1
00:00:15,250 --> 00:00:26,810
Speaker 1: Pushkin, you're listening to Brave New Planet, a podcast about

2
00:00:26,850 --> 00:00:31,010
Speaker 1: amazing new technologies that could dramatically improve our world. Or

3
00:00:31,410 --> 00:00:33,970
Speaker 1: if we don't make wise choices, could leave us a

4
00:00:33,970 --> 00:00:44,690
Speaker 1: lot worse off. Utopia or dystopia. It's up to us

5
00:00:45,810 --> 00:00:52,810
Speaker 1: pash the eye. On July sixteenth, nineteen sixty nine, Apollo

6
00:00:52,850 --> 00:00:58,290
Speaker 1: eleven blasted off from the Kennedy Space Center near Cape Canaveral, Florida.

7
00:00:58,650 --> 00:01:03,290
Speaker 1: Twenty five million Americans watched on television as the spacecraft

8
00:01:03,410 --> 00:01:08,890
Speaker 1: ascended toward the heavens, carrying commander Neil Armstrong, Lunar Module

9
00:01:08,930 --> 00:01:14,410
Speaker 1: pilot Buzz Aldron, and Command Module pilot Michael Collins their

10
00:01:14,410 --> 00:01:18,330
Speaker 1: mission to be the first humans in history to set

11
00:01:18,410 --> 00:01:23,810
Speaker 1: foot on the Moon. Four days later, on Sunday, July twentieth,

12
00:01:24,210 --> 00:01:27,970
Speaker 1: the lunar module separated from the command ship and soon

13
00:01:28,210 --> 00:01:34,490
Speaker 1: fired its rockets to begin its lunar descent. Five minutes later,

14
00:01:35,330 --> 00:01:40,250
Speaker 1: disaster struck about a mile above the Moon's surface. Program

15
00:01:40,290 --> 00:01:43,890
Speaker 1: alarms twelve O one and twelve O two sounded loudly,

16
00:01:43,930 --> 00:01:49,810
Speaker 1: indicating that the mission computer was overloaded, and then, well,

17
00:01:51,130 --> 00:01:59,530
Speaker 1: every American knows what happened next. Lost date of five

18
00:02:05,770 --> 00:02:10,130
Speaker 1: good evening, my fellow Americans, President Richard Nick addressed a

19
00:02:10,250 --> 00:02:14,970
Speaker 1: grieving nation. Fades has ordained that the men who went

20
00:02:15,010 --> 00:02:19,010
Speaker 1: to the Moon to explore in peace will stay on

21
00:02:19,050 --> 00:02:25,930
Speaker 1: the Moon to rest in peace. These brave men, Neil

22
00:02:26,250 --> 00:02:32,130
Speaker 1: Armstrong and Edwin Auburn, know that there's no hope for

23
00:02:32,210 --> 00:02:37,290
Speaker 1: their recovery, but they also know that there is hope

24
00:02:37,410 --> 00:02:42,450
Speaker 1: for mankind in their sacrifice. He ended with the now

25
00:02:42,570 --> 00:02:46,370
Speaker 1: famous words for every human being who looks up at

26
00:02:46,370 --> 00:02:49,530
Speaker 1: the Moon and the nights to come, will know that

27
00:02:49,530 --> 00:02:54,610
Speaker 1: there is some corner another word that is forever mankind.

28
00:02:58,130 --> 00:03:02,130
Speaker 1: Wait a minute, that never happened. The Moon mission was

29
00:03:02,170 --> 00:03:07,170
Speaker 1: a historic success. The three astronauts returned safely to ticker

30
00:03:07,210 --> 00:03:11,130
Speaker 1: tape parades and a celebrity thirty eight day world tour.

