1 00:00:18,396 --> 00:00:21,116 Speaker 1: Noah Feldman here. I'm excited to tell you about a 2 00:00:21,156 --> 00:00:25,236 Speaker 1: special five part Deep Background mini series called Deep Bench. 3 00:00:25,956 --> 00:00:28,556 Speaker 1: The first episodes will appear in your feed on Saturday, 4 00:00:28,596 --> 00:00:32,316 Speaker 1: October seventeenth. The battle for the Supreme Court has become 5 00:00:32,356 --> 00:00:35,956 Speaker 1: a huge issue in the presidential election. In many ways, 6 00:00:36,236 --> 00:00:40,676 Speaker 1: it's a culmination of a conservative legal revolution spearheaded by 7 00:00:40,676 --> 00:00:44,956 Speaker 1: the Federalist Society. Deep Bench is the inside story of 8 00:00:45,036 --> 00:00:48,476 Speaker 1: how these legal conservatives gained power and how, at the 9 00:00:48,516 --> 00:00:52,556 Speaker 1: height of their influence, they're actually in danger of splitting apart. 10 00:00:53,436 --> 00:00:57,156 Speaker 1: But first, we're presenting an episode from Pushkin Industry's newest show, 11 00:00:57,556 --> 00:01:01,756 Speaker 1: Brave New Planet. Every day we see how powerful technologies 12 00:01:01,796 --> 00:01:06,156 Speaker 1: are advancing at a breathtaking pace. They have amazing potential upsides, 13 00:01:06,636 --> 00:01:09,356 Speaker 1: but if we're not careful, some might leave us a 14 00:01:09,356 --> 00:01:13,556 Speaker 1: lot worse off. In Brave New Planet, doctor Eric Lander 15 00:01:13,596 --> 00:01:16,036 Speaker 1: and his guests weigh the pros and cons of a 16 00:01:16,116 --> 00:01:20,916 Speaker 1: wide range of powerful innovations in science and technology. Doctor 17 00:01:20,996 --> 00:01:24,516 Speaker 1: Lander directs the Broad Institute of MIT and Harvard. He 18 00:01:24,636 --> 00:01:27,796 Speaker 1: was a leader of the Human Genome Project for eight years. 19 00:01:27,796 --> 00:01:31,156 Speaker 1: He served as a science advisor to President Obama's White House. 20 00:01:31,676 --> 00:01:36,596 Speaker 1: In this episode, doctor Lander explores deep fakes. Deep fakes 21 00:01:36,636 --> 00:01:40,156 Speaker 1: can be useful in art, education and therapy, but could 22 00:01:40,156 --> 00:01:43,596 Speaker 1: they be weaponized to provoke international conflicts or swing elections? 23 00:01:44,036 --> 00:01:46,036 Speaker 1: And where does the right to free speech fit in? 24 00:01:46,996 --> 00:01:50,716 Speaker 1: Every episode of Brave New Planet will grapple with opportunities 25 00:01:50,716 --> 00:01:53,316 Speaker 1: and challenges that are too big to fit in a tweet, 26 00:01:53,436 --> 00:01:57,316 Speaker 1: but will shape our future. You can subscribe in Apple Podcasts. 27 00:01:57,836 --> 00:02:08,796 Speaker 1: Here's the brilliant Eric Lander and Brave New Planet. Your 28 00:02:09,396 --> 00:02:13,436 Speaker 1: to Brave New Planet, a podcast about amazing new technologies 29 00:02:13,516 --> 00:02:17,076 Speaker 1: that could dramatically improve our world, or, if we don't 30 00:02:17,076 --> 00:02:19,796 Speaker 1: make wise choices, could leave us a lot worse off 31 00:02:20,596 --> 00:02:35,236 Speaker 1: Utopia or dystopia. It's up to us. On July sixteenth, 32 00:02:35,396 --> 00:02:40,476 Speaker 1: nineteen sixty nine, Apollo eleven blasted off from the Kennedy 33 00:02:40,516 --> 00:02:45,756 Speaker 1: Space Center near Cape Canaveral, Florida. Twenty five million Americans 34 00:02:45,836 --> 00:02:49,956 Speaker 1: watched on television as the spacecraft ascended toward the heavens, 35 00:02:50,476 --> 00:02:55,996 Speaker 1: carrying Commander Neil Armstrong, Lunar Module pilot Buzz Aldron, and 36 00:02:56,076 --> 00:03:00,716 Speaker 1: Command Module pilot Michael Collins their mission to be the 37 00:03:00,796 --> 00:03:04,556 Speaker 1: first humans in history to set foot on the Moon. 38 00:03:05,636 --> 00:03:10,236 Speaker 1: Four days later, on Sunday, July twentieth, The lunar module 39 00:03:10,356 --> 00:03:14,716 Speaker 1: separated from the command ship and soon fired its rockets 40 00:03:15,076 --> 00:03:21,636 Speaker 1: to begin its lunar descent. Five minutes later, disaster struck 41 00:03:22,556 --> 00:03:26,236 Speaker 1: about a mile above the Moon's surface. Program alarms twelve 42 00:03:26,276 --> 00:03:29,556 Speaker 1: oh one and twelve O two sounded loudly, indicating that 43 00:03:29,596 --> 00:03:36,996 Speaker 1: the mission computer was overloaded, and then well, every American 44 00:03:37,076 --> 00:03:51,396 Speaker 1: knows what happened next lost date of five good evening, 45 00:03:51,676 --> 00:03:56,196 Speaker 1: my fellow Americans. President Richard Nixon addressed a grieving nation. 46 00:03:57,316 --> 00:04:00,276 Speaker 1: Fates has ordained that the men who went to the 47 00:04:00,276 --> 00:04:04,516 Speaker 1: Moon to explore in peace will stay on the Moon 48 00:04:05,036 --> 00:04:13,556 Speaker 1: to rest in peace. These men, Neil Armstrong and Edwin Auburn, 49 00:04:14,916 --> 00:04:19,756 Speaker 1: know that there is no pope for their recovery, but 50 00:04:19,876 --> 00:04:23,996 Speaker 1: they also know that there is hook for mankind in 51 00:04:24,076 --> 00:04:29,236 Speaker 1: their sacrifice. He ended with the now famous words for 52 00:04:29,396 --> 00:04:31,996 Speaker 1: every human being who looks up at the Moon and 53 00:04:32,036 --> 00:04:35,156 Speaker 1: the nights to come, will know that there is some 54 00:04:35,276 --> 00:04:44,156 Speaker 1: pun or another word that is forever mankind. Wait a minute, 55 00:04:44,676 --> 00:04:48,636 Speaker 1: that never happened. The Moon mission was a historic success. 56 00:04:49,236 --> 00:04:53,316 Speaker 1: The three astronauts returned safely to ticker tape parades and 57 00:04:53,356 --> 00:04:58,316 Speaker 1: a celebratory thirty eight day World Tour. Those alarms actually 58 00:04:58,436 --> 00:05:02,836 Speaker 1: did sound, but they turned out to be harmless. Nixon 59 00:05:02,916 --> 00:05:06,716 Speaker 1: never delivered that speech. His speech writer had written it, 60 00:05:07,036 --> 00:05:10,476 Speaker 1: but it sat in a folder labeled an event of 61 00:05:10,716 --> 00:05:16,596 Speaker 1: Moon disaster until now. The Nixon you just heard is 62 00:05:16,596 --> 00:05:20,156 Speaker 1: a deep fake, part of a seven minute film created 63 00:05:20,156 --> 00:05:25,276 Speaker 1: by artificial intelligence deep learning algorithms. The fake was made 64 00:05:25,356 --> 00:05:29,476 Speaker 1: by the Center for Advanced Virtuality at the Massachusetts Institute 65 00:05:29,476 --> 00:05:32,716 Speaker 1: of Technology as part of an art exhibit to raise 66 00:05:32,756 --> 00:05:36,636 Speaker 1: awareness about the power of synthesized media. Not long ago, 67 00:05:37,276 --> 00:05:40,316 Speaker 1: something like this would have taken a lot of time 68 00:05:40,316 --> 00:05:43,836 Speaker 1: and money, But now it's getting easy. You can make 69 00:05:43,916 --> 00:05:47,436 Speaker 1: new paintings in the style of French Impressionism, revived dead 70 00:05:47,556 --> 00:05:52,316 Speaker 1: movie stars, help patience with ner degenerative disease, or soon 71 00:05:52,436 --> 00:05:55,396 Speaker 1: maybe take a class on a tour of ancient Rome. 72 00:05:55,836 --> 00:05:59,596 Speaker 1: But as the technology quickly becomes democratized, we're getting to 73 00:05:59,636 --> 00:06:03,396 Speaker 1: the point where almost anyone can create a fake video 74 00:06:03,436 --> 00:06:06,876 Speaker 1: of a friend, an ex lover, a stranger, or a 75 00:06:06,916 --> 00:06:12,316 Speaker 1: public figure that's embarrassing, pornographic, or perhaps capable of causing 76 00:06:12,436 --> 00:06:17,196 Speaker 1: international chaos. Some argue that in a culture where faked 77 00:06:17,196 --> 00:06:22,196 Speaker 1: news spreads like wildfire, and political leaders deny the veracity 78 00:06:22,236 --> 00:06:25,996 Speaker 1: of hard facts. Deep faked media may do a lot 79 00:06:26,116 --> 00:06:33,396 Speaker 1: more harm than good. Today's big question will synthesized media 80 00:06:33,556 --> 00:06:38,276 Speaker 1: unleash a new wave of creativity or will it erode 81 00:06:38,316 --> 00:06:42,476 Speaker 1: the already tenuous role of truth in our democracy? And 82 00:06:43,596 --> 00:06:46,316 Speaker 1: is there anything we can do to keep it in check. 83 00:06:55,196 --> 00:06:57,556 Speaker 1: My name is Eric Lander. I'm a scientist who works 84 00:06:57,596 --> 00:07:00,316 Speaker 1: on ways to improve human health. I helped lead the 85 00:07:00,396 --> 00:07:03,756 Speaker 1: Human Genome Project, and today I lead the Broad Institute 86 00:07:03,756 --> 00:07:08,036 Speaker 1: of MIT and Harvard. In the twenty first century, powerful 87 00:07:08,116 --> 00:07:11,956 Speaker 1: technologies have been peering at a breathtaking pace related to 88 00:07:11,956 --> 00:07:16,476 Speaker 1: the Internet, artificial intelligence, genetic engineering, and more. They have 89 00:07:16,716 --> 00:07:20,836 Speaker 1: amazing potential upsides, but we can't ignore the risks that 90 00:07:20,956 --> 00:07:24,716 Speaker 1: come with them. The decisions aren't just up to scientists 91 00:07:25,036 --> 00:07:28,756 Speaker 1: or politicians. Whether we like it or not, we all 92 00:07:28,756 --> 00:07:31,596 Speaker 1: of us are the stewards of a brave New planet. 93 00:07:32,116 --> 00:07:35,916 Speaker 1: This generation's choices will shape the future as never before. 94 00:07:38,476 --> 00:07:42,436 Speaker 1: Coming up on today's episode of Brave New Planet, I 95 00:07:42,556 --> 00:07:46,916 Speaker 1: speak with some of the leaders behind advances in synthesized media. 96 00:07:47,116 --> 00:07:50,236 Speaker 1: You could, certainly, by the way, generate stories that could 97 00:07:50,636 --> 00:07:54,276 Speaker 1: be fresh and interesting and new and personal for every child. 98 00:07:54,636 --> 00:07:58,476 Speaker 1: We got emails from people who were quadruplegic and they 99 00:07:58,476 --> 00:08:01,156 Speaker 1: asked us if we could make them dance. We hear 100 00:08:01,236 --> 00:08:04,876 Speaker 1: from experts about some of the frightening ways that bad 101 00:08:04,876 --> 00:08:08,756 Speaker 1: actors can use deep fakes. Creditors would chime in and say, 102 00:08:09,196 --> 00:08:11,996 Speaker 1: you can absolutely make a deep fake sex video of 103 00:08:12,036 --> 00:08:14,836 Speaker 1: your ex with thirty pictures. I've done it with twenty. 104 00:08:15,076 --> 00:08:16,876 Speaker 1: Here's the things that keep me up at night. Right 105 00:08:17,556 --> 00:08:20,436 Speaker 1: a video of Donald Trump saying I've launched nuclear weapons 106 00:08:20,436 --> 00:08:23,836 Speaker 1: against Iran. And before anybody gets around to figuring out 107 00:08:23,836 --> 00:08:25,956 Speaker 1: whether this is real or not, we have global nuclear 108 00:08:26,036 --> 00:08:30,476 Speaker 1: mountdown and we explore how we might prevent the worst abuses. 109 00:08:31,316 --> 00:08:37,076 Speaker 1: It's important that younger people advocate for the Internet that 110 00:08:37,116 --> 00:08:39,716 Speaker 1: they want. We have to fight for it. We have 111 00:08:39,836 --> 00:08:49,676 Speaker 1: to ask for different things. Stay with us, Chapter one, 112 00:08:50,276 --> 00:08:55,196 Speaker 1: Abraham Lincoln's Head. To begin to understand the significance of 113 00:08:55,276 --> 00:08:58,956 Speaker 1: deep fake technology, I went to San Francisco to speak 114 00:08:58,956 --> 00:09:02,476 Speaker 1: with a world expert on synthetic media. My name is 115 00:09:02,716 --> 00:09:07,796 Speaker 1: alexey Or sometimes called alyosha Eros, and I'm a professor 116 00:09:07,836 --> 00:09:11,716 Speaker 1: at UC Berkeley and Computer Science and Electrical Engineering Department. 117 00:09:12,236 --> 00:09:18,116 Speaker 1: My research is on computer vision, computer graphics, machine learning, 118 00:09:18,836 --> 00:09:24,636 Speaker 1: various aspects of artificial intelligence. Where'd you grow up? I 119 00:09:24,716 --> 00:09:28,596 Speaker 1: grew up in Saint Petersburg in Russia. I was one 120 00:09:28,636 --> 00:09:33,036 Speaker 1: of those geeky kids playing around with computers or dreaming 121 00:09:33,036 --> 00:09:40,196 Speaker 1: about computers. My first computer was actually the first Soviet 122 00:09:40,716 --> 00:09:44,676 Speaker 1: personal computer. So you actually are involved in making sort 123 00:09:44,676 --> 00:09:49,916 Speaker 1: of synthetic content, synthetic media. That's right. Alexei has invented 124 00:09:49,996 --> 00:09:53,876 Speaker 1: powerful artificial intelligence tools, but his lab also has a 125 00:09:53,916 --> 00:09:58,196 Speaker 1: wonderful ability to use computers to enhance the human experience. 