07:45.600 AI programs that this this guy was smart enough to 0:07:45.600 --> 0:07:49.800 figure out how to use to train to create UM 0:07:49.920 --> 0:07:52.840 these videos where you take a celebrity's face and put 0:07:52.880 --> 0:07:55.560 it on a clip from a porn movie and it 0:07:55.640 --> 0:08:00.840 looks like the celebrity is doing what you're seeing. And 0:08:00.920 --> 0:08:03.280 at first it was kind of hokey and not very 0:08:03.440 --> 0:08:07.040 it was very obviously not real. Yeah, I think the 0:08:07.760 --> 0:08:12.920 scary part was how quickly and easily it could be done. Motherboard, 0:08:13.040 --> 0:08:15.200 who we we used to write for every now and then. 0:08:15.760 --> 0:08:18.680 Remember that I tried to forget. They tried to forget 0:08:18.720 --> 0:08:21.080 for sure. Yeah, if I hit us up, what like, 0:08:21.280 --> 0:08:23.679 I feel like seven or eight years ago, I'm trying 0:08:23.680 --> 0:08:26.000 to forget and you're making it really hard. Said you 0:08:26.000 --> 0:08:28.360 guys want to write some blogs for Motherboard. He said, sure, 0:08:28.440 --> 0:08:31.120 so we did. We did. We wrote ten. Yeah, you 0:08:31.120 --> 0:08:32.640 can probably go find those on the internet if you 0:08:32.640 --> 0:08:34.480 want to learn how to they drive a stick shift 0:08:34.600 --> 0:08:37.079 or something. The fine people that motherboards scrubbed those from 0:08:37.120 --> 0:08:40.960 the internet, let's hope. So so Uh, this deep fake 0:08:41.080 --> 0:08:45.760 character figures this out. Another guy released a downloadable desktop 0:08:45.800 --> 0:08:49.840 software that said here, you can do this awful thing too. Right, 0:08:50.080 --> 0:08:53.240 within like two months of of deep Fake coming out 0:08:53.240 --> 0:08:54.680 and saying look what I did and here's how I 0:08:54.720 --> 0:08:57.040 did it, somebody said that's a really good idea. I'm 0:08:57.040 --> 0:08:58.880 going to turn it into an app and make it 0:08:59.400 --> 0:09:03.400 give it to everybody. That's right, And now, uh, people. 0:09:03.440 --> 0:09:05.640 Can you know at this time, this was a very 0:09:05.640 --> 0:09:07.760 short time ago, people, and it's really come a long 0:09:07.760 --> 0:09:11.319 way in the past whatever, not even two years, because 0:09:11.320 --> 0:09:15.520 there's a late right early two eighteen when it really 0:09:15.559 --> 0:09:17.800 first popped up. Yeah. So this thing was downloaded a 0:09:17.840 --> 0:09:21.800 hundred thousand times in the first month alone, and some 0:09:21.840 --> 0:09:24.920 people used it for fun stuff, like uh, putting Nick 0:09:24.960 --> 0:09:27.240 Cage in movies he wasn't in. Yeah, those are called 0:09:27.280 --> 0:09:32.040 dirt fakes. Yeah, they've all got fun names, don't they. Dude. 0:09:32.840 --> 0:09:37.400 Nicholas Cages Yoda is patently objectively hilarious. I didn't like 0:09:37.480 --> 0:09:41.200 that one. I thought the I don't know, the rages 0:09:41.240 --> 0:09:44.040 of all stark thing was interesting, I guess, but none 0:09:44.040 --> 0:09:47.480 of them made me laugh. Like maybe I just don't 0:09:47.520 --> 0:09:50.880 have that kind of sense of humor. Yeah, but yeah, 0:09:50.920 --> 0:09:53.240 I did. Never was like, oh my god, that's hysterical. 