31
00:03:11,930 --> 00:03:15,730
Speaker 1: Those alarms actually did sound, but they turned out to

32
00:03:15,730 --> 00:03:21,130
Speaker 1: be harmless. Nixon never delivered that speech. His speechwriter had

33
00:03:21,130 --> 00:03:24,450
Speaker 1: written it, but it sat in a folder labeled an

34
00:03:24,530 --> 00:03:30,650
Speaker 1: event of Moon disaster until now. The Nixon you just

35
00:03:30,810 --> 00:03:34,050
Speaker 1: heard is a deep fake, part of a seven minute

36
00:03:34,090 --> 00:03:39,650
Speaker 1: film created by artificial intelligence deep learning algorithms. The fake

37
00:03:39,810 --> 00:03:43,090
Speaker 1: was made by the Center for Advanced Virtuality at the

38
00:03:43,090 --> 00:03:46,930
Speaker 1: Massachusetts Institute of Technology as part of an art exhibit

39
00:03:47,210 --> 00:03:50,930
Speaker 1: to raise awareness about the power of synthesized media. Not

40
00:03:51,050 --> 00:03:54,570
Speaker 1: long ago, something like this would have taken a lot

41
00:03:54,730 --> 00:03:58,410
Speaker 1: of time and money, But now it's getting easy. You

42
00:03:58,450 --> 00:04:01,450
Speaker 1: can make new paintings in the style of French Impressionism,

43
00:04:01,770 --> 00:04:06,330
Speaker 1: revived dead movie stars, help patients with nor degenerative disease,

44
00:04:06,810 --> 00:04:09,570
Speaker 1: or soon maybe take a class on a tour of

45
00:04:09,610 --> 00:04:14,090
Speaker 1: ancient rome. But as the technology quickly becomes democratized, we're

46
00:04:14,130 --> 00:04:17,690
Speaker 1: getting to the point where almost anyone can create a

47
00:04:17,730 --> 00:04:21,210
Speaker 1: fake video of a friend, an ex lover, a stranger,

48
00:04:21,650 --> 00:04:26,650
Speaker 1: or a public figure that's embarrassing, pornographic, or perhaps capable

49
00:04:26,690 --> 00:04:31,450
Speaker 1: of causing international chaos. Some argue that in a culture

50
00:04:31,490 --> 00:04:36,290
Speaker 1: where fake news spreads like wildfire and political leaders deny

51
00:04:36,410 --> 00:04:40,530
Speaker 1: the veracity of hard facts, deep fake media may do

52
00:04:40,570 --> 00:04:47,090
Speaker 1: a lot more harm than good. Today's big question will

53
00:04:47,170 --> 00:04:52,570
Speaker 1: synthesized media unleash a new wave of creativity or will

54
00:04:52,570 --> 00:04:57,170
Speaker 1: it erode the already tenuous role of truth in our democracy?

55
00:04:57,290 --> 00:05:00,850
Speaker 1: And is there anything we can do to keep it

56
00:05:00,890 --> 00:05:12,090
Speaker 1: in check. My name is Eric Lander. I'm a scientist

57
00:05:12,090 --> 00:05:15,010
Speaker 1: who works on ways to improve human health. I helped

58
00:05:15,050 --> 00:05:17,770
Speaker 1: lead the Human Genome Project, and today I lead the

59
00:05:17,810 --> 00:05:21,930
Speaker 1: Broad Institute of MIT and Harvard. In the twenty first century,

60
00:05:22,290 --> 00:05:26,770
Speaker 1: powerful technologies have been appearing at a breathtaking pace related

61
00:05:26,810 --> 00:05:31,290
Speaker 1: to the Internet, artificial intelligence, genetic engineering, and more. They

62
00:05:31,330 --> 00:05:35,650
Speaker 1: have amazing potential upsides, but we can't ignore the risks

63
00:05:35,730 --> 00:05:38,770
Speaker 1: that come with them. The decisions aren't just up to

64
00:05:38,890 --> 00:05:42,730
Speaker 1: scientists or politicians. Whether we like it or not, we

65
00:05:43,450 --> 00:05:46,530
Speaker 1: all of us are the stewards of a brave New Planet.

66
00:05:47,090 --> 00:05:50,850
Speaker 1: This generation's choices will shape the future as never before.

67
00:05:53,410 --> 00:05:57,490
Speaker 1: Coming up on today's episode of Brave New Planet, I

68
00:05:57,530 --> 00:06:01,850
Speaker 1: speak with some of the leaders behind advances in synthesized media.

69
00:06:02,050 --> 00:06:05,050
Speaker 1: You could, certainly, by the way, generate stories that could

70
00:06:05,570 --> 00:06:09,370
Speaker 1: be fresh and interesting and new and personal for every child.