126 00:09:58,876 --> 00:10:02,276 Speaker 1: I was struck by a remarkable video on YouTube created 127 00:10:02,316 --> 00:10:06,116 Speaker 1: by his team at Berkeley. So this was a project 128 00:10:06,196 --> 00:10:11,556 Speaker 1: that actually was done by my students who didn't even 129 00:10:11,596 --> 00:10:15,356 Speaker 1: think of this as anything but a silly little toy 130 00:10:15,436 --> 00:10:19,396 Speaker 1: project of trying to see if we could get a 131 00:10:19,516 --> 00:10:24,316 Speaker 1: geeky computer science student to move like a ballerina. In 132 00:10:24,356 --> 00:10:28,116 Speaker 1: the video, one of the students, Carolyn cham dances with 133 00:10:28,156 --> 00:10:31,796 Speaker 1: a skill and grace of a professional, despite never having 134 00:10:31,796 --> 00:10:35,956 Speaker 1: studied ballet. The idea is you take a source actor 135 00:10:36,036 --> 00:10:40,156 Speaker 1: like a ballerina. There is a way to detect the 136 00:10:40,316 --> 00:10:43,596 Speaker 1: limbs of the dancer have a kind of a skeleton 137 00:10:44,236 --> 00:10:48,956 Speaker 1: extracted and also have my student just move around and 138 00:10:49,036 --> 00:10:52,716 Speaker 1: do some geeky moves. And now we're basically just going 139 00:10:52,796 --> 00:10:58,596 Speaker 1: to try to sympathize the appearance of my student driven 140 00:10:58,956 --> 00:11:01,716 Speaker 1: by the skeleton of the ballerina. Put it all together, 141 00:11:01,876 --> 00:11:05,716 Speaker 1: and then we have our grad student dancing pirouets like 142 00:11:05,876 --> 00:11:12,276 Speaker 1: a ballerina through artificial intelligence. Carolyn's body is puppeteered by 143 00:11:12,316 --> 00:11:15,396 Speaker 1: the dancer. We weren't even going to publish it, but 144 00:11:15,476 --> 00:11:19,916 Speaker 1: we just released a video on YouTube called Everybody Dance Now, 145 00:11:20,356 --> 00:11:25,516 Speaker 1: and somehow it really touched the nerve. Well, there's been 146 00:11:25,556 --> 00:11:29,596 Speaker 1: an explosion recently a new ways to manipulate media. Alexei 147 00:11:29,676 --> 00:11:33,196 Speaker 1: notes that the idea itself isn't new, it has a 148 00:11:33,276 --> 00:11:37,236 Speaker 1: long history. I can't help but ask, given that you 149 00:11:37,236 --> 00:11:41,916 Speaker 1: come from Russia. One of the premier users of doctoring 150 00:11:41,956 --> 00:11:47,196 Speaker 1: photographs I think was Stalin, who used the ability to 151 00:11:47,236 --> 00:11:51,716 Speaker 1: manipulate images for political effect. How did they do that? 152 00:11:52,276 --> 00:11:54,236 Speaker 1: Can you think of examples of this and like, what 153 00:11:54,316 --> 00:12:01,156 Speaker 1: was the technology? Then? The urge to change photographs has 154 00:12:01,236 --> 00:12:05,196 Speaker 1: been around basically since the invention of photography. For example, 155 00:12:05,196 --> 00:12:08,796 Speaker 1: there is a photograph of Abraham Lincoln that still hangs 156 00:12:08,796 --> 00:12:13,836 Speaker 1: in may classrooms. That's fake. It's actually Calhoun with Lincoln's 157 00:12:13,836 --> 00:12:17,956 Speaker 1: head attached it. Alexei's referring to John C. Calhoun, the 158 00:12:17,996 --> 00:12:22,396 Speaker 1: South Carolina senator and champion of slavery. A Civil War 159 00:12:22,516 --> 00:12:27,636 Speaker 1: portrait artist superimposed a photo of Lincoln's head onto an 160 00:12:27,636 --> 00:12:32,396 Speaker 1: engraving of Calhoun's body because they thought Lincoln's gangly frame 161 00:12:32,956 --> 00:12:35,836 Speaker 1: wasn't dignified enough, and so they just said, okay, we 162 00:12:35,916 --> 00:12:39,436 Speaker 1: can use Calhoun. Let's slap the Lincoln's head on his body. 163 00:12:39,676 --> 00:12:42,956 Speaker 1: And then, of course, as soon as you go into 164 00:12:43,276 --> 00:12:46,796 Speaker 1: the twentieth century, as soon as you get to dictatorships, 165 00:12:47,196 --> 00:12:50,876 Speaker 1: this is a wonderful toy for a dictator to use. 166 00:12:51,356 --> 00:12:55,156 Speaker 1: So again, Stalin was a big fan of this. He 167 00:12:55,236 --> 00:12:59,716 Speaker 1: would get rid of people in photographs once they were 168 00:12:59,716 --> 00:13:02,676 Speaker 1: out of favor or once they got jailed or killed. 169 00:13:03,156 --> 00:13:09,436 Speaker 1: He would just basically get them scratched out with reasonably cruetics. 170 00:13:10,076 --> 00:13:13,316 Speaker 1: Hitler did it, Mao did it, Castro did it, Bresnev 171 00:13:13,436 --> 00:13:16,316 Speaker 1: did it. I'm sure US agencies have done it also. 172 00:13:16,836 --> 00:13:20,876 Speaker 1: We have always manipulated images with a desire to change history. 173 00:13:21,356 --> 00:13:24,356 Speaker 1: This is Honi for Reed. He's also a professor at 174 00:13:24,356 --> 00:13:27,276 Speaker 1: Berkeley and a friend of Alexei's. I'm a professor of 175 00:13:27,316 --> 00:13:31,196 Speaker 1: computer science and I'm an expert in digital forensics, where 176 00:13:31,196 --> 00:13:35,436 Speaker 1: Alexei works on making synthetic media. Honey has devoted his 177 00:13:35,516 --> 00:13:39,476 Speaker 1: career to identifying when synthetic media is being used to 178 00:13:39,556 --> 00:13:45,196 Speaker 1: fool people, that is spotting fakes. He regularly collaborates on 179 00:13:45,236 --> 00:13:49,356 Speaker 1: this mission with Alexei. So I met Alyosha efros ten 180 00:13:49,436 --> 00:13:54,036 Speaker 1: twenty years ago. He is a really incredibly creative and 181 00:13:54,316 --> 00:13:58,196 Speaker 1: clever guy, and he has done what I consider some 182 00:13:58,236 --> 00:14:00,716 Speaker 1: of the most interesting work in computer vision and computer 183 00:14:00,796 --> 00:14:04,476 Speaker 1: graphics over the last two decades. And if you really 184 00:14:04,516 --> 00:14:06,996 Speaker 1: want to do forensics, well, you have to partner with 185 00:14:06,996 --> 00:14:10,116 Speaker 1: somebody like Alyosha. You have to partner with world class 186 00:14:10,196 --> 00:14:13,756 Speaker 1: mind who knows how to think about the synthesiside so 187 00:14:13,956 --> 00:14:17,236 Speaker 1: that you can synthesize the absolute best content and then 188 00:14:17,356 --> 00:14:20,036 Speaker 1: think about how to detect it. I think it's interesting 189 00:14:20,076 --> 00:14:22,516 Speaker 1: that if you're somebody on the synthesis side and developing 190 00:14:22,596 --> 00:14:24,836 Speaker 1: the forensic there's a little bit of a jekylin hide there, 191 00:14:24,836 --> 00:14:27,676 Speaker 1: and I think it's really fascinating. You know, the idea 192 00:14:27,836 --> 00:14:32,396 Speaker 1: of altering photos, it's not entirely new How far back 193 00:14:32,436 --> 00:14:35,676 Speaker 1: does this go? So we used to have in the 194 00:14:35,716 --> 00:14:40,796 Speaker 1: days of Stalin, highly talented, highly skilled, time consuming, difficult 195 00:14:40,876 --> 00:14:46,116 Speaker 1: process of manipulating images, removing somebody, erasing something from the image, 196 00:14:46,436 --> 00:14:51,036 Speaker 1: splicing faces together. And then we moved into the digital age, 197 00:14:51,316 --> 00:14:55,036 Speaker 1: where now a highly talented digital artist could remove one 198 00:14:55,036 --> 00:14:57,356 Speaker 1: face and add another face, but it was still a 199 00:14:57,436 --> 00:15:01,636 Speaker 1: time consuming and required scale. In nineteen ninety four, the 200 00:15:01,676 --> 00:15:04,916 Speaker 1: makers of the movie Forrest Gump won an Oscar for 201 00:15:05,036 --> 00:15:09,676 Speaker 1: Visual Effects for their representations of the title character interacting 202 00:15:09,676 --> 00:15:14,476 Speaker 1: with historical figures like President John F. Kennedy gratulate. How 203 00:15:14,516 --> 00:15:18,316 Speaker 1: does it feel to be an all Americans? Very good gratulation? 204 00:15:18,516 --> 00:15:26,276 Speaker 1: How do you feel? I believe it. Now computers are 205 00:15:26,316 --> 00:15:28,236 Speaker 1: doing all of the heavy lifting of what used to 206 00:15:28,236 --> 00:15:31,836 Speaker 1: be relegated to talented artists. The average person now can 207 00:15:31,916 --> 00:15:35,276 Speaker 1: use sophisticated technology to not just capture the recording, but 208 00:15:35,356 --> 00:15:38,676 Speaker 1: also manipulate it and then distribute it. The tools used 209 00:15:38,676 --> 00:15:42,316 Speaker 1: to create synthetic media have grown by leaps and bounds, 210 00:15:42,476 --> 00:15:45,396 Speaker 1: especially in the past few years, and so now we 211 00:15:45,436 --> 00:15:49,356 Speaker 1: have technology broadly called deep fake, but more specifically should 212 00:15:49,356 --> 00:15:53,556 Speaker 1: be called synthesized content where you point an image or 213 00:15:53,596 --> 00:15:56,916 Speaker 1: a video or an audio to an AI or machine 214 00:15:56,996 --> 00:16:00,276 Speaker 1: learning system, and it will replace the face for you. 215 00:16:00,516 --> 00:16:01,916 Speaker 1: I mean it can do that in an image, it 216 00:16:01,916 --> 00:16:04,276 Speaker 1: can do that in a video, or it can synthesize 217 00:16:04,276 --> 00:16:10,956 Speaker 1: audio for you in a particular person's voice. It's becomes 218 00:16:10,996 --> 00:16:15,836 Speaker 1: straightforward to swap people's faces. There's a popular YouTube video 219 00:16:16,236 --> 00:16:20,236 Speaker 1: that features tech pioneer Elon Musk's adult face on a 220 00:16:20,276 --> 00:16:24,316 Speaker 1: baby's body, and there's a famous meme where actor Nicholas 221 00:16:24,396 --> 00:16:28,636 Speaker 1: Cage's face replaces those of leading movie actors, both male 222 00:16:28,676 --> 00:16:32,156 Speaker 1: and female. You can put words into people's mouths and 223 00:16:32,276 --> 00:16:35,996 Speaker 1: make them jump and dance and run. You can even 224 00:16:36,076 --> 00:16:40,356 Speaker 1: resurrect powerful figures and have them deliver a fake speech 225 00:16:40,836 --> 00:16:49,596 Speaker 1: about a fake tragedy. From an Altered History, Chapter two, 226 00:16:50,596 --> 00:16:55,916 Speaker 1: Creating Nixon. The text of Nixon's Moon disaster speech that 227 00:16:55,956 --> 00:16:58,356 Speaker 1: we heard at the top of the show is actually 228 00:16:58,396 --> 00:17:01,396 Speaker 1: not fake. As I mentioned, it was written for President 229 00:17:01,476 --> 00:17:05,676 Speaker 1: Nixon as a contingency speech and thankfully never had to 230 00:17:05,716 --> 00:17:08,956 Speaker 1: be delivered. It's an amazing piece of writing. It was 231 00:17:09,156 --> 00:17:13,156 Speaker 1: written by Bill Safire, who was one of Nixon's speech writers. 232 00:17:13,556 --> 00:17:17,116 Speaker 1: This is artist in journalist Francesco Panetta. She's the co 233 00:17:17,316 --> 00:17:21,916 Speaker 1: director of the Nixon Fake or MTS Moon Disaster Team. 234 00:17:22,556 --> 00:17:27,476 Speaker 1: She's also the creative director in MIT's Center for Advanced Virtuality. 235 00:17:27,836 --> 00:17:32,396 Speaker 1: I was doing experimental journalism at the Guardian newspaper. I 236 00:17:32,476 --> 00:17:35,556 Speaker 1: ran the Guardians Virtual Reality studio for the last three years. 237 00:17:35,836 --> 00:17:38,316 Speaker 1: The second half of the Moon Disaster team is sound 238 00:17:38,436 --> 00:17:42,796 Speaker 1: artist Halsey Bergund. My name is Halsey Bergund. I'm a 239 00:17:42,956 --> 00:17:45,796 Speaker 1: sound artist and technologist, and I've had a lot of 240 00:17:45,836 --> 00:17:49,956 Speaker 1: experience with lots of sorts of audio enhanced with technology, 241 00:17:50,236 --> 00:17:53,596 Speaker 1: though this is my first experience with synthetic media, especially 242 00:17:53,676 --> 00:17:57,156 Speaker 1: since I typically focus on authenticity of voice and now 243 00:17:57,196 --> 00:18:00,636 Speaker 1: I'm kind of doing the opposite. So together Halsey and 244 00:18:00,716 --> 00:18:04,556 Speaker 1: Francesca chose to automate a tragic moment in history that 245 00:18:04,836 --> 00:18:07,996 Speaker 1: never actually happened. I think it all started with it 246 00:18:08,036 --> 00:18:10,396 Speaker 1: being the fiftieth anniversary of the Moon landing last year, 247 00:18:10,556 --> 00:18:13,356 Speaker 1: and add on top of that an election cycle in 248 00:18:13,396 --> 00:18:16,556 Speaker 1: this country, and dealing with this information, which is obviously 249 00:18:17,196 --> 00:18:21,276 Speaker 1: very important in election cycles. It was like light bulbs 250 00:18:21,596 --> 00:18:25,076 Speaker 1: went on and we got very excited about pursuing it. 251 00:18:25,076 --> 00:18:28,436 Speaker 1: It's possible to make mediocre fakes pretty quickly and cheaply, 252 00:18:28,796 --> 00:18:32,596 Speaker 1: but Francesca and Halsey wanted high production values. So how 253 00:18:32,596 --> 00:18:36,476 Speaker 1: does one go about making a first rate fake presidential address. 254 00:18:36,956 --> 00:18:40,116 Speaker 1: There are two components. There's the visuals and there's the audio, 255 00:18:40,156 --> 00:18:44,396 Speaker 1: and they are completely different processes. So we decided to 256 00:18:44,436 --> 00:18:48,716 Speaker 1: go with a video dialogue replacement company called Kenny Ai, 257 00:18:49,156 --> 00:18:51,716 Speaker 1: who would do the visuals for us, and then we 258 00:18:51,796 --> 00:18:56,436 Speaker 1: decided to go with Respeecher, who are a dialogue replacement 259 00:18:56,476 --> 00:19:01,556 Speaker 1: company for the voice of Nixon. They tackled the voice first, 260 00:19:01,716 --> 00:19:04,476 Speaker 1: the more challenging of the two mediums. What we were 261 00:19:04,516 --> 00:19:07,596 Speaker 1: told to do was to get two to three hours 262 00:19:07,676 --> 00:19:10,876 Speaker 1: worth of Nixon talking. That was pretty easy because the 263 00:19:10,996 --> 00:19:14,996 Speaker 1: Nixon Library has hours and hours of Nixon, mainly giving 264 00:19:15,076 --> 00:19:19,276 Speaker 1: Vietnam's speeches. The Communist armies of North Vietnam launched a 265 00:19:19,356 --> 00:19:23,036 Speaker 1: massive inversion of South Vietnam. That audio was then chopped 266 00:19:23,116 --> 00:19:26,916 Speaker 1: up into chunks between one and three seconds long. We 267 00:19:27,036 --> 00:19:32,676 Speaker 1: found this incredibly patient actor called Lewis D. Wheeler. Lewis 268 00:19:32,716 --> 00:19:36,396 Speaker 1: would listen to the one second clip and then he 269 00:19:37,036 --> 00:19:45,796 Speaker 1: would repeat that and do what I believe was right. Speech. 