0:09:53.360 --> 0:09:55.800 It's Nick Cage's face. I understand. I understand where you 0:09:55.800 --> 0:09:57.440 come from. I don't think I was, like, you know, 0:09:57.760 --> 0:10:01.920 in stitches or anything like that. But it's pretty great. Okay, Okay, 0:10:01.960 --> 0:10:04.200 it's just not my thing. You're not a Gizmoto reader, 0:10:04.240 --> 0:10:06.480 are you. No, none of this is my thing. But 0:10:06.520 --> 0:10:09.720 that doesn't mean we can't report on it. However, since 0:10:09.800 --> 0:10:13.400 it started happening that it became pretty clear pretty quickly 0:10:13.960 --> 0:10:16.440 that this could be a bad thing in the future, 0:10:16.920 --> 0:10:21.839 and not just for putting your ex girlfriend's face on 0:10:22.000 --> 0:10:25.840 a sex video, you know, saying look what she did. 0:10:26.000 --> 0:10:28.800 It could you could put a world leader up there 0:10:28.840 --> 0:10:33.160 and uh really cause a lot of problems. Yes, hypothetically 0:10:33.240 --> 0:10:37.280 you could. And that's that's really as we'll see this, 0:10:37.280 --> 0:10:41.839 this new technology, this deep fake technology, it poses at 0:10:41.880 --> 0:10:46.120 least two risks to immediately obvious risks, and they're hyper 0:10:46.160 --> 0:10:52.800 individualized and hyper macro social. That's risks, but they both 0:10:52.840 --> 0:10:56.079 stem from the same thing, from the same route or 0:10:56.720 --> 0:10:59.600 same seat. To keep the metaphor going and on track, 0:10:59.760 --> 0:11:03.040 that's so, let's talk about the technology behind this, because 0:11:03.080 --> 0:11:07.160 this stuff is just totally fascinating. Surely you agree it 0:11:07.280 --> 0:11:10.200 is AI. It was created by a guy named Ian Goodfellow, 0:11:11.480 --> 0:11:13.640 just this particular type of AI. He didn't make the 0:11:13.679 --> 0:11:17.959 deep fake stuff, no, no, no. But basically what this 0:11:18.000 --> 0:11:21.600 model you know, everyone knows AI is basically when you 0:11:21.640 --> 0:11:25.520 teach a machine to start teaching itself, starts learning on 0:11:25.559 --> 0:11:29.480 its own, which is a little creepy. Um. But the 0:11:29.520 --> 0:11:31.920 model that they're using these days is called artificial neural 0:11:31.960 --> 0:11:35.680 net which is uh, machine learning. And basically what they've 0:11:35.720 --> 0:11:37.560 done in this case is all you have to do 0:11:37.600 --> 0:11:40.760 is show something a lot of data for it to 0:11:40.760 --> 0:11:43.040 start to be able to recognize that data when you 0:11:43.080 --> 0:11:45.840 aren't showing it that data. Yeah, and it learns on 0:11:45.920 --> 0:11:50.400 its own. What makes the classic example is um AI 0:11:50.480 --> 0:11:53.960 that can pick out pictures of cats, and it's easy enough, 0:11:54.000 --> 0:11:57.160 but you don't tell the AI, here's what a cat is. 0:11:57.400 --> 0:12:00.880 Find pictures of cats in this data set. It's here's 0:12:00.880 --> 0:12:03.640 a bunch of stuff, and figure out what a cat is, 0:12:03.679 --> 0:12:06.440 and they get really good at picking it out. You 0:12:06.480 --> 0:12:08.600 can also turn it the opposite way, once you have 0:12:08.640 --> 0:12:11.800 an AI trained on identifying cats and get it to 0:12:11.960 --> 0:12:17.120 produce pictures of cats, but they're usually terrible and often 0:12:17.360 --> 0:12:20.160 very very bizarre, like anyone would look at it and 0:12:20.200 --> 0:12:22.840 be like a human didn't make this. It's just off 0:12:23.000 --> 0:12:27.080 in some really obvious ways. And what you and Goodfellow 0:12:27.120 --> 0:12:31.199 figured out was a way around that problem. Yeah. So, uh, 0:12:31.360 --> 0:12:34.160 I'm not sure I agree with his wording here, um, 0:12:34.360 --> 0:12:36.680 but we'll we'll say what he calls it, uh, he 0:12:36.880 --> 0:12:40.560 set up two teams and one is a generator and 0:12:40.600 --> 0:12:45.880 one is a discriminator, and he calls it generative adversarial network. 