71
00:06:09,610 --> 00:06:13,410
Speaker 1: We got emails from people who were quadruplegic and they

72
00:06:13,450 --> 00:06:16,090
Speaker 1: asked us if we could make them dance. We hear

73
00:06:16,170 --> 00:06:19,810
Speaker 1: from experts about some of the frightening ways that bad

74
00:06:19,850 --> 00:06:23,850
Speaker 1: actors can use deep fakes. Creditors would chime in and say,

75
00:06:24,090 --> 00:06:26,970
Speaker 1: you can absolutely make a deep fake sex video of

76
00:06:26,970 --> 00:06:29,770
Speaker 1: your ex with thirty pictures. I've done it with twenty.

77
00:06:30,010 --> 00:06:31,850
Speaker 1: Here's the things that keep me up at night right

78
00:06:32,530 --> 00:06:35,410
Speaker 1: a video of Donald Trump saying I've launched nuclear weapons

79
00:06:35,410 --> 00:06:38,770
Speaker 1: against Iran, and before anybody gets around to firing out

80
00:06:38,810 --> 00:06:40,890
Speaker 1: whether this is real or not, we have global nuclear

81
00:06:40,970 --> 00:06:45,410
Speaker 1: outdown and we explore how we might prevent the worst abuses.

82
00:06:46,250 --> 00:06:52,050
Speaker 1: It's important that younger people advocate for the Internet that

83
00:06:52,090 --> 00:06:54,690
Speaker 1: they want. We have to fight for it. We have

84
00:06:54,770 --> 00:07:04,650
Speaker 1: to ask for different things. Stay with us, Chapter one,

85
00:07:05,210 --> 00:07:10,050
Speaker 1: Abraham Lincoln's Head. To begin to understand and the significance

86
00:07:10,050 --> 00:07:13,530
Speaker 1: of deep fake technology, I went to San Francisco to

87
00:07:13,610 --> 00:07:17,170
Speaker 1: speak with a world expert on synthetic media. My name

88
00:07:17,330 --> 00:07:22,170
Speaker 1: is Alexei or sometimes called Alyosha Afros, and I'm a

89
00:07:22,170 --> 00:07:26,650
Speaker 1: professor at UC Berkeley and Computer Science and Lexical Engineering Department.

90
00:07:27,170 --> 00:07:33,050
Speaker 1: My research is on computer vision, computer graphics, machine learning,

91
00:07:33,770 --> 00:07:39,570
Speaker 1: various aspects of artificial intelligence. Where'd you grow up. I

92
00:07:39,650 --> 00:07:43,530
Speaker 1: grew up in Saint Petersburg in Russia. I was one

93
00:07:43,570 --> 00:07:47,970
Speaker 1: of those geeky kids playing around with computers or dreaming

94
00:07:47,970 --> 00:07:55,170
Speaker 1: about computers. My first computer was actually the first Soviet

95
00:07:55,650 --> 00:07:59,610
Speaker 1: personal computer. So you actually are involved in making sort

96
00:07:59,650 --> 00:08:04,890
Speaker 1: of synthetic content, synthetic media, that's right. Alexei has invented

97
00:08:04,970 --> 00:08:08,970
Speaker 1: powerful artificial intelligence tools, but his lab also has one

98
00:08:09,250 --> 00:08:13,130
Speaker 1: full ability to use computers to enhance the human experience.

99
00:08:13,850 --> 00:08:17,250
Speaker 1: I was struck by a remarkable video on YouTube created

100
00:08:17,290 --> 00:08:21,090
Speaker 1: by his team at Berkeley. So this was a project

101
00:08:21,130 --> 00:08:26,530
Speaker 1: that actually was done by my students who didn't even

102
00:08:26,570 --> 00:08:30,290
Speaker 1: think of this as anything but a silly little toy

103
00:08:30,370 --> 00:08:34,370
Speaker 1: project of trying to see if we could get a

104
00:08:34,490 --> 00:08:39,290
Speaker 1: geeky computer science student to move like a ballerina. In