270 00:19:45,876 --> 00:19:47,996 Speaker 1: Would say to us things like we need to change 271 00:19:48,036 --> 00:19:52,636 Speaker 1: the diagonal attention, which meant nothing to us. Yes, we 272 00:19:52,676 --> 00:19:59,196 Speaker 1: have a whole lot of potential band names going forward. Yeah, 273 00:19:59,276 --> 00:20:02,436 Speaker 1: Synthetic Nixon is another good one. So once we have 274 00:20:02,836 --> 00:20:06,236 Speaker 1: our Nixon model made out of these thousands of tiny clips, 275 00:20:06,596 --> 00:20:10,956 Speaker 1: it means that whatever our act says will come out 276 00:20:11,076 --> 00:20:14,476 Speaker 1: then in Nixon's voice. So then what we did was 277 00:20:14,596 --> 00:20:19,196 Speaker 1: record the contingency speech of Nixon, and it meant that 278 00:20:19,236 --> 00:20:24,796 Speaker 1: we got Lewis's actually performance, but in Nixon's voice. What 279 00:20:24,876 --> 00:20:27,996 Speaker 1: about the video part, I mean, the video was much easier. 280 00:20:27,996 --> 00:20:30,036 Speaker 1: We're talking a couple of days here and a tiny 281 00:20:30,076 --> 00:20:35,676 Speaker 1: amount of data just with Lewis's iPhone. We filmed him 282 00:20:35,716 --> 00:20:39,436 Speaker 1: reading the contingency speech once a couple of minutes of 283 00:20:39,516 --> 00:20:42,956 Speaker 1: him just chatting to camera, and that was it fate 284 00:20:44,196 --> 00:20:47,316 Speaker 1: that the men who went to the Moon to explore 285 00:20:48,636 --> 00:20:52,796 Speaker 1: will stay on the moon. You know, we were told 286 00:20:52,796 --> 00:20:56,116 Speaker 1: by Kenny Ai that everything would be the same in 287 00:20:56,156 --> 00:20:59,476 Speaker 1: the video apart from just the area around the mouth. 288 00:20:59,916 --> 00:21:03,076 Speaker 1: So every gesture of the hand, every blink, every time 289 00:21:03,076 --> 00:21:05,876 Speaker 1: he moved his face, all of that would stay the same, 290 00:21:06,396 --> 00:21:11,036 Speaker 1: but just the mouth basically would change. So we used 291 00:21:11,276 --> 00:21:15,196 Speaker 1: Nixon's resignation speech. To have served in this office, it 292 00:21:15,356 --> 00:21:19,916 Speaker 1: is to have felt a very personal sense of It 293 00:21:19,996 --> 00:21:22,876 Speaker 1: was the speech of Nixon that looked to the most 294 00:21:22,916 --> 00:21:25,316 Speaker 1: somber way. He seemed to have the most emotion in 295 00:21:25,356 --> 00:21:28,636 Speaker 1: his face. So what actually went on in the computer? 296 00:21:29,516 --> 00:21:34,996 Speaker 1: Artificial intelligence sometimes sounds inscrutable, but the basic ideas are 297 00:21:35,076 --> 00:21:38,316 Speaker 1: quite simple. In this case, it uses a type of 298 00:21:38,356 --> 00:21:42,716 Speaker 1: computer program called an auto encoder. It's trained to take 299 00:21:42,876 --> 00:21:48,476 Speaker 1: complicated things, say spoken sentences or pictures, encode them in 300 00:21:48,516 --> 00:21:52,676 Speaker 1: a much simpler form, and then decode them to recover 301 00:21:52,716 --> 00:21:56,356 Speaker 1: the original as best it can. The encoder tries to 302 00:21:56,396 --> 00:22:00,036 Speaker 1: reduce things to their essence, throwing away most of the 303 00:22:00,116 --> 00:22:03,036 Speaker 1: information but keeping enough to do a good job of 304 00:22:03,076 --> 00:22:06,596 Speaker 1: reconstructing it to make a deep fake. Here's the trick. 305 00:22:07,316 --> 00:22:11,116 Speaker 1: Train a speech auto encode for Nixon to Nixon, and 306 00:22:11,236 --> 00:22:15,476 Speaker 1: a speech auto encoder for actor to actor, but force 307 00:22:15,596 --> 00:22:20,956 Speaker 1: them to use the same encoder. Then you can input 308 00:22:21,076 --> 00:22:26,036 Speaker 1: actor and decoded as Nixon. If you have enough data. 309 00:22:26,236 --> 00:22:32,156 Speaker 1: It's a piece of cake. Around their carefully created video, 310 00:22:32,516 --> 00:22:36,916 Speaker 1: the Moon Disaster team created an entire art installation a 311 00:22:36,996 --> 00:22:40,916 Speaker 1: nineteen sixties living room with a fake vintage newspaper sharing 312 00:22:40,916 --> 00:22:45,756 Speaker 1: the fake tragic news while a fake Nixon speak solemnly 313 00:22:46,036 --> 00:22:49,916 Speaker 1: on a vintage black and white television. Some people, when 314 00:22:49,996 --> 00:22:52,996 Speaker 1: they were watching the installation, they watched a number of times. 315 00:22:53,036 --> 00:22:54,996 Speaker 1: You'd see them, they'd watch it once, then they would 316 00:22:54,996 --> 00:22:57,916 Speaker 1: watch it again, staring at the lips to see if 317 00:22:57,956 --> 00:23:01,356 Speaker 1: they could see any lack of synchronicity. We had some 318 00:23:01,396 --> 00:23:05,316 Speaker 1: people who thought that perhaps Nixon had actually recorded this 319 00:23:05,556 --> 00:23:09,076 Speaker 1: speech as a contingency speech for it to go onto television. 320 00:23:09,516 --> 00:23:12,916 Speaker 1: Lots of folks who were listening, viewing, and even press 321 00:23:12,996 --> 00:23:15,636 Speaker 1: folks just immediately said, oh, the voice is real or whatever, 322 00:23:15,636 --> 00:23:19,076 Speaker 1: you know, said these things that weren't accurate because they 323 00:23:19,156 --> 00:23:21,596 Speaker 1: just felt like there wasn't even a question. I suppose 324 00:23:22,076 --> 00:23:24,036 Speaker 1: that is what we wanted to achieve. But at the 325 00:23:24,076 --> 00:23:26,676 Speaker 1: same time, it's it was a little bit eye opening 326 00:23:26,676 --> 00:23:29,436 Speaker 1: and like a little scary, you know that that could happen. 327 00:23:33,076 --> 00:23:39,876 Speaker 1: Chapter three, Everybody dance. What do you see as just 328 00:23:39,956 --> 00:23:45,636 Speaker 1: the wonderful upsides of having technologies like this? Yeah, I 329 00:23:45,676 --> 00:23:49,916 Speaker 1: mean ai A art is becoming a whole field in itself, 330 00:23:50,396 --> 00:23:54,996 Speaker 1: so creatively there is enormous potential. One of the potential 331 00:23:55,156 --> 00:23:58,396 Speaker 1: positive educational uses of deep fake technology would be to 332 00:23:58,516 --> 00:24:02,796 Speaker 1: bring historical figures back to life to make learning more durable. 333 00:24:02,876 --> 00:24:05,916 Speaker 1: I think one could do that with bringing Abraham Lincoln 334 00:24:05,916 --> 00:24:08,636 Speaker 1: back to life and having him deliver speeches. Film companies 335 00:24:08,636 --> 00:24:11,676 Speaker 1: and radios cited about re enactments. We're already beginning to 336 00:24:11,716 --> 00:24:15,796 Speaker 1: see this in films like Star Wars, when we're bringing 337 00:24:15,836 --> 00:24:18,556 Speaker 1: people like Carrie Fisher back to life. I mean that 338 00:24:18,716 --> 00:24:22,156 Speaker 1: is at the moment not being done through deep fake technologies. 339 00:24:22,156 --> 00:24:25,716 Speaker 1: This is using fairy traditional techniques of CGI at the moment. 340 00:24:25,916 --> 00:24:28,596 Speaker 1: So we still have to see our first deep fake 341 00:24:28,836 --> 00:24:32,396 Speaker 1: big cinema screen release, but this is just to come. 342 00:24:32,516 --> 00:24:35,516 Speaker 1: Like the technology is getting better and better. Not only 343 00:24:35,556 --> 00:24:38,596 Speaker 1: will we be able to potentially bring back actors and 344 00:24:38,636 --> 00:24:42,236 Speaker 1: actresses who are no longer alive and have them star movies, 345 00:24:42,276 --> 00:24:45,036 Speaker 1: but an actor could make a model of their own 346 00:24:45,116 --> 00:24:47,956 Speaker 1: voice and then sell the use of that voice to 347 00:24:47,996 --> 00:24:52,236 Speaker 1: anybody to do a voiceover of whatever is wanted, and 348 00:24:52,276 --> 00:24:54,716 Speaker 1: so they could have twenty of these going on at 349 00:24:54,716 --> 00:24:57,156 Speaker 1: the same time, and the sort of restriction of their 350 00:24:57,196 --> 00:25:00,716 Speaker 1: physical presence is no longer there. And that might mean 351 00:25:00,756 --> 00:25:03,756 Speaker 1: that you know, Brad Pitt is in everything, or it 352 00:25:03,876 --> 00:25:07,116 Speaker 1: might just mean that lower budget films can afford to 353 00:25:07,156 --> 00:25:10,036 Speaker 1: have some of the higher cost talent. That point, you know, 354 00:25:10,076 --> 00:25:13,356 Speaker 1: the top twenty actors could just do everything. Yes, there's 355 00:25:13,396 --> 00:25:15,796 Speaker 1: no doubt that there will be winners and losers from 356 00:25:16,076 --> 00:25:19,596 Speaker 1: these technologies, but the potential of synthetic media goes way 357 00:25:19,636 --> 00:25:23,876 Speaker 1: beyond the arts. There are possible medical and therapeutic applications. 358 00:25:24,236 --> 00:25:27,596 Speaker 1: There are companies that are working very hard to allow 359 00:25:27,676 --> 00:25:30,196 Speaker 1: people who have either lost their voice or who never 360 00:25:30,236 --> 00:25:32,716 Speaker 1: had a voice to be able to speak in a 361 00:25:32,996 --> 00:25:35,836 Speaker 1: way that is either how they used to speak or 362 00:25:36,556 --> 00:25:39,556 Speaker 1: in a way that isn't a canned voice that everybody has. 363 00:25:39,956 --> 00:25:43,796 Speaker 1: Alexei ePROs and his students discovered potential uses of synthetic 364 00:25:43,836 --> 00:25:48,996 Speaker 1: media and medicine quite unintentionally while working on their Everybody 365 00:25:49,116 --> 00:25:53,276 Speaker 1: Danced Now project that could turn anyone into a ballerina. 366 00:25:53,476 --> 00:25:57,836 Speaker 1: We were kind of surprised for all the positive feedback. 367 00:25:57,876 --> 00:26:01,556 Speaker 1: We god, we've got emails from people who were quadruplegic 368 00:26:01,636 --> 00:26:03,916 Speaker 1: and they asked us if we could make them dance, 369 00:26:03,996 --> 00:26:08,116 Speaker 1: and it was very unexpected. So now we are trying 370 00:26:08,156 --> 00:26:10,316 Speaker 1: to get the soft where to be in a state 371 00:26:10,316 --> 00:26:13,996 Speaker 1: where people can use that because yeah, it's somehow it 372 00:26:14,036 --> 00:26:20,756 Speaker 1: did hit a nerve with folks. Chapter four Unicorns in 373 00:26:20,796 --> 00:26:26,116 Speaker 1: the Andes. The past few years have seen amazing advances 374 00:26:26,196 --> 00:26:30,196 Speaker 1: in the creation of synthetic media through artificial intelligence. The 375 00:26:30,236 --> 00:26:33,836 Speaker 1: technology now goes far beyond fitting one face over another 376 00:26:33,876 --> 00:26:37,276 Speaker 1: face in a video. A recent breakthrough has made it 377 00:26:37,276 --> 00:26:43,196 Speaker 1: possible to create entirely new and very convincing content out 378 00:26:43,236 --> 00:26:49,356 Speaker 1: of thin air. The breakthrough, called generative adversarial networks or GAMS, 379 00:26:49,956 --> 00:26:53,796 Speaker 1: came from a machine learning researcher at Google named Ian Goodfellow. 380 00:26:54,596 --> 00:26:58,916 Speaker 1: Like auto encoders, the basic idea is simple but brilliant. 381 00:26:59,836 --> 00:27:03,596 Speaker 1: Suppose you want to create amazingly realistic photos of people 382 00:27:03,636 --> 00:27:07,356 Speaker 1: who don't exist. While you build a GAM consisting of 383 00:27:07,476 --> 00:27:12,276 Speaker 1: two computer programs, a photo generator that learns to generate 384 00:27:12,316 --> 00:27:17,436 Speaker 1: fake photos and a photo discriminator that learns to discriminate 385 00:27:17,516 --> 00:27:22,596 Speaker 1: or identify fake photos from a vast collection of real photos. 386 00:27:23,276 --> 00:27:26,996 Speaker 1: You then let the two programs compete, continually tweaking their 387 00:27:26,996 --> 00:27:30,636 Speaker 1: code to outsmart each other. By the time they're done, 388 00:27:31,036 --> 00:27:35,556 Speaker 1: the GAN can generate amazingly convincing fakes. You can see 389 00:27:35,596 --> 00:27:38,596 Speaker 1: for yourself if you go to the website this person 390 00:27:38,716 --> 00:27:43,476 Speaker 1: does not exist dot com. Every time you refresh the page, 391 00:27:43,836 --> 00:27:47,716 Speaker 1: you're shown a new uncanny image of a person who, 392 00:27:48,076 --> 00:27:51,836 Speaker 1: as the website says, does not and never did exist. 393 00:27:52,636 --> 00:28:00,796 Speaker 1: Francescan I actually tried out the website. This young Asian woman. 394 00:28:00,956 --> 00:28:04,436 Speaker 1: She's got great complexion, envious of that, neat black hair, 395 00:28:04,516 --> 00:28:08,276 Speaker 1: with a fringe pink lipstick and a slightly dreamy look 396 00:28:08,396 --> 00:28:14,116 Speaker 1: as she's kind of gazing off to her left. Oh, 397 00:28:14,156 --> 00:28:16,316 Speaker 1: here's a woman who looks like she could be a 398 00:28:16,356 --> 00:28:20,836 Speaker 1: neighbor of mine in Cambridge, probably about sixty five. She's 399 00:28:20,836 --> 00:28:26,276 Speaker 1: got nice wire framed glasses, layered hair. Her earrings don't 400 00:28:26,316 --> 00:28:30,196 Speaker 1: actually match, but that could just be her distinctive style. 