0:12:46.040 --> 0:12:49.400 So basically his contention is that these two are adversarial. 0:12:49.840 --> 0:12:57.040 I saw it more as like a managerial in nature. Okay, bureaucratic, Yeah, 0:12:57.120 --> 0:12:58.720 I mean, isn't that what it felt like you or 0:12:59.000 --> 0:13:02.480 the the scriminer is like, yeah, I'm gonna need you 0:13:02.520 --> 0:13:05.439 to come in on Saturdays. It kind of felt like, 0:13:05.800 --> 0:13:08.000 So you've got these two networks and they're both trained 0:13:08.040 --> 0:13:10.520 on the same data set, but the generator is the 0:13:10.559 --> 0:13:15.040 one that's producing these fake cats, and then there's a 0:13:15.120 --> 0:13:18.360 discriminator or what I like to call a manager saying 0:13:18.760 --> 0:13:21.280 these look good. These don't look so good? Right. The 0:13:21.280 --> 0:13:23.600 other way, the way that good Fellow has proposed it, 0:13:23.679 --> 0:13:27.360 is that the discriminator is going through and looking at 0:13:27.400 --> 0:13:31.240 these generated pictures and trying to figure out if it's 0:13:31.320 --> 0:13:33.640 real or if it's from the data, or if it's 0:13:33.679 --> 0:13:37.439 fake the generator created it, or if the or if 0:13:37.440 --> 0:13:40.600 it comes from the data set, And based on the 0:13:40.679 --> 0:13:45.000 feedback that the manager gives the generator um the generator 0:13:45.040 --> 0:13:47.160 is going to adjust as parameters so that it gets 0:13:47.240 --> 0:13:52.160 better and better at putting out more realistic pictures of cats. Yeah. 0:13:52.160 --> 0:13:54.320 I don't get the adversarial part unless it gets mean 0:13:54.559 --> 0:13:57.000 and how it delivers that message the way the reason 0:13:57.040 --> 0:14:00.840 they call it ever sincere is I saw it, but like, um, 0:14:00.880 --> 0:14:05.880 it's like an art forger and detective and the art 0:14:05.880 --> 0:14:08.679 forgers putting out or an appraiser is a better way 0:14:08.720 --> 0:14:11.199 to put it. No art forgers putting out forged art, 0:14:11.280 --> 0:14:13.040 and the appraisers like this is fake, this is fake, 0:14:13.120 --> 0:14:15.440 this is fake. Well, I'm not sure about this one. 0:14:16.559 --> 0:14:18.760 I don't know if this one is fake. This is real, 0:14:18.960 --> 0:14:21.160 this is real, this is real. And then at that point, 0:14:21.360 --> 0:14:26.120 the generator has become adept at fooling an AI that's 0:14:26.120 --> 0:14:29.360 trained to identify pictures of cats creating pictures of cats 0:14:29.400 --> 0:14:34.160 that don't exist. Okay, it's adversarial. They're trying, the generators 0:14:34.160 --> 0:14:38.200 trying to fool the discriminator, and the discriminators trying to 0:14:38.240 --> 0:14:41.240 thwart the generator. That's the adversarial part. But in the 0:14:41.320 --> 0:14:44.280 end they're really on the same team. Yeah, okay, I 0:14:44.320 --> 0:14:46.360 guess that's where it loses me. You have a really 0:14:46.400 --> 0:14:50.680 positive image of corporate America. What is that the two 0:14:50.680 --> 0:14:54.200 with anything. The managers on the same team as everybody. 0:14:54.200 --> 0:14:56.960 Come on, get on board, Get on the trolley, Josh. So, 0:14:57.200 --> 0:15:00.200 if you want to look up Mona Lisa talking, there 0:15:00.200 --> 0:15:03.520 was a video this year from Samsung that showed how 0:15:03.640 --> 0:15:07.360 you could do this, and all the stuff is on YouTube. 