105
00:08:39,330 --> 00:08:43,050
Speaker 1: the video, one of the students, Carolyn cham dances with

106
00:08:43,130 --> 00:08:46,730
Speaker 1: a skill and grace of a professional despite never having

107
00:08:46,730 --> 00:08:50,890
Speaker 1: studied ballet. The idea is, you take a source actor

108
00:08:50,970 --> 00:08:55,090
Speaker 1: like a ballerina. There is a way to detect the

109
00:08:55,250 --> 00:08:59,770
Speaker 1: limbs of the dancer, have a kind of a skeleton extracted,

110
00:09:00,210 --> 00:09:04,090
Speaker 1: and also have my student just move around and do

111
00:09:04,210 --> 00:09:08,050
Speaker 1: some geeky moves. And now we're basically just going to

112
00:09:08,090 --> 00:09:14,090
Speaker 1: try to sympathize the appearance of my student driven by

113
00:09:14,130 --> 00:09:16,930
Speaker 1: the skeleton of the ballerina. Put it all together, and

114
00:09:16,970 --> 00:09:21,810
Speaker 1: then we have our grad student dancing pirouets like a ballerina.

115
00:09:23,290 --> 00:09:27,850
Speaker 1: Through artificial intelligence, Carolyn's body is puppeteered by the dancer.

116
00:09:28,130 --> 00:09:30,890
Speaker 1: We weren't even going to publish it, but we just

117
00:09:31,370 --> 00:09:35,610
Speaker 1: released a video on YouTube called Everybody Dance Now, and

118
00:09:36,490 --> 00:09:40,650
Speaker 1: somehow it really touched the nerve. Well, there's been an

119
00:09:40,650 --> 00:09:44,890
Speaker 1: explosion recently a new ways to manipulate media. Alexei notes

120
00:09:44,970 --> 00:09:49,410
Speaker 1: that the idea itself isn't new, It has a long history.

121
00:09:49,770 --> 00:09:53,450
Speaker 1: I can't help but ask, given that you come from Russia.

122
00:09:53,770 --> 00:09:58,770
Speaker 1: One of the premier users of doctoring photographs I think

123
00:09:58,930 --> 00:10:03,930
Speaker 1: was Stalin, who used the ability to manipulate images for

124
00:10:04,130 --> 00:10:07,650
Speaker 1: political effect. How did they do that? Can you think

125
00:10:07,650 --> 00:10:10,650
Speaker 1: of examples of this and like what was the technology? Then?

126
00:10:11,370 --> 00:10:17,890
Speaker 1: The urge to change photographs has been around basically since

127
00:10:17,890 --> 00:10:20,930
Speaker 1: the invention of photography. For example, there is a photograph

128
00:10:21,010 --> 00:10:25,650
Speaker 1: of Abraham Lincoln that still hangs in many classrooms. That's fake.

129
00:10:25,730 --> 00:10:30,450
Speaker 1: It's actually Calhoun with Lincoln's head attached to it. Alexei's

130
00:10:30,490 --> 00:10:34,570
Speaker 1: referring to John C. Calhoun, the South Carolina senator and

131
00:10:34,770 --> 00:10:39,970
Speaker 1: champion of slavery. A Civil War portrait artist superimposed a

132
00:10:40,050 --> 00:10:44,530
Speaker 1: photo of Lincoln's head onto an engraving of Calhoun's body

133
00:10:45,090 --> 00:10:49,650
Speaker 1: because he thought Lincoln's gangly frame wasn't dignified enough, and

134
00:10:49,730 --> 00:10:52,570
Speaker 1: so they just said Okay, we can use Calhoun. Let's

135
00:10:52,730 --> 00:10:55,690
Speaker 1: slap the Lincoln's head on his body. And then, of course,

136
00:10:56,250 --> 00:10:59,250
Speaker 1: as soon as you go into the twentieth century, as

137
00:10:59,250 --> 00:11:03,370
Speaker 1: soon as you get to dictatorships, this is a wonderful

138
00:11:04,050 --> 00:11:07,850
Speaker 1: toy for a dictator to use. So again, Stalin was

139
00:11:08,690 --> 00:11:12,370
Speaker 1: big fan of this. He would get rid of people