401 00:28:30,996 --> 00:28:35,636 Speaker 1: I mean, of course, she doesn't really exist. It's hard 402 00:28:35,676 --> 00:28:40,236 Speaker 1: to argue that gans aren't creating original art. In fact, 403 00:28:40,636 --> 00:28:44,356 Speaker 1: an artist collective recently used a GAM to create a 404 00:28:44,396 --> 00:28:50,036 Speaker 1: French Impressionist style portrait. When Christie's sold it at auction, 405 00:28:50,596 --> 00:28:53,836 Speaker 1: it fetched an eye popping four hundred and thirty two 406 00:28:54,076 --> 00:28:59,996 Speaker 1: thousand dollars. Alexeiefros, the Berkeley professor, recently pushed gans a 407 00:29:00,076 --> 00:29:05,516 Speaker 1: step further, creating something called cycle gans by connecting two 408 00:29:05,636 --> 00:29:10,236 Speaker 1: gans together in a clever way. Cycle. Gans can transform 409 00:29:10,276 --> 00:29:14,156 Speaker 1: a money painting into what's seemingly a photograph of the 410 00:29:14,196 --> 00:29:18,396 Speaker 1: same scene, or turn a summer landscape into a winter 411 00:29:18,556 --> 00:29:23,356 Speaker 1: landscape of the same view. Alexei's psychogans seem like magic. 412 00:29:23,796 --> 00:29:27,836 Speaker 1: If you were to add in virtual reality, the possibilities 413 00:29:28,196 --> 00:29:35,236 Speaker 1: become mind blowing. You maybe reminiscing about walking down Saint 414 00:29:35,316 --> 00:29:38,756 Speaker 1: German and Paris, and with a few clicks, you are there, 415 00:29:38,876 --> 00:29:41,596 Speaker 1: and you're walking down the boulevard and you're looking at 416 00:29:41,596 --> 00:29:45,356 Speaker 1: all the buildings, and maybe you can even switch to 417 00:29:45,396 --> 00:29:49,596 Speaker 1: a different year. And I think that is I think 418 00:29:49,756 --> 00:29:54,796 Speaker 1: very exciting as a way to mentally travel to different places. 419 00:29:55,276 --> 00:29:57,676 Speaker 1: So if you do this in VR, I mean imagine 420 00:29:58,076 --> 00:30:01,876 Speaker 1: classes going on a class visit to ancient Rome. That's right. 421 00:30:02,196 --> 00:30:06,996 Speaker 1: You could imagine from how a particular city like Chrome 422 00:30:07,356 --> 00:30:11,196 Speaker 1: looks now. Trying to extrapolate we looked into past. It 423 00:30:11,276 --> 00:30:15,396 Speaker 1: turns out that gans aren't just transforming images. I spoke 424 00:30:15,436 --> 00:30:19,156 Speaker 1: with a friend who's very familiar with another remarkable application 425 00:30:19,196 --> 00:30:22,716 Speaker 1: of the technology. My name is Reid Hoffman. I'm a 426 00:30:22,716 --> 00:30:25,676 Speaker 1: podcaster of Master's a Scale. I'm a partner at Greylock, 427 00:30:25,716 --> 00:30:28,236 Speaker 1: which is where we're sitting right now co founder of 428 00:30:28,276 --> 00:30:32,716 Speaker 1: LinkedIn and then a variety of other eccentric hobbies. Reid 429 00:30:32,836 --> 00:30:37,236 Speaker 1: is a board member of an unusual organization called open ai. 430 00:30:37,676 --> 00:30:40,876 Speaker 1: Open a Eyes is highly concerned with artificial general intelligence 431 00:30:40,956 --> 00:30:45,036 Speaker 1: human level intelligence. I helped Sam Altman and Elon Musk 432 00:30:45,156 --> 00:30:50,436 Speaker 1: standing up. The basic concern was that if one company 433 00:30:50,516 --> 00:30:55,076 Speaker 1: created and deployed that that could be disbalancing in all 434 00:30:55,156 --> 00:30:58,116 Speaker 1: kinds of ways. And so the thought is, if it 435 00:30:58,196 --> 00:31:00,956 Speaker 1: could be created, we should make sure that there is 436 00:31:01,036 --> 00:31:04,356 Speaker 1: essentially a nonprofit that is creating this and that can 437 00:31:04,436 --> 00:31:09,756 Speaker 1: make that technology available at selective time slices to industry 438 00:31:09,756 --> 00:31:13,716 Speaker 1: as a whole government sut C. Last year, open ai 439 00:31:13,916 --> 00:31:18,556 Speaker 1: released a program that uses gams to write language from 440 00:31:18,556 --> 00:31:23,516 Speaker 1: a short opening prompt. The system, called GPT two, can 441 00:31:23,556 --> 00:31:27,156 Speaker 1: spin a convincing article or story instead of a deep 442 00:31:27,196 --> 00:31:32,076 Speaker 1: fake video. It's deep fake text. It's pretty amazing. Actually. 443 00:31:32,516 --> 00:31:37,836 Speaker 1: For example, open ai researchers gave the program the following prompt. 444 00:31:38,716 --> 00:31:42,036 Speaker 1: In a shocking finding, scientists discovered a herd of unicorns 445 00:31:42,076 --> 00:31:45,676 Speaker 1: living in a remote, previously unexplored valley in the Andes Mountains. 446 00:31:46,316 --> 00:31:49,316 Speaker 1: Even more surprising to the researchers was the fact that 447 00:31:49,316 --> 00:31:53,756 Speaker 1: the unicorns spoke perfect English. GPT two took it from there, 448 00:31:53,956 --> 00:31:59,356 Speaker 1: delivering nine crisp paragraphs on the landmark discovery. I asked 449 00:31:59,396 --> 00:32:02,956 Speaker 1: Fran to read a bit from the story. Doctor Jorge Perez, 450 00:32:03,156 --> 00:32:06,956 Speaker 1: an evolutionary biologist from the University of Lapaz, and several 451 00:32:06,996 --> 00:32:10,796 Speaker 1: companions were exploring the these mountains when they found a 452 00:32:10,836 --> 00:32:15,156 Speaker 1: small valley with no other animals or humans. Perez noticed 453 00:32:15,196 --> 00:32:17,596 Speaker 1: that the valley had what appeared to be a natural 454 00:32:17,676 --> 00:32:21,236 Speaker 1: fountains surrounded by two peaks of rock and silver snow. 455 00:32:22,036 --> 00:32:25,076 Speaker 1: Perez and the others then ventured further into the valley. 456 00:32:25,596 --> 00:32:27,436 Speaker 1: By the time we reached the top of one peak, 457 00:32:27,516 --> 00:32:30,556 Speaker 1: the water looked blue with some crystals on top, said Perez. 458 00:32:30,916 --> 00:32:34,076 Speaker 1: Perez and his friends were astonished to see the unicorn herd. 459 00:32:39,156 --> 00:32:40,556 Speaker 1: Don't tell me some of the great things you can 460 00:32:40,596 --> 00:32:45,236 Speaker 1: do with language generation, well, say, for example, entertainment, Generate 461 00:32:45,356 --> 00:32:48,796 Speaker 1: stories that could be fresh and interesting and new and 462 00:32:48,836 --> 00:32:53,596 Speaker 1: personal for every child. Embed educational things in those stories 463 00:32:53,596 --> 00:32:56,356 Speaker 1: so they're drawn into the fact that the story is 464 00:32:56,396 --> 00:32:59,916 Speaker 1: involving them and their friends, but also now brings in 465 00:33:00,076 --> 00:33:04,716 Speaker 1: grammar and math and other kinds of things. As are 466 00:33:04,756 --> 00:33:09,716 Speaker 1: doing it. Generate explanatory material of this kind of education 467 00:33:10,276 --> 00:33:13,316 Speaker 1: that works best for this audience, for this kind of people, 468 00:33:13,356 --> 00:33:14,836 Speaker 1: like we want to have this kind of math or 469 00:33:14,876 --> 00:33:17,076 Speaker 1: this kind of physics, or this kind of history or 470 00:33:17,076 --> 00:33:19,996 Speaker 1: this kind of poetry explained in the right way. And 471 00:33:20,076 --> 00:33:23,596 Speaker 1: also the sile language right like you know native city 472 00:33:23,796 --> 00:33:27,596 Speaker 1: x language. When open ai announced its breakthrough program for 473 00:33:27,676 --> 00:33:31,796 Speaker 1: text generation, it took the unusual step of not releasing 474 00:33:31,796 --> 00:33:34,556 Speaker 1: the full powered version because it was worried about the 475 00:33:34,556 --> 00:33:38,476 Speaker 1: possible consequences. Now, part of the open ai decision to 476 00:33:38,516 --> 00:33:41,916 Speaker 1: say we're going to release a smaller model than the 477 00:33:41,916 --> 00:33:44,836 Speaker 1: one we did is because we think that the deep 478 00:33:44,876 --> 00:33:47,676 Speaker 1: fake problem hasn't been solved. And by the way, some 479 00:33:47,716 --> 00:33:49,876 Speaker 1: people complained about that because they said, well, you're slowing 480 00:33:49,916 --> 00:33:52,516 Speaker 1: down our ability to do progress and so far. The 481 00:33:52,516 --> 00:33:55,036 Speaker 1: answer to say, look, when these are released to the 482 00:33:55,276 --> 00:33:58,396 Speaker 1: entire public, we cannot control the downside as well as 483 00:33:58,436 --> 00:34:06,556 Speaker 1: the upsides. Downsides from art to therapy to virtual time travel, 484 00:34:06,876 --> 00:34:12,756 Speaker 1: personalized stories and education, Synthetic media has amazing upsides. What 485 00:34:13,476 --> 00:34:20,316 Speaker 1: could possibly go wrong? Chapter five? What could possibly go wrong? 486 00:34:21,876 --> 00:34:25,556 Speaker 1: The downsides are actually not hard to find. The ability 487 00:34:25,556 --> 00:34:30,836 Speaker 1: to reshape reality brings extraordinary power, and people inevitably use 488 00:34:30,956 --> 00:34:35,476 Speaker 1: power to control other people. It should be no surprise, therefore, 489 00:34:35,516 --> 00:34:39,236 Speaker 1: that ninety six percent of fake videos posted online are 490 00:34:39,356 --> 00:34:45,236 Speaker 1: non consensual pornography videos, almost always of women manipulated to 491 00:34:45,356 --> 00:34:49,796 Speaker 1: depict sex acts that never actually occurred. I spoke with 492 00:34:49,836 --> 00:34:53,716 Speaker 1: a professor who studies deep fakes, including digital attempts to 493 00:34:53,836 --> 00:34:57,676 Speaker 1: control women's bodies. I'm Danielle Citron, and I am a 494 00:34:57,756 --> 00:35:01,036 Speaker 1: law professor at Boston University School of Law. I write 495 00:35:01,076 --> 00:35:06,396 Speaker 1: about privacy, technology, automation. My newest work and my next 496 00:35:06,396 --> 00:35:09,636 Speaker 1: book is going to be about sexual privacy. So I've 497 00:35:09,676 --> 00:35:13,196 Speaker 1: worked in and around consumer privacy, individual rights, civil rights. 498 00:35:13,236 --> 00:35:16,756 Speaker 1: I write a lot about free speech and then automated systems. 499 00:35:17,116 --> 00:35:20,676 Speaker 1: When did you first become aware of deep fakes? Do 500 00:35:20,676 --> 00:35:23,076 Speaker 1: you remember when this cross your rit? I did so. 501 00:35:23,596 --> 00:35:26,796 Speaker 1: There was a Reddit thread devoted to, you know, fake 502 00:35:27,156 --> 00:35:31,596 Speaker 1: pornography movies of Galjado Emma Watson. But the reddit thread 503 00:35:32,036 --> 00:35:35,716 Speaker 1: sort of spooled not just from celebrities but ordinary people, 504 00:35:36,236 --> 00:35:38,636 Speaker 1: and so you had rereditors asking each other, how do 505 00:35:38,676 --> 00:35:40,956 Speaker 1: I make a deep fake sex video of max girlfriend? 506 00:35:40,956 --> 00:35:43,836 Speaker 1: I have thirty pictures, and then other reditors would chime 507 00:35:43,836 --> 00:35:46,876 Speaker 1: in and say, look at this YouTube tutorial. You can 508 00:35:46,956 --> 00:35:49,916 Speaker 1: absolutely make a deep fake sex video of your ex 509 00:35:50,356 --> 00:35:53,316 Speaker 1: with thirty pictures. I've done it with twenty. In November 510 00:35:53,436 --> 00:35:59,276 Speaker 1: twenty seventeen, an anonymous redditor began posting synthesized porn videos 511 00:35:59,476 --> 00:36:03,356 Speaker 1: under the pseudonym deep fakes, perhaps a nod to the 512 00:36:03,436 --> 00:36:06,516 Speaker 1: deep learning technology you used to create them as well 513 00:36:06,556 --> 00:36:11,996 Speaker 1: as the nineteen seventies porn film throat. The Internet quickly 514 00:36:12,036 --> 00:36:16,716 Speaker 1: adopted the term deep fakes and broadened its meanings beyond pornography. 515 00:36:17,356 --> 00:36:20,676 Speaker 1: To create the videos, he used celebrity faces from Google 516 00:36:20,756 --> 00:36:25,276 Speaker 1: image search and YouTube videos, and then trains an algorithm 517 00:36:25,276 --> 00:36:29,876 Speaker 1: on that content together with pornographic videos. Have you seen 518 00:36:30,556 --> 00:36:34,916 Speaker 1: deep fake pornography videos? Yes, so still pretty crude, so 519 00:36:34,996 --> 00:36:38,556 Speaker 1: you probably can tell that it's a fake, But for 520 00:36:38,596 --> 00:36:43,156 Speaker 1: the person who's inserted into pornography, it's devastating. You use 521 00:36:43,276 --> 00:36:48,756 Speaker 1: the neural network technology, the artificial intelligence technology to create 522 00:36:48,956 --> 00:36:54,596 Speaker 1: out of digital whole cloth pornography videos using probably real 523 00:36:54,636 --> 00:36:58,356 Speaker 1: pornography and then inserting the person in the pornography so 524 00:36:58,436 --> 00:37:01,196 Speaker 1: they become the female actress. If it's a female, it's 525 00:37:01,276 --> 00:37:06,276 Speaker 1: usually a female in that video. My name is Noel 526 00:37:06,476 --> 00:37:14,236 Speaker 1: Martin and I am an activist and Lauraform campaigner in Australia. 527 00:37:14,476 --> 00:37:18,876 Speaker 1: Noel is twenty six years old and she lives in Perth, Australia. 528 00:37:19,156 --> 00:37:25,796 Speaker 1: So the first time that I discovered myself on pornographic 529 00:37:25,876 --> 00:37:31,956 Speaker 1: sites was when I was eighteen and out of curiosity, 530 00:37:32,036 --> 00:37:36,516 Speaker 1: decided to Google image reverse search myself. In an instant, 531 00:37:37,276 --> 00:37:41,796 Speaker 1: like in a less than a millisecond, my life completely changed. 