0:15:07.360 --> 0:15:09.920 If you want to see Nick Cage as Indiana Jones, 0:15:10.040 --> 0:15:13.760 which is pretty funny, or Yoda, which is hilarious, or 0:15:13.800 --> 0:15:16.560 if you want to see Mona Lisa talking, it looks 0:15:16.640 --> 0:15:19.640 it looks fairly realistic, like hey, they brought that painting 0:15:19.640 --> 0:15:22.200 to life, right yeah, yeah, And if you scroll down 0:15:22.240 --> 0:15:24.360 a little bit, they did one with Marilyn Monro too, 0:15:24.840 --> 0:15:28.440 they brought her to life. He did interesting. Well, I 0:15:28.440 --> 0:15:33.840 mean this is like just set up for for TV commercials. Yes, 0:15:33.880 --> 0:15:36.400 they've already done stuff like this, right, This is like 0:15:36.440 --> 0:15:40.480 fredis Stared dancing with a dirt devil or whatever that was. 0:15:40.640 --> 0:15:45.920 What this could bring is creating entirely new movies and 0:15:45.960 --> 0:15:49.400 bringing back dead actors and actresses or I guess as 0:15:49.440 --> 0:15:53.080 actors now these days. Right, Uh well, I mean you've 0:15:53.080 --> 0:15:54.960 seen some of them. Are you talking about the the 0:15:55.040 --> 0:15:59.320 aging people are just creating, like bringing back someone that's 0:15:59.360 --> 0:16:02.040 been long dead. No, No, I'm saying like you call 0:16:02.160 --> 0:16:04.920 actors and actresses just actors these days. Oh that part. 0:16:05.720 --> 0:16:09.560 Uh you can do whatever you want, okay, Uh, motion 0:16:09.600 --> 0:16:13.440 picture performers, maybe even television performers, but bringing them back 0:16:13.480 --> 0:16:16.400 and giving them like they could star in an entirely 0:16:17.680 --> 0:16:21.080 because it's so realistic in life. Yeah. Um, they're not 0:16:21.120 --> 0:16:23.920 at that point yet because they're just now getting to 0:16:24.000 --> 0:16:28.240 where the d aging looks decent depending on who it is. Uh, 0:16:28.280 --> 0:16:32.920 Like the Sam Jackson stuff and Captain Marvel, Uh, my 0:16:32.960 --> 0:16:36.040 friend looked pretty good. It looked amazing. Yeah. And this 0:16:36.160 --> 0:16:39.080 Will Smith stuff in this new angry movie looks really good. 0:16:39.240 --> 0:16:42.920 What's that one? He's an angry movie where he plays 0:16:42.960 --> 0:16:48.640 a some sort of assassin that uh Aladdin? Yeah, that's it. 0:16:49.040 --> 0:16:51.720 Have you seen that? It was good? No? I have 0:16:51.760 --> 0:16:55.680 no interest, um to go back and kill with the 0:16:55.760 --> 0:16:57.480 younger version of himself or the young ones trying to 0:16:57.560 --> 0:16:59.040 kill the older one. Is what it is. Man. It's 0:16:59.040 --> 0:17:02.760 a lot like Looper, uh sort of, but it looks 0:17:02.760 --> 0:17:06.600 pretty good. Like it looks like young Will Smith. Slightly uncanny, 0:17:07.040 --> 0:17:08.879 but not as bad as I think. Some people are 0:17:08.920 --> 0:17:11.200 easier than others, like the Michael Douglas stuff and aunt 0:17:11.280 --> 0:17:14.000 Man and the marble stuff is kind of creepy looking. 0:17:14.119 --> 0:17:16.320 I haven't seen that. I mean I've seen parts of it, 0:17:16.359 --> 0:17:18.600 but I didn't notice that they were d aging Michael Douglas. 0:17:18.680 --> 0:17:21.400 They took Michael Douglas back in scenes to like the seventies. 