140
00:11:12,450 --> 00:11:15,850
Speaker 1: in photographs once they were out of favor, or once

141
00:11:15,890 --> 00:11:20,170
Speaker 1: they got jailed or killed. He would just basically get

142
00:11:20,170 --> 00:11:25,810
Speaker 1: them scratched out with reasonably crude techniques. Hitler did it,

143
00:11:25,930 --> 00:11:29,090
Speaker 1: Mao did it, Castro did it, Bresnev did it. I'm

144
00:11:29,090 --> 00:11:32,530
Speaker 1: sure US agencies have done it. Also, we have always

145
00:11:32,570 --> 00:11:36,610
Speaker 1: manipulated images with a desire to change history. This is

146
00:11:36,650 --> 00:11:39,930
Speaker 1: Honi f Reed. He's also a professor at Berkeley and

147
00:11:39,970 --> 00:11:43,130
Speaker 1: a friend of Alexey's. I'm a professor of computer science

148
00:11:43,130 --> 00:11:47,210
Speaker 1: and I'm an expert in digital forensics, where Alexei works

149
00:11:47,250 --> 00:11:50,930
Speaker 1: on making synthetic media. Honey has devoted his career to

150
00:11:51,170 --> 00:11:55,250
Speaker 1: identifying when synthetic media is being used to fool people,

151
00:11:55,810 --> 00:12:00,730
Speaker 1: that is, spotting fakes. He regularly collaborates on this mission

152
00:12:00,730 --> 00:12:05,130
Speaker 1: with Alexey so I, met Alyosha efros Ten, twenty years ago.

153
00:12:05,570 --> 00:12:11,090
Speaker 1: He is really incredibly creative and clever guy, and he

154
00:12:11,690 --> 00:12:14,050
Speaker 1: has done what I consider some of the most interesting

155
00:12:14,090 --> 00:12:16,650
Speaker 1: work in computer vision and computer graphics over the last

156
00:12:16,690 --> 00:12:20,810
Speaker 1: two decades. And if you really want to do forensics, well,

157
00:12:20,930 --> 00:12:23,290
Speaker 1: you have to partner with somebody like Aliosha. You have

158
00:12:23,370 --> 00:12:26,570
Speaker 1: to partner with a world class mind who knows how

159
00:12:26,570 --> 00:12:29,410
Speaker 1: to think about the synthesis side so that you can

160
00:12:29,450 --> 00:12:32,890
Speaker 1: synthesize the absolute best content and then think about how

161
00:12:32,930 --> 00:12:35,370
Speaker 1: to detect it. I think it's interesting that if you're

162
00:12:35,410 --> 00:12:38,250
Speaker 1: somebody on the synthesis side and developing the forensic there's

163
00:12:38,290 --> 00:12:39,890
Speaker 1: a little bit of a jekylin hide there, and I

164
00:12:39,930 --> 00:12:44,290
Speaker 1: think it's really fascinating. You know, the idea of altering photos,

165
00:12:44,930 --> 00:12:47,850
Speaker 1: it's not entirely new. How far back does this go?

166
00:12:48,730 --> 00:12:51,450
Speaker 1: So we used to have in the days of Stalin,

167
00:12:51,970 --> 00:12:57,690
Speaker 1: highly talented, highly skilled, time consuming, difficult process of manipulating images,

168
00:12:58,170 --> 00:13:03,090
Speaker 1: removing somebody, erasing something from the image, splicing faces together.

169
00:13:03,650 --> 00:13:06,810
Speaker 1: And then we moved into the digital age where now

170
00:13:06,850 --> 00:13:10,290
Speaker 1: a highly talented digital artist could remove one face and

171
00:13:10,330 --> 00:13:13,290
Speaker 1: add another phase, but it was still a time consuming

172
00:13:13,330 --> 00:13:17,090
Speaker 1: and required scale. In nineteen ninety four, The makers of

173
00:13:17,090 --> 00:13:21,170
Speaker 1: the movie Forrest Gump won an Oscar for Visual Effects

174
00:13:21,250 --> 00:13:25,410
Speaker 1: for their representations of the title character interacting with historical

175
00:13:25,530 --> 00:13:29,610
Speaker 1: figures like President John F. Kennedy gratulating how do the