532 00:37:42,196 --> 00:37:45,916 Speaker 1: At first, it started with photos still images stolen from 533 00:37:45,996 --> 00:37:51,396 Speaker 1: Noel's social media accounts. They were then doctoring my face 534 00:37:51,756 --> 00:37:57,996 Speaker 1: from ordinary images and superimposing those onto the bodies of 535 00:37:58,036 --> 00:38:02,516 Speaker 1: women depicting me having sexual intercourse. It proved impossible to 536 00:38:02,556 --> 00:38:06,316 Speaker 1: identify who was manipulating Noel's image in this way. It's 537 00:38:06,356 --> 00:38:09,316 Speaker 1: still unclear today, which made it difficult for her to 538 00:38:09,356 --> 00:38:13,516 Speaker 1: seek legal action. I went to the police soon after, 539 00:38:14,236 --> 00:38:20,636 Speaker 1: I contacted government agencies, tried getting a private investigator. Essentially, 540 00:38:20,956 --> 00:38:24,356 Speaker 1: there's nothing that they could do. The sites are hosted overseas, 541 00:38:24,676 --> 00:38:29,076 Speaker 1: the perpetrators are probably overseas. The reaction was at the 542 00:38:29,156 --> 00:38:31,836 Speaker 1: end of the day, I think you can contact the 543 00:38:31,916 --> 00:38:34,836 Speaker 1: webmasters to try and get things deleted. You know, you 544 00:38:34,876 --> 00:38:38,156 Speaker 1: can adjust your privacy setting so that you know nothing 545 00:38:38,276 --> 00:38:43,436 Speaker 1: is available to anyone publicly. It was an unwinnable situation. 546 00:38:44,036 --> 00:38:48,076 Speaker 1: Then things started to escalate. In twenty eighteen, who all 547 00:38:48,156 --> 00:38:52,316 Speaker 1: saw a synthesized pornographic video of herself And I believe 548 00:38:52,396 --> 00:38:56,556 Speaker 1: that it was done for the purposes of silencing me 549 00:38:57,356 --> 00:39:02,836 Speaker 1: because I've been very public about my story and advocating 550 00:39:02,916 --> 00:39:07,356 Speaker 1: for change. So I had actually gotten email from a 551 00:39:07,436 --> 00:39:11,076 Speaker 1: fake email address, and you know, I clicked the link 552 00:39:11,156 --> 00:39:15,076 Speaker 1: I was actually at work. It was a video of 553 00:39:15,116 --> 00:39:19,996 Speaker 1: me having sexual intercourse. The title had my name, the 554 00:39:20,036 --> 00:39:23,756 Speaker 1: face of the woman in it was edited so that 555 00:39:23,836 --> 00:39:27,196 Speaker 1: it was my face, and you know, all the tags 556 00:39:27,236 --> 00:39:34,076 Speaker 1: were like Noel Martin Australia, Feminist, and it didn't look real, 557 00:39:34,836 --> 00:39:39,156 Speaker 1: but the context of everything, with the title my face, 558 00:39:39,396 --> 00:39:43,996 Speaker 1: with the tags all points to me being depicted in 559 00:39:44,036 --> 00:39:47,436 Speaker 1: this video. The fakes were of poor quality, but poor 560 00:39:47,516 --> 00:39:52,116 Speaker 1: and consumers aren't a discriminating lot, and many people reacted 561 00:39:52,116 --> 00:39:54,396 Speaker 1: to them as if they were real. The public reaction 562 00:39:54,556 --> 00:39:57,916 Speaker 1: was horrifying to me. I was a victim, blamed and 563 00:39:57,996 --> 00:40:01,836 Speaker 1: slut shamed, and it's definitely limited the course of where 564 00:40:01,876 --> 00:40:06,476 Speaker 1: I can go in terms of career and employment. Noel 565 00:40:06,596 --> 00:40:09,796 Speaker 1: finished a degree in law and began a peening to 566 00:40:09,876 --> 00:40:14,716 Speaker 1: criminalize this sort of content. My advocacy and my activism 567 00:40:14,876 --> 00:40:17,756 Speaker 1: started off because I had a lived experience of this, 568 00:40:17,916 --> 00:40:21,596 Speaker 1: and I experienced it at a time where it wasn't 569 00:40:21,636 --> 00:40:28,516 Speaker 1: criminalized in Australia, the distribution of altered intimate images or 570 00:40:28,676 --> 00:40:34,156 Speaker 1: altered intimate videos, and so I had to petition, meet 571 00:40:34,196 --> 00:40:38,236 Speaker 1: with my politicians in my area. I wrote a number 572 00:40:38,276 --> 00:40:41,036 Speaker 1: of articles, I spoke to the media, and I was 573 00:40:41,516 --> 00:40:45,156 Speaker 1: involved in the law reform in Australia in a number 574 00:40:45,196 --> 00:40:49,116 Speaker 1: of jurisdictions in Western Australia and New South Wales, and 575 00:40:49,516 --> 00:40:53,356 Speaker 1: I ended up being involved in two press conferences with 576 00:40:53,436 --> 00:40:57,596 Speaker 1: the Attorney generals of each state at the announcement of 577 00:40:57,636 --> 00:41:03,556 Speaker 1: the law that was criminalizing this abuse. Today, in part 578 00:41:03,636 --> 00:41:07,156 Speaker 1: because of Noel's activism, it is illegal in Australia to 579 00:41:07,196 --> 00:41:11,676 Speaker 1: distribute intimate images without and scent, including intimate images and 580 00:41:11,836 --> 00:41:15,836 Speaker 1: videos that have been altered. Although it doesn't encompass all 581 00:41:15,996 --> 00:41:24,676 Speaker 1: malicious synthetic media, Noel has made a solid start. Chapter six, 582 00:41:25,236 --> 00:41:30,836 Speaker 1: Scissors and Glue. The videos depicting Noel Martin were nowhere 583 00:41:30,956 --> 00:41:34,876 Speaker 1: near as sophisticated as those made by the Moon disaster team. 584 00:41:35,356 --> 00:41:39,436 Speaker 1: They were more cheap fakes than deep fakes, and yet 585 00:41:39,556 --> 00:41:41,996 Speaker 1: the point didn't have to be perfect to be devastating. 586 00:41:42,756 --> 00:41:45,756 Speaker 1: The same turns out to be true in politics. To 587 00:41:45,836 --> 00:41:50,516 Speaker 1: understand the power of fakes, you have to understand human psychology. 588 00:41:50,636 --> 00:41:53,316 Speaker 1: It turns out that people are pretty easy to fool. 589 00:41:53,916 --> 00:41:56,676 Speaker 1: John Kerry I was running for president of the US. 590 00:41:57,116 --> 00:42:01,076 Speaker 1: His stance on the Vietnam War was controversial. Jane Fond, 591 00:42:01,076 --> 00:42:03,636 Speaker 1: of course, was a very controversial figure back then because 592 00:42:03,636 --> 00:42:06,596 Speaker 1: of her anti war stand. What have we become as 593 00:42:06,596 --> 00:42:08,556 Speaker 1: a nation if we call the men heroes that were 594 00:42:08,636 --> 00:42:10,516 Speaker 1: used by the had have gone to try to exterminate 595 00:42:10,516 --> 00:42:12,476 Speaker 1: an entire people? What business have we to try to 596 00:42:12,476 --> 00:42:15,156 Speaker 1: exterminate a people? And somebody had created a photo of 597 00:42:15,156 --> 00:42:17,356 Speaker 1: the two of them sharing a stage and an anti 598 00:42:17,396 --> 00:42:21,196 Speaker 1: war rally with the hopes of damaging the Carry campaign. 599 00:42:21,316 --> 00:42:23,756 Speaker 1: The photo was fake. They had never shared a stage together. 600 00:42:24,196 --> 00:42:26,756 Speaker 1: They just took two images, probably put it into some 601 00:42:26,796 --> 00:42:30,196 Speaker 1: standard photo editing software like a photoshop, and just put 602 00:42:30,196 --> 00:42:32,956 Speaker 1: a headline around it, and out to the world it went. 603 00:42:33,316 --> 00:42:35,916 Speaker 1: And I will tell you I remember the most fascinating 604 00:42:36,076 --> 00:42:39,156 Speaker 1: interview I've heard in a long time was right after 605 00:42:39,476 --> 00:42:42,956 Speaker 1: the election, Carry, of course lost, and a voter was 606 00:42:43,036 --> 00:42:45,916 Speaker 1: being interviewed and asked how they voted, and he said 607 00:42:45,956 --> 00:42:48,196 Speaker 1: he couldn't vote for Carry, and the interview said, well 608 00:42:48,196 --> 00:42:50,956 Speaker 1: why not? And the gentleman said, I couldn't get that 609 00:42:50,996 --> 00:42:53,236 Speaker 1: photo of John Carry and Jane Fonda out of my head. 610 00:42:53,476 --> 00:42:56,196 Speaker 1: And the interview said, well, you know that photo is fake, 611 00:42:56,436 --> 00:42:59,316 Speaker 1: and the guy said, much to my surprise, yes, but 612 00:42:59,396 --> 00:43:02,036 Speaker 1: I couldn't get it out of my mind. And this 613 00:43:02,156 --> 00:43:05,196 Speaker 1: is shows you the power of visual imagery, Like even 614 00:43:05,236 --> 00:43:07,916 Speaker 1: after I tell you something is fake, it still had 615 00:43:07,916 --> 00:43:10,836 Speaker 1: an impact on somebody. And I thought, wow, we're in 616 00:43:10,876 --> 00:43:14,156 Speaker 1: a lot of trouble because it's very very hard to 617 00:43:14,196 --> 00:43:16,476 Speaker 1: put the cat back into the bag. Once that content 618 00:43:16,596 --> 00:43:20,636 Speaker 1: is out there, you can't undo it. So seeing is believing, 619 00:43:21,076 --> 00:43:23,916 Speaker 1: even above thinking, Yeah, that seems to be the rule. 620 00:43:24,276 --> 00:43:27,356 Speaker 1: There is very good evidence from the social science literature 621 00:43:27,436 --> 00:43:30,556 Speaker 1: that it's very very difficult to correct the record after 622 00:43:30,596 --> 00:43:34,436 Speaker 1: the mistakes are out there. Law professor Danielle Citram also 623 00:43:34,516 --> 00:43:38,276 Speaker 1: notes that humans tend to pass on information without thinking, 624 00:43:38,796 --> 00:43:43,956 Speaker 1: which triggers what she calls information cascades. Information cascades is 625 00:43:43,996 --> 00:43:47,356 Speaker 1: a phenomenon where we have so much information overload that 626 00:43:47,476 --> 00:43:50,236 Speaker 1: when someone sends us something, some information, and we trust 627 00:43:50,276 --> 00:43:52,676 Speaker 1: that person, we pass it on. We don't even check 628 00:43:52,836 --> 00:43:57,116 Speaker 1: its veracity, and so information can go viral fairly quickly 629 00:43:57,756 --> 00:44:02,476 Speaker 1: because we're not terribly reflective, because we act on impulse. 630 00:44:03,676 --> 00:44:07,276 Speaker 1: Danielle says that information cascades have been given new life 631 00:44:07,316 --> 00:44:10,836 Speaker 1: in the twenty first century through social media. Think about 632 00:44:10,836 --> 00:44:13,516 Speaker 1: the twentieth century phenomenon where we get most of our 633 00:44:13,556 --> 00:44:20,476 Speaker 1: information from trusted sources, trusted newspapers, trusted major couple of 634 00:44:20,636 --> 00:44:22,756 Speaker 1: TV channels. Growing up, we only had to you know, 635 00:44:22,796 --> 00:44:27,036 Speaker 1: we didn't have a million, and they were adhering to 636 00:44:27,116 --> 00:44:31,156 Speaker 1: journalistic ethics and commitments to truth and neutrality and notion 637 00:44:31,236 --> 00:44:34,756 Speaker 1: that you can't publish something without checking it. Now we 638 00:44:35,636 --> 00:44:38,236 Speaker 1: are publishing information that most people say. We're lying on 639 00:44:38,276 --> 00:44:42,236 Speaker 1: our peers and our friends. Social media platforms are designed 640 00:44:42,636 --> 00:44:46,516 Speaker 1: to tailor our information diet to what we want and 641 00:44:46,636 --> 00:44:49,356 Speaker 1: to our pre existing views, so we're locked in a 642 00:44:49,436 --> 00:44:53,076 Speaker 1: digital echo chamber. We think everybody agrees with us. We 643 00:44:53,236 --> 00:44:57,316 Speaker 1: pass on that information we haven't checked the veracity, it 644 00:44:57,356 --> 00:45:00,476 Speaker 1: goes wild. And we're especially likely to pass it on 645 00:45:00,836 --> 00:45:04,196 Speaker 1: if it's negative and novel. Why's that, It's just like 646 00:45:04,436 --> 00:45:07,956 Speaker 1: it's one of our weaknesses. We know how gossip goes 647 00:45:07,956 --> 00:45:12,116 Speaker 1: like wildfire online. So like Hillary Clinton as running a 648 00:45:12,916 --> 00:45:16,716 Speaker 1: sex ring. That's crazy. Oh my god, Eric, did you 649 00:45:16,756 --> 00:45:19,316 Speaker 1: hear about that? I'll post it on Facebook. Eric, you 650 00:45:19,396 --> 00:45:22,876 Speaker 1: pass it on. We just can't help ourselves. And it 651 00:45:22,956 --> 00:45:25,836 Speaker 1: is much in the way that we love suits and 652 00:45:25,956 --> 00:45:30,036 Speaker 1: fats and pizza. You know, we indulge. We don't think 653 00:45:30,916 --> 00:45:34,636 Speaker 1: not some sense. This phenomenon is an old phenomenon, right, 654 00:45:34,676 --> 00:45:39,316 Speaker 1: there's the famous observation by Mark Twain about how a 655 00:45:39,396 --> 00:45:42,196 Speaker 1: lie gets halfway around the world before the truth gets 656 00:45:42,236 --> 00:45:43,756 Speaker 1: its pants hang. Yeah, the truth is still in the 657 00:45:43,756 --> 00:45:47,796 Speaker 1: bedroom getting dressed, and we often will see the lie, 658 00:45:47,996 --> 00:45:53,476 Speaker 1: but the rebuttal is not seen. It's often lost in 659 00:45:53,556 --> 00:45:57,276 Speaker 1: the noise of the defamatory statements. That is not new. 660 00:45:57,756 --> 00:46:00,996 Speaker 1: But what is new is a number of things about 661 00:46:00,996 --> 00:46:10,356 Speaker 1: our information ecosystem are our force multipliers. Chapter seven, Truth decay. 662 00:46:15,116 --> 00:46:18,676 Speaker 1: Many experts are worried that the rapid advances in making fakes, 663 00:46:19,116 --> 00:46:23,716 Speaker 1: combined with a catalyst of information cascades, will undermine democracy. 