0:17:22.040 --> 0:17:25.520 It's just like it doesn't look great. But anyway, that's 0:17:25.520 --> 0:17:28.200 sort of off track, um, but not really. I mean, 0:17:28.200 --> 0:17:30.240 it's kind of similar type of stuff. I guess, no 0:17:30.320 --> 0:17:32.879 more than that Gizmoto article to start. Yeah, that's a 0:17:32.880 --> 0:17:36.000 good point, um, But the whole reason we should point 0:17:36.000 --> 0:17:39.960 out that people are doing this stuff with celebrities and 0:17:40.000 --> 0:17:42.200 stuff like that, it's just because there's more data out there. 0:17:42.440 --> 0:17:45.600 It's a lot easier when you have a gazillion pictures 0:17:45.600 --> 0:17:49.240 of Brad Pitt on the internet to do a fake pit. Yeah, 0:17:49.280 --> 0:17:52.560 because the data set that the AI is trained on 0:17:52.760 --> 0:17:56.120 is just more and more robust. And the more pictures are, 0:17:56.160 --> 0:17:59.560 the more angles the AI has seen Brad Pitt, you know, 0:17:59.640 --> 0:18:03.560 look round at, and so can recreate this these faces 0:18:03.640 --> 0:18:08.040 because the AI seem like every possible pose or expression 0:18:08.119 --> 0:18:11.239 or whatever Brad Pitts ever made. But the thing is 0:18:11.280 --> 0:18:14.760 that mona Lisa and Marilyn Monroe. Thing that Samsung showed, 0:18:15.320 --> 0:18:18.960 They showed that you could make a pretty convincing deep 0:18:19.000 --> 0:18:23.560 fake with just one picture, one pose, right, So that's 0:18:23.560 --> 0:18:26.359 a big deal. But again, the bigger the data set, 0:18:26.400 --> 0:18:30.040 the better. And that's why, like you said, celebrities and 0:18:30.040 --> 0:18:35.280 and UM world leaders were the earliest targets. But over time, 0:18:35.760 --> 0:18:38.520 with the advent of other software and the fact that 0:18:38.560 --> 0:18:42.159 people now post tons of stuff about themselves and pictures 0:18:42.160 --> 0:18:46.439 of themselves on social media, um, it's become easier and 0:18:46.480 --> 0:18:51.040 easier to make a deep fake video of anybody. There's 0:18:51.080 --> 0:18:55.199 like holes, their software that scrapes social media accounts for 0:18:55.280 --> 0:19:00.240 every picture and video that's been posted ants, right, um distinction, Yeah, 0:19:00.240 --> 0:19:04.320 for sure. Um there's and then there's um other other 0:19:04.440 --> 0:19:07.960 sites and other apps that say, oh this this picture 0:19:07.960 --> 0:19:10.879 of this person. You're targeting, your your classmate or whatever. 0:19:11.480 --> 0:19:14.640 They probably have a pretty good match with this porn star, 0:19:14.760 --> 0:19:17.159 So go find videos of this porn star. And then 0:19:17.200 --> 0:19:18.920 the next thing, you know, you run it through that app, 0:19:19.000 --> 0:19:21.960 that deep fake app came out with, and you've got 0:19:21.960 --> 0:19:25.560 yourself a deep fake video and you've officially become a 0:19:25.600 --> 0:19:28.800 bad person. All right, that's a good place to take 0:19:28.800 --> 0:19:32.119 a break, and uh, we'll talk more about these bad people, 0:19:32.200 --> 0:19:53.919 right for this that's what sk as good should all right, 0:19:54.040 --> 0:19:58.359 So you mentioned before the break, Uh, this person's face 0:19:58.520 --> 0:20:01.800 that I just stole off the internet fits this porn 0:20:01.840 --> 0:20:06.320 actors body, which is a it's a consideration if you're 0:20:06.359 --> 0:20:10.080 making that, because to look right, that has to they 0:20:10.080 --> 0:20:13.160 have to bear a passing resemblance. I think that's right. Okay, 0:20:13.520 --> 0:20:16.600 So I was just about to say, so, yeah, So 0:20:16.680 --> 0:20:20.119 what they're doing now is they're browsing these applications with 0:20:20.200 --> 0:20:24.040 facial recognition software to make this a lot easier. And 0:20:24.080 --> 