WEBVTT 00:00:00.000 --> 00:00:04.480 The art of deception is as old as time itself and covers everything from counterintelligence 00:00:04.480 --> 00:00:10.000 to Halloween masks. But one incredibly convincing method of subterfuge that's rapidly becoming 00:00:10.000 --> 00:00:16.880 more common is the deep fake. The idea behind a deep fake is that it creates images that strongly 00:00:16.880 --> 00:00:22.800 resemble a certain subject, usually a person, that isn't actually them. But I'm not talking about 00:00:22.800 --> 00:00:28.320 some kind of cheap obvious Photoshop job that involves cut pasting your head onto a wait-lifter 00:00:28.320 --> 00:00:34.400 friend's vacation photos. Instead, deep fakes can be nearly indistinguishable from the real thing 00:00:34.400 --> 00:00:39.920 to the average human eye thanks to the use of artificial intelligence and machine learning. 00:00:39.920 --> 00:00:46.080 So here's how it works. Programs that generate deep fakes use not one but two different AIs 00:00:46.080 --> 00:00:52.560 working together. The first AI will scan many images of the subject to be faked and then create 00:00:52.560 --> 00:00:59.680 new, faked images. The second AI will then examine these fakes and compare them to real images, 00:00:59.680 --> 00:01:04.960 and if the differences are too stark, the second AI will mark the image as an obvious fake and 00:01:04.960 --> 00:01:11.280 tell the first AI. So the first AI takes this information and continually adjusts the fake 00:01:11.280 --> 00:01:16.960 images until the second AI's error rate hits a certain target. That is to say, until the second 00:01:16.960 --> 00:01:22.880 AI can't tell a fake from the real thing anymore. This system is called a generative adversarial 00:01:22.880 --> 00:01:28.240 network and although the idea behind it is fairly simple, advances in processing power have made 00:01:28.240 --> 00:01:34.560 it a very powerful tool for producing convincing looking fakes. And because it's a fairly general 00:01:34.560 --> 00:01:41.440 purpose technology, it has lots of cool applications. For example, an AI network like this can analyze 00:01:41.440 --> 00:01:47.840 lots of people with a similar look to create a fake model for advertising purposes. It's much 00:01:47.840 --> 00:01:52.240 cheaper than hiring a real model and paying them for a photo shoot, much of course to the chagrin of 00:01:52.240 --> 00:01:57.360 attractive yet unemployed people everywhere. Deep fakes could also help with age progression photos 00:01:57.360 --> 00:02:02.640 used by the police to help find missing people or even to upscale old school video games more 00:02:02.640 --> 00:02:08.240 realistically so they look nice on modern displays. But if you've seen stories about deep fakes in the 00:02:08.240 --> 00:02:13.280 news, you probably know that there's a much darker side to the technology. Unsurprisingly, 00:02:13.280 --> 00:02:18.400 the majority of the deep fake videos on the internet right now are pornographic in nature, 00:02:18.400 --> 00:02:23.040 with the idea often being to swap in a celebrity's face without their consent, 00:02:23.040 --> 00:02:27.760 resulting in several well-known actresses falling victim to the practice. And because 00:02:27.760 --> 00:02:32.480 deep fake technology has gotten good enough to create fake videos that make it appear as though 00:02:32.480 --> 00:02:37.120 people have been caught on camera saying things they haven't, seriously it can accurately match 00:02:37.120 --> 00:02:43.120 up lip movements to convincingly fake speech, there's a huge concern that it could be used 00:02:43.120 --> 00:02:48.560 to spread fake videos of politicians saying insightful or inflammatory things in order to 00:02:48.560 --> 00:02:52.960 push a certain agenda. Of course, this doesn't mean that everyone out there will be fooled, 00:02:52.960 --> 00:02:58.560 but with how realistic deep fakes can be and how polarized the political climate is in many parts 00:02:58.560 --> 00:03:04.880 of the world, political deep fakes could easily convince enough people. Recently, a video surfaced 00:03:04.880 --> 00:03:10.400 of Nancy Pelosi giving a speech while slurring her words apparently drunk. The video wasn't even 00:03:10.400 --> 00:03:15.920 a deep fake, it was merely slowed down, but it still racked up millions of views before finally 00:03:15.920 --> 00:03:21.680 being exposed as fake. And with accusations of political manipulation becoming more common these 00:03:21.680 --> 00:03:26.640 days, the rise of deep fake technology has led some observers to believe that it will have a 00:03:26.640 --> 00:03:32.960 tangible effect on political discourse and even election results. Although it remains to be seen 00:03:33.040 --> 00:03:38.640 if these fears actually come to pass, there are a number of cheaply accessible deep fake tools 00:03:38.640 --> 00:03:43.600 that allow people to create deep fakes without a ton of technical skills. So if you want to tinker 00:03:43.600 --> 00:03:49.600 with them yourself, I'd say just go ahead, but stay away from the politics and stick to swapping 00:03:49.600 --> 00:03:54.240 Linus Faces into your friends wedding video. So thanks for watching guys, like, dislike, 00:03:54.240 --> 00:03:58.080 check out our other videos, leave a comment if you have a suggestion for a future fast as possible, 00:03:58.080 --> 00:04:03.360 and don't forget to subscribe and follow. And that's not me being 00:04:03.360 --> 00:04:06.560 fakely manipulated to say that. I really think you should subscribe.