664 00:46:24,356 --> 00:46:29,276 Speaker 1: The biggest concerns have focused on elections. Globally, we are 665 00:46:29,316 --> 00:46:35,676 Speaker 1: looking at highly polarized situations where this kind of manipulated 666 00:46:35,756 --> 00:46:37,796 Speaker 1: media can be used as a weapon. One of the 667 00:46:37,836 --> 00:46:41,756 Speaker 1: main reasons Francesca and Halsey made their Nixon deep fake 668 00:46:42,356 --> 00:46:46,076 Speaker 1: was to spread awareness about the risks of misinformation campaigns 669 00:46:46,556 --> 00:46:50,756 Speaker 1: before the twenty twenty US presidential election. Similarly, a group 670 00:46:50,796 --> 00:46:54,116 Speaker 1: showcased the power of deep fakes by making videos in 671 00:46:54,116 --> 00:46:57,516 Speaker 1: the run up to the UK parliamentary election showing the 672 00:46:57,556 --> 00:47:02,396 Speaker 1: two bitter rivals Boris Johnson and Jeremy Corman, each endorsing 673 00:47:02,396 --> 00:47:05,596 Speaker 1: the other. I wish to rise above this divide and 674 00:47:05,756 --> 00:47:09,436 Speaker 1: indorse my worthy opponent, the right Honorable Jeremy Corbyn, to 675 00:47:09,716 --> 00:47:13,796 Speaker 1: be Prime Minister of our United Kingdom back Boris Johnson 676 00:47:13,876 --> 00:47:16,916 Speaker 1: to continue as our Prime Minister. But you know what, 677 00:47:17,676 --> 00:47:20,156 Speaker 1: don't listen to me. I think I may be one 678 00:47:20,196 --> 00:47:23,636 Speaker 1: of the thousands of deep fakes on the Internet, using 679 00:47:23,756 --> 00:47:28,156 Speaker 1: powerful technologies to tell stories that aren't so. This just 680 00:47:28,236 --> 00:47:34,076 Speaker 1: kind of indicates how candidates and political figures can be misrepresented, 681 00:47:34,636 --> 00:47:38,956 Speaker 1: and you just need to feed them into people's social 682 00:47:38,956 --> 00:47:41,876 Speaker 1: media feeds for them to be seeing this. At times 683 00:47:41,876 --> 00:47:44,956 Speaker 1: when the stakes are pretty high. So far, we haven't 684 00:47:45,036 --> 00:47:49,716 Speaker 1: yet seen sophisticated deep fakes in US or UK politics. 685 00:47:49,756 --> 00:47:53,156 Speaker 1: That might be because fakes will be most effective if 686 00:47:53,156 --> 00:47:56,916 Speaker 1: they're time for maximum chaos, say close to election day, 687 00:47:56,956 --> 00:48:00,356 Speaker 1: when newsrooms won't have the time to investigate and debunk them. 688 00:48:01,076 --> 00:48:04,756 Speaker 1: But another reason might be that well cheap fakes made 689 00:48:04,756 --> 00:48:09,676 Speaker 1: with basic video editing software or actually pretty effective. Remember 690 00:48:09,716 --> 00:48:13,116 Speaker 1: the video that surfaced of how Speaker Nancy Pelosi, in 691 00:48:13,116 --> 00:48:17,356 Speaker 1: which she appeared intoxicated and confused. We want to give 692 00:48:17,436 --> 00:48:25,396 Speaker 1: this president the opportunity do something historic for our country. 693 00:48:25,956 --> 00:48:29,276 Speaker 1: Both President Trump and Rudy Giuliani shared the video as 694 00:48:29,356 --> 00:48:32,636 Speaker 1: fact on Twitter. The video is just a cheap fake, 695 00:48:32,996 --> 00:48:37,036 Speaker 1: just slowed down Pelosi's speech to make her seem incompetent. 696 00:48:37,436 --> 00:48:42,196 Speaker 1: But maybe elections won't be the biggest targets. Some people 697 00:48:42,276 --> 00:48:47,356 Speaker 1: worry the deep fakes could be weaponized to foment international conflict. 698 00:48:47,876 --> 00:48:50,796 Speaker 1: Berkeley professor Honey f Reed has been working with US 699 00:48:50,836 --> 00:48:55,356 Speaker 1: government's Media Forensics program to address this issue. DARPA, on 700 00:48:55,436 --> 00:48:58,236 Speaker 1: the Defense Department's Research arm, has been pouring a lot 701 00:48:58,276 --> 00:49:00,516 Speaker 1: of money over the last five years into this program. 702 00:49:01,116 --> 00:49:05,636 Speaker 1: They are very concerned about how this technology can be 703 00:49:05,876 --> 00:49:08,876 Speaker 1: a threat to national security and also how when we 704 00:49:08,876 --> 00:49:11,156 Speaker 1: get images and videos from around the world in areas 705 00:49:11,156 --> 00:49:12,916 Speaker 1: of conflict, do we know if they're real or not? 706 00:49:13,436 --> 00:49:15,556 Speaker 1: Is this really an image of a US soldier who 707 00:49:15,596 --> 00:49:19,196 Speaker 1: has been taken hostage? How do we know? So? What 708 00:49:19,276 --> 00:49:22,196 Speaker 1: do you see as some of the worst case scenarios. 709 00:49:22,556 --> 00:49:24,396 Speaker 1: Here's the things that keep me up at night right 710 00:49:25,116 --> 00:49:27,996 Speaker 1: a video of Donald Trump saying I've launched nuclear weapons 711 00:49:27,996 --> 00:49:31,396 Speaker 1: against Iran, and before anybody gets around to figuring out 712 00:49:31,396 --> 00:49:33,076 Speaker 1: whether this is real or not, where we have global 713 00:49:33,196 --> 00:49:35,956 Speaker 1: nuclear mountdown. And here's the thing. I don't think that 714 00:49:35,956 --> 00:49:39,916 Speaker 1: that's likely, but I also don't think that the probability 715 00:49:39,956 --> 00:49:43,836 Speaker 1: of that is zero. And that should worry us because 716 00:49:44,196 --> 00:49:49,196 Speaker 1: while it's not likely, the consequences are spectacularly bad. Lawyer 717 00:49:49,276 --> 00:49:54,516 Speaker 1: Danielle Citrom worries about an even more plausible scenario. Imagine 718 00:49:54,556 --> 00:49:58,356 Speaker 1: a deep fake of a well known American general burning 719 00:49:58,356 --> 00:50:02,596 Speaker 1: a Koran, and it is timed at a very tense 720 00:50:02,716 --> 00:50:09,556 Speaker 1: moment in a particular most country, whether it's agha unerstand. 721 00:50:09,796 --> 00:50:13,316 Speaker 1: It could then lead to physical violence. And you think 722 00:50:13,356 --> 00:50:16,836 Speaker 1: this could be made. No general, no Quran actually used 723 00:50:16,876 --> 00:50:20,476 Speaker 1: in the video just programmed. You can use the technology 724 00:50:21,076 --> 00:50:24,316 Speaker 1: to mine existing photographs. Kind of easy, especially with someone 725 00:50:24,476 --> 00:50:27,996 Speaker 1: could take Jim Mattis when he was our defense secretary. 726 00:50:27,996 --> 00:50:31,076 Speaker 1: Of Jim Mattis, you know, actually taking a Koran and 727 00:50:31,156 --> 00:50:33,836 Speaker 1: ripping it in half and say all Muslims should die. 728 00:50:34,396 --> 00:50:40,116 Speaker 1: Imagine the chaos in diplomacy, the chaos of our soldiers 729 00:50:40,116 --> 00:50:46,156 Speaker 1: abroad in Muslim countries. It would be inciting violence without question. Well, 730 00:50:46,196 --> 00:50:50,116 Speaker 1: we haven't yet seen spectacular fake videos used to disrupt 731 00:50:50,196 --> 00:50:55,636 Speaker 1: elections or create international chaos. We have seen increasingly sophisticated 732 00:50:55,676 --> 00:51:00,116 Speaker 1: attacks on public policymaking. So we've got an example in 733 00:51:00,276 --> 00:51:04,916 Speaker 1: twenty seventeen where the FCC solicited public comment on the 734 00:51:04,956 --> 00:51:09,436 Speaker 1: proposal to repeal net neutrality. Net neutrality is the instable 735 00:51:09,476 --> 00:51:13,516 Speaker 1: that Internet service providers should be a neutral public utility. 736 00:51:14,116 --> 00:51:18,996 Speaker 1: They shouldn't discriminate between websites, say slowing down Netflix streaming 737 00:51:19,276 --> 00:51:22,716 Speaker 1: to encourage you to purchase a different online video service. 738 00:51:23,556 --> 00:51:27,476 Speaker 1: As President Barack Obama described in twenty fourteen, there are 739 00:51:27,476 --> 00:51:30,916 Speaker 1: no gatekeepers deciding which sites you get to access. There 740 00:51:30,916 --> 00:51:34,276 Speaker 1: are no toll roads on the information super Highway. Federal 741 00:51:34,316 --> 00:51:39,676 Speaker 1: communications policy had long supported net neutrality, but in twenty seventeen, 742 00:51:40,076 --> 00:51:44,396 Speaker 1: the Trump administration favored repealing the policy. There were twenty 743 00:51:44,396 --> 00:51:48,996 Speaker 1: two million comments that the e SEC received, but ninety 744 00:51:49,036 --> 00:51:53,956 Speaker 1: six percent of those were actually fake. The interesting thing 745 00:51:54,196 --> 00:51:58,676 Speaker 1: is the real comments were opposed to repeal, whereas the 746 00:51:58,756 --> 00:52:02,476 Speaker 1: fake comments were in favor. A Wall Street Journal investigation 747 00:52:02,716 --> 00:52:07,076 Speaker 1: exposed that the fake public comments were generated by bots. 748 00:52:07,716 --> 00:52:11,196 Speaker 1: It found similar problems with public comments about payday lending. 749 00:52:12,076 --> 00:52:16,236 Speaker 1: The bots varied their comments in a combinatorial fashion so 750 00:52:16,276 --> 00:52:20,036 Speaker 1: that the content wasn't identical. With a little sleuthing, though, 751 00:52:20,076 --> 00:52:23,516 Speaker 1: you could see that they were generated by computers. But 752 00:52:23,556 --> 00:52:28,076 Speaker 1: with a technology increasingly able to generate completely original writing, 753 00:52:28,636 --> 00:52:32,036 Speaker 1: like open ayes program that wrote the story about unicorns 754 00:52:32,036 --> 00:52:35,316 Speaker 1: in the ANDES, it's going to become hard to spot 755 00:52:35,356 --> 00:52:38,836 Speaker 1: the fakes. So there was this Harvest student, Max Weiss, 756 00:52:38,876 --> 00:52:42,476 Speaker 1: who used GPT two to kind of demonstrate this, And 757 00:52:42,556 --> 00:52:44,516 Speaker 1: I went on his site yesterday and he's got this 758 00:52:44,556 --> 00:52:49,516 Speaker 1: little test where you need to decide whether a comment 759 00:52:50,156 --> 00:52:53,036 Speaker 1: is real or fake. So you go on and you 760 00:52:53,076 --> 00:52:55,356 Speaker 1: read it, and you decide whether it's been written by 761 00:52:55,356 --> 00:52:58,316 Speaker 1: a bot or by a human. So I did this, 762 00:52:58,476 --> 00:53:02,196 Speaker 1: and the ones that seemed to be really well written 763 00:53:02,236 --> 00:53:05,036 Speaker 1: and quite narrative and discussive. Generally, I was picking them 764 00:53:05,036 --> 00:53:07,596 Speaker 1: as human. I was wrong almost all the time. It 765 00:53:07,716 --> 00:53:12,116 Speaker 1: was amazing and a lot. In our democracy, public commons 766 00:53:12,276 --> 00:53:14,756 Speaker 1: have been an important way in which citizens can make 767 00:53:14,796 --> 00:53:19,076 Speaker 1: their voices heard. But now it's becoming easy to drown 768 00:53:19,116 --> 00:53:23,356 Speaker 1: out those voices with millions of fake opinions. Now the 769 00:53:23,436 --> 00:53:27,076 Speaker 1: downfall of truth likely won't come with a bang, but 770 00:53:27,196 --> 00:53:33,276 Speaker 1: a whimper, a slow, steady erosion that some call truth decay. 771 00:53:33,476 --> 00:53:35,556 Speaker 1: If you can't believe anything you read, or hear or 772 00:53:35,556 --> 00:53:37,996 Speaker 1: see anymore, I don't know how you have a democracy, 773 00:53:38,036 --> 00:53:40,796 Speaker 1: an I don't know, frankly, how we have civilized society 774 00:53:41,156 --> 00:53:43,476 Speaker 1: if everybody's going to live in an echo chamber, believing 775 00:53:43,476 --> 00:53:46,116 Speaker 1: their own version of events. How do we have a 776 00:53:46,116 --> 00:53:49,476 Speaker 1: dialogue if we can't agree on basic facts. In the end, 777 00:53:49,756 --> 00:53:53,476 Speaker 1: the most insidious impact of deep fakes may not be 778 00:53:53,516 --> 00:53:57,396 Speaker 1: the deep fake content itself, but the ability to claim 779 00:53:57,476 --> 00:54:01,596 Speaker 1: that real content is fake. It's something that Danielle Citron 780 00:54:01,836 --> 00:54:06,236 Speaker 1: refers to as the liars dividend. The liars dividend is 781 00:54:06,276 --> 00:54:08,876 Speaker 1: that the more you educate people about the phenomenon of 782 00:54:08,916 --> 00:54:13,436 Speaker 1: deep fix, the more the wrongdoer can disclaim reality. Think 783 00:54:13,476 --> 00:54:17,596 Speaker 1: about what President Trump did with the Access Hollywood tape. 784 00:54:17,916 --> 00:54:19,836 Speaker 1: You know, I'm automatically attracted to be in the work. 785 00:54:19,876 --> 00:54:21,916 Speaker 1: I just started kissing them. It's like a magnet. You 786 00:54:22,036 --> 00:54:24,956 Speaker 1: just I don't even know. And when you were a start, 787 00:54:25,036 --> 00:54:27,276 Speaker 1: they let you do it. You can do anything whatever 788 00:54:27,276 --> 00:54:32,516 Speaker 1: you want, grabbed by the I can do anything. Initially, 789 00:54:32,756 --> 00:54:36,796 Speaker 1: Trump apologized for the remarks. Anyone who knows me knows 790 00:54:36,876 --> 00:54:40,716 Speaker 1: these words don't reflect who I am. I said it. 791 00:54:41,196 --> 00:54:45,276 Speaker 1: I was wrong, and I apologize. But in twenty seventeen, 792 00:54:45,796 --> 00:54:49,356 Speaker 1: a year after his initial apology and with the idea 793 00:54:49,396 --> 00:54:53,636 Speaker 1: of deep fake content starting to gain attention, Trump changed 794 00:54:53,716 --> 00:54:56,836 Speaker 1: his tune. Upon reflection, he said, they're not real. That 795 00:54:56,956 --> 00:54:59,396 Speaker 1: wasn't me. I don't think that was my voice. That's 796 00:54:59,396 --> 00:55:03,836 Speaker 1: the liar's dividend. In practice, the Trump commented about Access 797 00:55:03,876 --> 00:55:09,036 Speaker 1: Hollywood was remarkable. Slightly more subtle than that, He said, 798 00:55:09,556 --> 00:55:12,356 Speaker 1: I'm not sure that was me. Right. Well, that's the 799 00:55:12,396 --> 00:55:27,396 Speaker 1: corrosive gas lighting. Chapter eight, A Life Stored in the Cloud. 800 00:55:29,356 --> 00:55:33,676 Speaker 1: Deep fakes have the potential to devastate individuals and harms society. 