0:20:27.000 that's what most of this is about, is like, let's 0:20:27.000 --> 0:20:29.640 just see how easy we can make this and how 0:20:30.280 --> 0:20:33.440 how much we can democratize this where any shmoke can 0:20:33.480 --> 0:20:36.560 take any single photo and do the worst things possible 0:20:36.560 --> 0:20:41.840 with it, but also how convincing they've become as well. Yeah, 0:20:42.000 --> 0:20:44.120 I mean, so it's another big change. It's looking better 0:20:44.160 --> 0:20:47.800 and better, quicker and quicker, which is pretty scary. Did 0:20:47.840 --> 0:20:50.879 you see the Obama one. Had it not been for 0:20:51.480 --> 0:20:55.000 had it not been for Jordan's Peel's voice obviously being 0:20:55.000 --> 0:20:57.760 not Obama, I would have been like, Wow, this is 0:20:57.800 --> 0:21:00.800 really convincing. Really, Yeah, see how didn't think the lips 0:21:00.840 --> 0:21:03.600 matched up at all? Oh? I thought it looked pretty close. Yeah. 0:21:03.680 --> 0:21:06.639 So what we're talking about is Jordan Peel did a 0:21:06.640 --> 0:21:10.240 basically a demonstration video to raise awareness about how awful 0:21:10.280 --> 0:21:13.760 this is by doing it himself, and did a video 0:21:13.760 --> 0:21:17.520 of Obama like you know, referring to Trump as a 0:21:17.520 --> 0:21:21.199 curse word, a dipstick and basically saying like, hey, this 0:21:21.280 --> 0:21:24.320 is Obama and what you know people are doing. He 0:21:24.359 --> 0:21:28.000 basically is describing what's happening as you're watching it. And 0:21:28.119 --> 0:21:29.919 I thought it look kind of fake. He's describing a 0:21:29.960 --> 0:21:32.960 deep fake through deep fake in Jordan Peel, and he 0:21:33.000 --> 0:21:35.640 did it in conjunction with BuzzFeed and another production company. 0:21:35.880 --> 0:21:37.959 But in their defense, they were making this in like 0:21:38.040 --> 0:21:42.080 early two thou eighteen, like April two eighteen, and since then, 0:21:42.160 --> 0:21:45.920 even more technologies come out that is dedicated to matching 0:21:46.480 --> 0:21:49.520 the movement of the mouth to whatever words you want 0:21:49.520 --> 0:21:52.800 the person to say. Yeah, and you can also like 0:21:53.119 --> 0:21:55.760 use only parts of it, so it's even more convincing. 0:21:56.320 --> 0:21:59.920 So like if Obama had a lead in that actually worked, 0:22:00.000 --> 0:22:02.280 you could just keep that in there and then take 0:22:02.320 --> 0:22:05.280 out certain words, and you can manipulate it however you 0:22:05.320 --> 0:22:07.719 want to write, and the AI can go through and 0:22:07.760 --> 0:22:10.000 find like phone names and stuff like that to make 0:22:10.119 --> 0:22:14.360 the new words that the person never said. It's it's 0:22:14.359 --> 0:22:17.040 becoming extremely easy. Let's just put it like this. It's 0:22:17.080 --> 0:22:21.359 becoming extremely easy, and it's um widely available for anybody 0:22:21.560 --> 0:22:25.080 to make a video of somebody doing something or saying 0:22:25.119 --> 0:22:27.840 something that they never did or never said, and to 0:22:27.920 --> 0:22:32.960 make it convincing enough that you may believe it at first. Yeah, which, 0:22:32.960 --> 0:22:35.760 like we said that, you know, the obvious scariest implications 0:22:36.200 --> 0:22:40.320 that aren't just of the personal variety are in politics, 0:22:40.920 --> 0:22:45.000 when you could create real fake news that actually put 0:22:45.200 --> 0:22:48.159 people in jeopardy. They were put the entire world in 0:22:48.240 --> 0:22:52.320