176
00:13:29,690 --> 00:13:33,690
Speaker 1: field being all Americans? It's very good congratulation. How do

177
00:13:33,730 --> 00:13:40,090
Speaker 1: you feel I got? I believe that he had. Now

178
00:13:40,450 --> 00:13:42,770
Speaker 1: computers are doing all of the heavy lifting of what

179
00:13:42,930 --> 00:13:46,290
Speaker 1: used to be relegated to talented artists. The average person

180
00:13:46,370 --> 00:13:50,130
Speaker 1: now can use sophisticated technology to not just capture the recording,

181
00:13:50,170 --> 00:13:53,170
Speaker 1: but also manipulate it and then distribute it. The tools

182
00:13:53,330 --> 00:13:57,290
Speaker 1: used to create synthetic media have grown by leaps and bounds,

183
00:13:57,410 --> 00:14:00,330
Speaker 1: especially in the past few years, and so now we

184
00:14:00,410 --> 00:14:04,290
Speaker 1: have technology broadly called deep fake, but more specifically should

185
00:14:04,330 --> 00:14:08,490
Speaker 1: be called synthesized content, where you point an image or

186
00:14:08,530 --> 00:14:11,850
Speaker 1: a video or an audio to an AI or machine

187
00:14:11,930 --> 00:14:15,210
Speaker 1: learning system and it will replace the face for you.

188
00:14:15,450 --> 00:14:16,850
Speaker 1: I mean it can do that in an image, it

189
00:14:16,890 --> 00:14:19,210
Speaker 1: can do that in a video, or it can synthesize

190
00:14:19,250 --> 00:14:25,890
Speaker 1: audio for you in a particular person's voice. It's becomes

191
00:14:25,970 --> 00:14:30,810
Speaker 1: straightforward to swap people's faces. There's a popular YouTube video

192
00:14:31,170 --> 00:14:35,170
Speaker 1: that features tech pioneer Elon Musk's adult face on a

193
00:14:35,250 --> 00:14:39,290
Speaker 1: baby's body, and there's a famous meme where actor Nicholas

194
00:14:39,330 --> 00:14:43,570
Speaker 1: Cage's face replaces those of leading movie actors, both male

195
00:14:43,650 --> 00:14:47,130
Speaker 1: and female. You can put words into people's mouths and

196
00:14:47,210 --> 00:14:50,970
Speaker 1: make them jump and dance and run. You can even

197
00:14:51,010 --> 00:14:55,290
Speaker 1: resurrect powerful figures and have them deliver a fake speech

198
00:14:55,810 --> 00:15:04,570
Speaker 1: about a fake tragedy. From an Altered History, Chapter two,

199
00:15:05,530 --> 00:15:10,890
Speaker 1: Creating Nixon. The text of Nixon's Moon disaster speech that

200
00:15:10,930 --> 00:15:13,330
Speaker 1: we heard at the top of the show is actually

201
00:15:13,370 --> 00:15:16,330
Speaker 1: not fake. As I mentioned, it was written for President

202
00:15:16,450 --> 00:15:20,690
Speaker 1: Nixon as a contingency speech and thankfully never had to

203
00:15:20,690 --> 00:15:23,970
Speaker 1: be delivered. It's an amazing piece of writing. It was

204
00:15:24,010 --> 00:15:28,130
Speaker 1: written by Bill Safire, who was one of Nixon's speech writers.

205
00:15:28,530 --> 00:15:32,090
Speaker 1: This is artist in journalist Francesca Panetta. She's the co

206
00:15:32,250 --> 00:15:36,850
Speaker 1: director of the Nixon Fake or MIT's Moon Disaster Team.

207
00:15:37,490 --> 00:15:42,410
Speaker 1: She's also the creative director in MIT's Center for Advanced Virtuality.

208
00:15:42,770 --> 00:15:47,330
Speaker 1: I was doing experimental journalism at the Guardian newspaper. I

209
00:15:47,410 --> 00:15:50,490
Speaker 1: ran the Guardians Virtual Reality studio for the last three years.