801 00:55:34,276 --> 00:55:38,116 Speaker 1: The question is can we stop them from spreading before 802 00:55:38,116 --> 00:55:41,396 Speaker 1: they get out of control. To do so, we need 803 00:55:41,476 --> 00:55:45,596 Speaker 1: reliable ways to spot deep fakes. So the good news 804 00:55:45,716 --> 00:55:48,596 Speaker 1: is there are still artifacts in the synthesized content, whether 805 00:55:48,596 --> 00:55:51,196 Speaker 1: those are images, audio, or a video, that we as 806 00:55:51,236 --> 00:55:54,196 Speaker 1: the experts, can tell apart. So when for example, the 807 00:55:54,196 --> 00:55:56,396 Speaker 1: New York Times wants to run a story with a video, 808 00:55:56,956 --> 00:55:59,676 Speaker 1: we can help them validate it. What are the real 809 00:55:59,756 --> 00:56:03,996 Speaker 1: sophisticated experts looking gat So the eyes are really wonderful 810 00:56:04,156 --> 00:56:07,876 Speaker 1: forensically because they reflect back to you what is in 811 00:56:07,916 --> 00:56:11,356 Speaker 1: the scene. I'm sitting now right now in a studio. 812 00:56:11,476 --> 00:56:13,836 Speaker 1: There's maybe about a dozen or so lights around me, 813 00:56:13,876 --> 00:56:16,236 Speaker 1: and you can see this very complex set of reflections 814 00:56:16,236 --> 00:56:20,276 Speaker 1: in my eyes. So we can analyze fairly complex lighting patterns, 815 00:56:20,276 --> 00:56:23,236 Speaker 1: for example, to determine if this is one person's head 816 00:56:23,276 --> 00:56:26,236 Speaker 1: spliced onto another person's body, or if the two people 817 00:56:26,236 --> 00:56:30,596 Speaker 1: standing next to each other were digitally inserted from another photograph. 818 00:56:30,676 --> 00:56:33,316 Speaker 1: I could spend another hour telling you about the many 819 00:56:33,316 --> 00:56:36,796 Speaker 1: different forensic techniques that we've developed. There's no silver bullet here. 820 00:56:37,196 --> 00:56:40,116 Speaker 1: Really is a sort of a time consuming and deliberate 821 00:56:40,156 --> 00:56:43,476 Speaker 1: and thoughtful and it requires many many tools, and it 822 00:56:43,516 --> 00:56:45,636 Speaker 1: requires people with a fair amount of skill to do this. 823 00:56:46,156 --> 00:56:49,436 Speaker 1: Honey Freed also has quite a few detection techniques that 824 00:56:49,516 --> 00:56:52,076 Speaker 1: he won't speak about publicly for fear of the deep 825 00:56:52,116 --> 00:56:55,316 Speaker 1: fake creators will learn how to beat his tests. I 826 00:56:55,316 --> 00:56:57,636 Speaker 1: don't create a GitHub repository and give my code to 827 00:56:57,676 --> 00:57:00,956 Speaker 1: all my adversaries. I don't have just one forensic techniques. 828 00:57:01,036 --> 00:57:03,796 Speaker 1: I have a couple dozen of them. So that means you, 829 00:57:04,036 --> 00:57:06,276 Speaker 1: as the person creating this now have to go back 830 00:57:06,316 --> 00:57:09,636 Speaker 1: and implement twenty different techniques. You have to do it 831 00:57:09,716 --> 00:57:12,516 Speaker 1: just perfectly, and that makes the landscape a little bit 832 00:57:12,556 --> 00:57:15,756 Speaker 1: more tricky for you to manage. As technology makes it 833 00:57:15,796 --> 00:57:19,236 Speaker 1: easier to create deep fakes, a big problem will be 834 00:57:19,276 --> 00:57:22,836 Speaker 1: the sheer amounts of content to review. So the average 835 00:57:22,876 --> 00:57:26,476 Speaker 1: person can download software repositories, and so it's getting to 836 00:57:26,476 --> 00:57:29,756 Speaker 1: the point now where the average person can just run 837 00:57:29,796 --> 00:57:32,196 Speaker 1: these as if they're running any standard piece of software. 838 00:57:32,236 --> 00:57:35,116 Speaker 1: There's also websites that have propped up where you can 839 00:57:35,116 --> 00:57:37,556 Speaker 1: pay them twenty bucks and you tell them, please put 840 00:57:37,596 --> 00:57:39,956 Speaker 1: this person's face into this person's video, and they will 841 00:57:39,956 --> 00:57:42,396 Speaker 1: do that for you. And so it doesn't take a 842 00:57:42,396 --> 00:57:44,916 Speaker 1: lot to get access to these tools. Now, I will 843 00:57:44,956 --> 00:57:47,236 Speaker 1: say that the output of those are not quite as 844 00:57:47,276 --> 00:57:49,756 Speaker 1: good as what we can create inside the lab. And 845 00:57:49,796 --> 00:57:51,556 Speaker 1: you just know what the trend is. You just know 846 00:57:51,596 --> 00:57:53,876 Speaker 1: it's going to get better and cheaper and faster and 847 00:57:53,956 --> 00:57:57,196 Speaker 1: easier to use. Detecting few fakes will be a never 848 00:57:57,396 --> 00:58:02,236 Speaker 1: ending cat and mouse game. Remember how generative adversarial networks 849 00:58:02,316 --> 00:58:06,716 Speaker 1: or gams are built by training a fake generator to 850 00:58:06,796 --> 00:58:12,316 Speaker 1: outsmart a detector. Well, as detectors get better, fake generators 851 00:58:12,396 --> 00:58:16,956 Speaker 1: will be trained to keep pays still. Detectives like Honey 852 00:58:17,036 --> 00:58:21,116 Speaker 1: and platforms like Facebook are working to develop automated ways 853 00:58:21,156 --> 00:58:25,876 Speaker 1: to spot deep fakes rapidly and reliably. That's important because 854 00:58:25,916 --> 00:58:30,036 Speaker 1: more than five hundred additional hours of video are being 855 00:58:30,116 --> 00:58:33,756 Speaker 1: uploaded to YouTube every minute. I don't mean to sound 856 00:58:33,796 --> 00:58:36,796 Speaker 1: defeatist about this, but I'm going to lose this war. 857 00:58:37,036 --> 00:58:39,916 Speaker 1: I know this because it's always going to be easier 858 00:58:39,916 --> 00:58:42,196 Speaker 1: to create content than it is to detect it. But 859 00:58:43,156 --> 00:58:45,276 Speaker 1: here's where I will win. I will take it out 860 00:58:45,316 --> 00:58:48,076 Speaker 1: of the hands of the average person. So think about, 861 00:58:48,156 --> 00:58:51,716 Speaker 1: for example, the creation of counterfeit currency. With the latest 862 00:58:51,836 --> 00:58:55,156 Speaker 1: innovations brought on by the Treasure Department, it is hard 863 00:58:55,156 --> 00:58:57,356 Speaker 1: for the average person to take their inkjet printer and 864 00:58:57,396 --> 00:59:00,756 Speaker 1: create compelling fake currency, and I think that's going to 865 00:59:00,836 --> 00:59:03,196 Speaker 1: be the same trend here is that if you're using 866 00:59:03,196 --> 00:59:05,396 Speaker 1: some off the shelf tool, if you're paying somebody on 867 00:59:05,396 --> 00:59:07,156 Speaker 1: the website, we're going to find you, and we're going 868 00:59:07,196 --> 00:59:09,556 Speaker 1: to find you quickly. But if you are a deadated 869 00:59:09,716 --> 00:59:12,476 Speaker 1: highly skilled of the time and the effort to create it, 870 00:59:12,756 --> 00:59:14,596 Speaker 1: we are going to have to work really hard to 871 00:59:14,596 --> 00:59:19,276 Speaker 1: detect those. Given the challenges of detecting fake content, some 872 00:59:19,356 --> 00:59:23,236 Speaker 1: people envision a different kind of techno fix. They proposed 873 00:59:23,316 --> 00:59:27,876 Speaker 1: developing airtight ways for content creators to mark their own 874 00:59:27,996 --> 00:59:33,436 Speaker 1: original video as real. That way, we could instantly recognize 875 00:59:33,476 --> 00:59:36,916 Speaker 1: an altered version if it wasn't identical. Now, there's ways 876 00:59:36,916 --> 00:59:39,476 Speaker 1: of authenticating at the point of recording, and these are 877 00:59:39,516 --> 00:59:43,036 Speaker 1: what it called control capture system. So here's the idea. 878 00:59:43,316 --> 00:59:46,396 Speaker 1: You use a special app on your mobile device that 879 00:59:46,476 --> 00:59:50,316 Speaker 1: at the point of capture a cryptographically signs the image 880 00:59:50,316 --> 00:59:52,996 Speaker 1: of the video or the audio. It puts that signature 881 00:59:53,036 --> 00:59:54,876 Speaker 1: onto the blockchain. And the only thing you have to 882 00:59:54,876 --> 00:59:57,156 Speaker 1: know about the blockchain is that it is an immutable 883 00:59:57,436 --> 01:00:01,236 Speaker 1: distributed ledger, which means that that signature is essentially impossible 884 01:00:01,276 --> 01:00:05,036 Speaker 1: to manipulate, and now all of that happened at the 885 01:00:05,036 --> 01:00:08,076 Speaker 1: point of recording. If I was running a campaign today 886 01:00:08,076 --> 01:00:12,036 Speaker 1: and I was worried about candidates likeness being misused, absolutely 887 01:00:12,116 --> 01:00:14,596 Speaker 1: every public event that they were at, I would record 888 01:00:14,596 --> 01:00:16,396 Speaker 1: with a control capture system, and I'd be able to 889 01:00:16,436 --> 01:00:19,676 Speaker 1: prove what they actually said or did at any point 890 01:00:19,716 --> 01:00:22,796 Speaker 1: in the future. So this approach would shift the burden 891 01:00:22,796 --> 01:00:27,196 Speaker 1: of authentication to the people creating the videos rather than 892 01:00:27,276 --> 01:00:31,956 Speaker 1: publishers or consumers. Law professor Danielle Citron has explored how 893 01:00:31,956 --> 01:00:36,116 Speaker 1: this solution could quickly become dystopium. We might see the 894 01:00:36,116 --> 01:00:39,276 Speaker 1: emergence of an essentially an audit trail of everything you 895 01:00:39,396 --> 01:00:41,956 Speaker 1: do and say all of the time. Danielle refers to 896 01:00:41,956 --> 01:00:46,716 Speaker 1: the business model as immutable lifelogs in the cloud. In 897 01:00:46,716 --> 01:00:48,836 Speaker 1: a way we sort of already seen it. There are 898 01:00:48,876 --> 01:00:51,716 Speaker 1: health plans that if you wear a fitbit all the 899 01:00:51,756 --> 01:00:54,556 Speaker 1: time and you let yourself be monitored, it lowers your 900 01:00:54,556 --> 01:00:57,956 Speaker 1: insurance your health insurance rates. But you can see how 901 01:00:58,116 --> 01:01:02,076 Speaker 1: if the incentives are there in the market to self surveil, 902 01:01:02,356 --> 01:01:06,956 Speaker 1: whether it's for health insurance, life insurance, car insurance, we're 903 01:01:06,996 --> 01:01:10,716 Speaker 1: going to see the unraveling of Brian to say by ourselves. 904 01:01:11,076 --> 01:01:15,916 Speaker 1: You know, corporations may very well, because the CEO is 905 01:01:15,956 --> 01:01:20,196 Speaker 1: so valuable, they may say, you've got to have a log, 906 01:01:20,396 --> 01:01:22,556 Speaker 1: an immutable audit trail of everything you do and say 907 01:01:22,596 --> 01:01:24,796 Speaker 1: so when that deep fake comes up the night before 908 01:01:24,836 --> 01:01:29,036 Speaker 1: the IPO, you can say, look, the CEO wasn't taking 909 01:01:29,036 --> 01:01:32,316 Speaker 1: the bribe, wasn't having sex with a prostitute. And so 910 01:01:32,356 --> 01:01:36,076 Speaker 1: we have proof, we have an auto trail, we have 911 01:01:36,076 --> 01:01:39,276 Speaker 1: a log. So when we were imagining, we were imagining 912 01:01:39,276 --> 01:01:43,676 Speaker 1: a business model that hasn't quite come up, but we 913 01:01:43,716 --> 01:01:48,236 Speaker 1: have gotten a number of requests from insurance companies as 914 01:01:48,276 --> 01:01:51,796 Speaker 1: well as companies to say we're interested in this idea. 915 01:01:51,916 --> 01:01:53,556 Speaker 1: So how much has to be in that log? Does 916 01:01:53,556 --> 01:01:55,636 Speaker 1: this have to be a whole video of your life? 917 01:01:55,716 --> 01:01:58,636 Speaker 1: That is a great question, one that terrifies us. So 918 01:01:58,676 --> 01:02:04,156 Speaker 1: it may be that you're logging locate geolocation, you're logging videos. 919 01:02:04,236 --> 01:02:07,116 Speaker 1: You see people talking and who they're interacting with, and 920 01:02:07,196 --> 01:02:09,516 Speaker 1: that might be good enough to prevent the miss Jeff 921 01:02:09,996 --> 01:02:14,556 Speaker 1: that would hijack the IPM. Your whole life online, yes, 922 01:02:14,756 --> 01:02:19,716 Speaker 1: stored securely, clock down, protected in the cloud. It is 923 01:02:20,116 --> 01:02:22,956 Speaker 1: at least for a privacy scholar. There are so many 924 01:02:22,956 --> 01:02:26,396 Speaker 1: reasons why we ought to have privacy that aren't about 925 01:02:26,516 --> 01:02:31,756 Speaker 1: hiding things. It's about creating spaces and managing boundaries around 926 01:02:31,836 --> 01:02:35,636 Speaker 1: ourselves and our intimates and our loved ones. So I 927 01:02:35,876 --> 01:02:39,876 Speaker 1: worry that if we entirely unravel privacy a in the 928 01:02:39,916 --> 01:02:45,436 Speaker 1: wrong hands is very dangerous right B It changes how 929 01:02:45,556 --> 01:02:54,316 Speaker 1: we think about ourselves and humanity. Chapter nine, Section two thirty. 930 01:02:55,676 --> 01:02:59,716 Speaker 1: So techno fixes are complicated. What about passing laws to 931 01:03:00,076 --> 01:03:03,076 Speaker 1: band deep fakes or at least deep fakes that don't 932 01:03:03,116 --> 01:03:07,196 Speaker 1: disclose their fake So the video and audio is speech? 933 01:03:07,596 --> 01:03:10,436 Speaker 1: In our First Amendment doctrine is very much a protective 934 01:03:10,956 --> 01:03:13,956 Speaker 1: of free speech, and the Supreme Court has explained that 935 01:03:14,756 --> 01:03:19,036 Speaker 1: lies just lies themselves without harm is protected speech. When 936 01:03:19,116 --> 01:03:21,596 Speaker 1: lies cause certain kinds of harm, we can regulate it 937 01:03:21,876 --> 01:03:29,716 Speaker 1: defamation of private people, threats, incitement, fraud, impersonation of government officials. 