210
00:15:50,770 --> 00:15:53,250
Speaker 1: The second half of the Moon Disaster team is sound

211
00:15:53,370 --> 00:15:57,650
Speaker 1: artist Halsey Bergund. My name is Halsey Bergund. I am

212
00:15:57,690 --> 00:16:00,690
Speaker 1: a sound artist and technologist, and I've had a lot

213
00:16:00,690 --> 00:16:04,890
Speaker 1: of experience with lots of sorts of audio enhanced with technology,

214
00:16:05,170 --> 00:16:08,530
Speaker 1: though this is my first experience with synthetic media, especially

215
00:16:08,610 --> 00:16:12,090
Speaker 1: since I typically focus on authenticity of voice and now

216
00:16:12,130 --> 00:16:15,570
Speaker 1: I'm kind of doing the opposite. So together, Halsey and

217
00:16:15,650 --> 00:16:19,530
Speaker 1: Francesca chose to automate a tragic moment in history that

218
00:16:19,770 --> 00:16:22,930
Speaker 1: never actually happened. I think it all started with it

219
00:16:23,010 --> 00:16:25,370
Speaker 1: being the fiftieth anniversary of the moon landing last year,

220
00:16:25,490 --> 00:16:28,290
Speaker 1: and add on top of that an election cycle in

221
00:16:28,330 --> 00:16:31,530
Speaker 1: this country, and dealing with this information, which is obviously

222
00:16:32,170 --> 00:16:36,730
Speaker 1: very important in election cycles. It was like lightbulbs went

223
00:16:36,810 --> 00:16:40,250
Speaker 1: on and we got very excited about pursuing it. It's

224
00:16:40,290 --> 00:16:43,810
Speaker 1: possible to make mediocre fakes pretty quickly and cheaply, but

225
00:16:43,930 --> 00:16:47,690
Speaker 1: Francesca and Halsey wanted high production values. So how does

226
00:16:47,730 --> 00:16:51,410
Speaker 1: one go about making a first rate fake presidential address?

227
00:16:51,890 --> 00:16:55,090
Speaker 1: There are two components. There's the visuals and there's the audio,

228
00:16:55,130 --> 00:16:59,370
Speaker 1: and they are completely different processes. So we decided to

229
00:16:59,410 --> 00:17:03,650
Speaker 1: go with a video dialogue replacement company called Kenny Ai,

230
00:17:04,090 --> 00:17:06,650
Speaker 1: who would do the visuals for us and then we

231
00:17:06,730 --> 00:17:10,810
Speaker 1: decided to go with re Speech, who are a dialogue

232
00:17:10,850 --> 00:17:15,690
Speaker 1: replacement company for the voice of Nixon. They tackled the

233
00:17:15,770 --> 00:17:19,090
Speaker 1: voice first, the more challenging of the two mediums. What

234
00:17:19,130 --> 00:17:21,850
Speaker 1: we were told to do was to get two to

235
00:17:21,930 --> 00:17:25,450
Speaker 1: three hours worth of Nixon talking. That was pretty easy

236
00:17:25,450 --> 00:17:28,810
Speaker 1: because the Nixon Library has hours and hours of Nixon,

237
00:17:29,370 --> 00:17:33,690
Speaker 1: mainly giving Vietnam's speeches. The Communist armies of North Vietnam

238
00:17:33,810 --> 00:17:37,290
Speaker 1: launched a massive inversion of South Vietnam. That audio was

239
00:17:37,370 --> 00:17:41,050
Speaker 1: then chopped up into chunks between one and three seconds long.

240
00:17:41,770 --> 00:17:46,610
Speaker 1: We found this incredibly patient actor called Lewis D. Wheeler.

241
00:17:47,250 --> 00:17:51,050
Speaker 1: Lewis would listen to the one second clip and then

242
00:17:51,210 --> 00:17:59,850
Speaker 1: he would repeat that and do what I believe was right.