938 01:03:29,876 --> 01:03:35,716 Speaker 1: What about lies concerning public figures like politicians? California and Texas, 939 01:03:35,796 --> 01:03:40,076 Speaker 1: for instance, recently pass laws making it illegal to publish 940 01:03:40,116 --> 01:03:42,876 Speaker 1: deep fakes of a candidate in the weeks leading up 941 01:03:42,916 --> 01:03:46,156 Speaker 1: to an election. It's not clear yet whether the laws 942 01:03:46,196 --> 01:03:50,916 Speaker 1: will pass constitutional muster. As you're saying in an American content, 943 01:03:51,716 --> 01:03:55,716 Speaker 1: we are just not going to be able to outlaw fakes. Yeah, 944 01:03:55,716 --> 01:03:57,436 Speaker 1: we can't have a flat van, and I don't think 945 01:03:57,436 --> 01:04:01,396 Speaker 1: we should. It would fail on doctrinal grounds, but ultimately 946 01:04:01,636 --> 01:04:08,796 Speaker 1: it would prevent the positive uses. Interestingly, in January twenty twenty, China, 947 01:04:09,196 --> 01:04:14,036 Speaker 1: which has no First Amendment protecting free speech, promulgated regulations 948 01:04:14,436 --> 01:04:18,516 Speaker 1: banning deep fakes. The use of AI or virtuality now 949 01:04:18,596 --> 01:04:21,716 Speaker 1: needs to be clearly marked in a prominent manner, and 950 01:04:21,796 --> 01:04:25,156 Speaker 1: the failure to do so is considered a criminal offense. 951 01:04:25,956 --> 01:04:29,036 Speaker 1: To explore other options for the US, I went to 952 01:04:29,076 --> 01:04:32,676 Speaker 1: speak with a public policy expert. My name is Joan Donovan, 953 01:04:32,836 --> 01:04:36,396 Speaker 1: and I work at Harvard Kennedy Shorenstein Center, where I 954 01:04:36,516 --> 01:04:39,756 Speaker 1: lead a team of researchers looking at medium manipulation and 955 01:04:39,836 --> 01:04:43,596 Speaker 1: disinformation campaigns. Joan is head of the Technology and Social 956 01:04:43,676 --> 01:04:48,236 Speaker 1: Change Research Project, and her staff studies how social media 957 01:04:48,596 --> 01:04:52,676 Speaker 1: gives rise to hoaxes and scams. Her team is particularly 958 01:04:52,676 --> 01:04:58,516 Speaker 1: interested and precisely how misinformation spreads across the Internet. Ultimately, 959 01:04:58,636 --> 01:05:01,756 Speaker 1: underneath all of this is the distribution mechanism, which is 960 01:05:02,036 --> 01:05:08,356 Speaker 1: social media and platforms and platforms have to rethink the 961 01:05:08,756 --> 01:05:12,596 Speaker 1: open of their design because that has now become a 962 01:05:12,676 --> 01:05:17,516 Speaker 1: territory for information warfare. In early twenty twenty, Facebook announced 963 01:05:17,516 --> 01:05:23,396 Speaker 1: a major policy change about synthesized content. Facebook preissued policies 964 01:05:23,516 --> 01:05:26,556 Speaker 1: now on deep fakes, saying that if it is an 965 01:05:26,596 --> 01:05:31,676 Speaker 1: AI generated video and it's misleading in some other contextual way, 966 01:05:32,476 --> 01:05:37,996 Speaker 1: then they will remove it. Interestingly, Facebook banned the Moon 967 01:05:38,116 --> 01:05:41,436 Speaker 1: Disaster Team's Nixon video even though it was made for 968 01:05:41,596 --> 01:05:45,836 Speaker 1: educational purposes, but didn't remove the slowed down version of 969 01:05:45,956 --> 01:05:50,756 Speaker 1: Nancy Pelosi, which was made to mislead the public. Why 970 01:05:50,876 --> 01:05:55,676 Speaker 1: because the Pelosi video wasn't created with artificial intelligence. For now, 971 01:05:56,236 --> 01:05:59,996 Speaker 1: Facebook is choosing to target deep fakes, but not cheap fakes. 972 01:06:00,476 --> 01:06:03,236 Speaker 1: One way to push platforms to take a stronger stance 973 01:06:03,356 --> 01:06:06,196 Speaker 1: might be to remove some of the legal protections that 974 01:06:06,316 --> 01:06:11,196 Speaker 1: they currently enjoy. Undersection two thirty of the Communication's Decency 975 01:06:11,276 --> 01:06:16,236 Speaker 1: Act past in nineteen ninety six, platforms aren't legally liable 976 01:06:16,476 --> 01:06:20,716 Speaker 1: for content posted by its users. The fact that platforms 977 01:06:20,756 --> 01:06:24,636 Speaker 1: have no responsibility for the content they host has an upside. 978 01:06:25,076 --> 01:06:28,196 Speaker 1: It's led to the massive diversity of online content we 979 01:06:28,316 --> 01:06:32,916 Speaker 1: enjoyed today. But it also allows a dangerous escalation of 980 01:06:32,956 --> 01:06:37,356 Speaker 1: fake news. Is it time to change section to thirty 981 01:06:37,516 --> 01:06:42,236 Speaker 1: to create incentives for platforms to police false content? I 982 01:06:42,316 --> 01:06:45,556 Speaker 1: asked the former head of a major platform, LinkedIn co 983 01:06:45,716 --> 01:06:49,476 Speaker 1: founder Reid Hoffman. For example, let's take my view of 984 01:06:49,516 --> 01:06:52,636 Speaker 1: what the response to the christ Church shooting should be 985 01:06:52,876 --> 01:06:55,996 Speaker 1: is to say, well, we want you to solve not 986 01:06:56,156 --> 01:07:02,036 Speaker 1: having terrorism, murderer or murderers displayed to people. So we're 987 01:07:02,036 --> 01:07:04,676 Speaker 1: simply going to do a fine of ten thousand dollars 988 01:07:04,676 --> 01:07:08,836 Speaker 1: per view. Two shootings occurred at mosques in Christchurch, New 989 01:07:08,916 --> 01:07:13,076 Speaker 1: Zealand in March twenty nineteen. Graphic videos of the event 990 01:07:13,516 --> 01:07:17,276 Speaker 1: were soon posted online. Five people saw it, that's fifty 991 01:07:17,276 --> 01:07:20,116 Speaker 1: thousand dollars. But if he becomes a meme and a 992 01:07:20,196 --> 01:07:24,756 Speaker 1: million people see it, that's ten billion dollars. Yes, right, 993 01:07:24,956 --> 01:07:27,516 Speaker 1: So what's really trying to do is get you to say, 994 01:07:28,076 --> 01:07:30,956 Speaker 1: let's make sure that the meme never happens. Okay, So 995 01:07:30,996 --> 01:07:35,796 Speaker 1: that's a governance mechanism there is you find the channel 996 01:07:35,836 --> 01:07:39,196 Speaker 1: the platform based on number of views would be a 997 01:07:39,476 --> 01:07:42,356 Speaker 1: very general way to say, now you guys have to solve. 998 01:07:42,516 --> 01:07:46,396 Speaker 1: Now you solve, you figure it out. What about other solutions. 999 01:07:46,676 --> 01:07:50,076 Speaker 1: If we are to make regulation, it should be about 1000 01:07:50,156 --> 01:07:54,436 Speaker 1: the amount of staff in proportion to the amount of 1001 01:07:54,556 --> 01:07:58,116 Speaker 1: users so that they can get a handle on the content. 1002 01:07:58,596 --> 01:08:02,076 Speaker 1: But can they be fast enough. Maybe the viral spread 1003 01:08:02,116 --> 01:08:06,076 Speaker 1: should be slowed down enough to allow them to moderate. 1004 01:08:06,156 --> 01:08:10,236 Speaker 1: Let's put it this way. The stock market has certain 1005 01:08:10,916 --> 01:08:14,036 Speaker 1: governor is built in when there's massive changes in a 1006 01:08:14,076 --> 01:08:17,796 Speaker 1: stock price. There are decelerators that kick in, breaks that 1007 01:08:17,996 --> 01:08:20,956 Speaker 1: kick in. Should the platforms have breaks that kick in 1008 01:08:21,356 --> 01:08:26,076 Speaker 1: before something can go fully viral? So in terms of deceleration, 1009 01:08:26,876 --> 01:08:29,516 Speaker 1: there are things that they do already that accelerate the 1010 01:08:29,556 --> 01:08:33,156 Speaker 1: process that they need to think differently about, especially when 1011 01:08:33,196 --> 01:08:37,596 Speaker 1: it comes to something turning into a trending topic. So 1012 01:08:37,636 --> 01:08:41,876 Speaker 1: there needs to be an intervening moment before things get 1013 01:08:41,916 --> 01:08:45,076 Speaker 1: to the homepage and get to trending, where there is 1014 01:08:45,076 --> 01:08:49,076 Speaker 1: a content review. So much to say here, but I 1015 01:08:49,156 --> 01:08:52,676 Speaker 1: want to think particularly about listeners who are in their 1016 01:08:52,716 --> 01:08:56,876 Speaker 1: twenties and thirties and are very tech savvy. They're going 1017 01:08:56,956 --> 01:09:00,076 Speaker 1: to be part of the solution here. What would you 1018 01:09:00,116 --> 01:09:04,476 Speaker 1: say to them about what they can do? I think 1019 01:09:05,156 --> 01:09:11,156 Speaker 1: it's important that younger people advocate for the Internet that 1020 01:09:11,196 --> 01:09:13,596 Speaker 1: they want we have to fight for it. We have 1021 01:09:13,716 --> 01:09:18,036 Speaker 1: to ask for different things, and that kind of agitation 1022 01:09:18,476 --> 01:09:22,716 Speaker 1: can come in the form of posting on the platform, 1023 01:09:22,756 --> 01:09:27,156 Speaker 1: writing letters, joining groups like Fight for the Future, and 1024 01:09:27,236 --> 01:09:32,836 Speaker 1: trying to work on getting platforms to do better and 1025 01:09:32,916 --> 01:09:35,796 Speaker 1: to advocate for the kind of content that you want 1026 01:09:35,836 --> 01:09:40,636 Speaker 1: to see more of. The important thing is that our 1027 01:09:40,676 --> 01:09:45,436 Speaker 1: society is shaped by these platforms, and so we're not 1028 01:09:45,516 --> 01:09:48,276 Speaker 1: going to do away with them, but we don't have 1029 01:09:48,396 --> 01:09:59,396 Speaker 1: to make do with them either. Conclusion, choose your planet. 1030 01:10:01,636 --> 01:10:04,596 Speaker 1: So there you have it, Stewards of the Brave New Planet. 1031 01:10:05,156 --> 01:10:10,916 Speaker 1: Synthetic media or deep fakes, have been manipulating content for 1032 01:10:11,036 --> 01:10:14,836 Speaker 1: more than a hundred years, but recent advances in AI 1033 01:10:14,956 --> 01:10:18,236 Speaker 1: have taken it to a whole new level of verisimilitude. 1034 01:10:18,836 --> 01:10:23,596 Speaker 1: The technology could transform movies and television, favored actors from 1035 01:10:23,676 --> 01:10:27,316 Speaker 1: years past starring in new narratives, along with actors who 1036 01:10:27,356 --> 01:10:31,516 Speaker 1: never existed, patients regaining the ability to speak in their 1037 01:10:31,556 --> 01:10:37,316 Speaker 1: own voices, personalized stories created on demand for any child 1038 01:10:37,396 --> 01:10:41,396 Speaker 1: around the globe, matching their interests, written in their dialect, 1039 01:10:41,756 --> 01:10:46,716 Speaker 1: representing their communities. But there's also great potential for harm. 1040 01:10:47,356 --> 01:10:52,276 Speaker 1: The ability to cast anyone in a pornographic video, weaponized 1041 01:10:52,356 --> 01:10:56,836 Speaker 1: media dropping days before an election, or provoking international conflicts. 1042 01:10:57,716 --> 01:11:00,676 Speaker 1: Are we going to be able to tell fact from fiction? 1043 01:11:01,236 --> 01:11:06,956 Speaker 1: Will truth survive? And what does it mean for our democracy? Better? 1044 01:11:06,956 --> 01:11:10,036 Speaker 1: Fake detection may help, but it'll be hard for it 1045 01:11:10,076 --> 01:11:13,636 Speaker 1: to keep up, and logging our lives in blockchain to 1046 01:11:13,716 --> 01:11:19,996 Speaker 1: protect against misrepresentation doesn't sound like an attractive idea. Outright 1047 01:11:20,156 --> 01:11:22,916 Speaker 1: bands on deep fakes are being tried in some countries, 1048 01:11:23,236 --> 01:11:26,596 Speaker 1: but they're tricky in the US given our constitutional protections 1049 01:11:26,636 --> 01:11:30,516 Speaker 1: for free speech. Maybe the best solution is to put 1050 01:11:30,516 --> 01:11:35,116 Speaker 1: the liability on platforms like Facebook and YouTube. If we 1051 01:11:35,276 --> 01:11:39,156 Speaker 1: can joan Donovan's right to get the future you want, 1052 01:11:39,556 --> 01:11:42,076 Speaker 1: you're going to have to fight for it. You don't 1053 01:11:42,116 --> 01:11:44,476 Speaker 1: have to be an expert, and you don't have to 1054 01:11:44,476 --> 01:11:47,916 Speaker 1: do it alone. When enough people get engaged, we make 1055 01:11:47,956 --> 01:11:52,236 Speaker 1: wise choices. Deep fakes are a problem that everyone can 1056 01:11:52,276 --> 01:11:56,116 Speaker 1: engage with. Brainstorm with your friends about what should be done. 1057 01:11:56,556 --> 01:12:00,356 Speaker 1: Use social media. Tweet at your elected representatives to ask 1058 01:12:00,396 --> 01:12:03,796 Speaker 1: if they're working on laws, like in California and Texas, 1059 01:12:04,556 --> 01:12:08,276 Speaker 1: And if you work for a tech company, ask yourself 1060 01:12:08,516 --> 01:12:12,316 Speaker 1: and your colleagues if you're doing enough. You can find 1061 01:12:12,476 --> 01:12:16,116 Speaker 1: lots of resources and ideas at our website Brave New 1062 01:12:16,196 --> 01:12:20,876 Speaker 1: Planet dot org. It's time to choose our planet. The 1063 01:12:20,996 --> 01:12:34,916 Speaker 1: future is up to us. Brave New Planet is a 1064 01:12:34,956 --> 01:12:37,636 Speaker 1: co production of the Broad Institute of MT and Harvard 1065 01:12:37,716 --> 01:12:41,156 Speaker 1: Pushkin Industries in the Boston Globe, with support from the 1066 01:12:41,196 --> 01:12:44,876 Speaker 1: Alfred P. Sloan Foundation. Our show is produced by Rebecca 1067 01:12:44,916 --> 01:12:49,196 Speaker 1: Lee Douglas with Mary Doo theme song composed by Ned Porter, 1068 01:12:49,796 --> 01:12:53,916 Speaker 1: mastering and sound designed by James Garver, fact checking by 1069 01:12:53,996 --> 01:12:58,316 Speaker 1: Joseph Fridman, and a Stitt and Enchant. Special Thanks to 1070 01:12:58,436 --> 01:13:03,076 Speaker 1: Christine Heenan and Rachel Roberts at Clarendon Communications, to Lee McGuire, 1071 01:13:03,236 --> 01:13:06,596 Speaker 1: Kristen Zarelli and Justine Levin Allerhand at the Broad, to 1072 01:13:06,756 --> 01:13:10,956 Speaker 1: Milobelle and Heather Faine at Pushkin, and to Eli and 1073 01:13:11,116 --> 01:13:15,196 Speaker 1: Edy Brode, who made the Brode Institute possible. This is 1074 01:13:15,276 --> 01:13:17,716 Speaker 1: Brave New Planet. I am her Aclander