243
00:18:00,250 --> 00:18:02,410
Speaker 1: Re Speech would say to us things like we need

244
00:18:02,450 --> 00:18:07,370
Speaker 1: to change the diagonal attention, which meant nothing to us. Yes,

245
00:18:07,530 --> 00:18:12,450
Speaker 1: we have a whole lot of potential band name going forward. Yeah,

246
00:18:14,210 --> 00:18:17,410
Speaker 1: Synthetic Nixon is another good one. So once we have

247
00:18:17,810 --> 00:18:21,210
Speaker 1: our Nixon model made out of these thousands of tiny clips,

248
00:18:21,530 --> 00:18:25,890
Speaker 1: it means that whatever our actor says will come out

249
00:18:26,050 --> 00:18:29,450
Speaker 1: then in Nixon's voice. So then what we did was

250
00:18:29,530 --> 00:18:34,130
Speaker 1: record the contingency speech of Nixon, and it meant that

251
00:18:34,170 --> 00:18:39,810
Speaker 1: we got Lewis's actually performance but in Nixon's voice. What

252
00:18:39,850 --> 00:18:42,930
Speaker 1: about the video part? I mean, the video was much easier.

253
00:18:42,970 --> 00:18:45,010
Speaker 1: We're talking a couple of days here and a tiny

254
00:18:45,010 --> 00:18:50,610
Speaker 1: amount of data just with Lewis's iPhone. We filmed him

255
00:18:50,650 --> 00:18:54,370
Speaker 1: reading the contingency speech once a couple of minutes of

256
00:18:54,450 --> 00:18:57,930
Speaker 1: him just chatting to camera, and that was it fate

257
00:18:59,130 --> 00:19:01,650
Speaker 1: that the men who went to the Moon to explore

258
00:19:01,810 --> 00:19:07,770
Speaker 1: in peace will stay on. You know. We were told

259
00:19:07,770 --> 00:19:11,050
Speaker 1: by Kenny Ai that everything would be the same in

260
00:19:11,090 --> 00:19:14,410
Speaker 1: the video apart from just the area around the mouth.

261
00:19:14,850 --> 00:19:18,010
Speaker 1: So every gesture of the hand, every blink, every time

262
00:19:18,050 --> 00:19:20,810
Speaker 1: he moved his face, all of that would stay the same,

263
00:19:21,330 --> 00:19:25,970
Speaker 1: but just the mouth basically would change. So we used

264
00:19:26,210 --> 00:19:30,330
Speaker 1: Nixon's resignation speech to have served in this office, it's

265
00:19:30,450 --> 00:19:35,050
Speaker 1: to have felt a very personal sense of it was

266
00:19:35,370 --> 00:19:38,570
Speaker 1: the speech of Nixon that looked the most somber, where

267
00:19:38,610 --> 00:19:40,810
Speaker 1: he seemed to have the most emotion in his face.

268
00:19:41,490 --> 00:19:45,690
Speaker 1: So what actually went on in the computer? Artificial intelligence

269
00:19:45,970 --> 00:19:50,610
Speaker 1: sometimes sounds inscrutable, but the basic ideas are quite simple.

270
00:19:51,290 --> 00:19:54,290
Speaker 1: In this case, it uses a type of computer program

271
00:19:54,330 --> 00:19:59,010
Speaker 1: called an auto encoder. It's trained to take complicated things,

272
00:19:59,450 --> 00:20:04,010
Speaker 1: say spoken sentences or pictures, encode them in a much

273
00:20:04,050 --> 00:20:08,250
Speaker 1: simpler form, and then decode them to recover the original

274
00:20:08,290 --> 00:20:12,210
Speaker 1: as bested. Care the encoder tries to reduce things to

275
00:20:12,250 --> 00:20:16,490
Speaker 1: their essence, throwing away most of the information but keeping

276
00:20:16,610 --> 00:20:19,690
Speaker 1: enough to do a good job of reconstructing it to

277
00:20:19,690 --> 00:20:23,330
Speaker 1: make a deep fake. Here's the trick. Train a speech

278
00:20:23,370 --> 00:20:27,090
Speaker 1: auto encoder for Nixon to Nixon, and a speech auto

279
00:20:27,210 --> 00:20:31,330
Speaker 1: encoder for actor to actor, but force them to use

280
00:20:31,450 --> 00:20:37,970
Speaker 1: the same encoder. Then you can input actor and decoded

281
00:20:38,330 --> 00:20:41,730
Speaker 1: as Nixon. If you have enough data. It's a piece

282
00:20:41,770 --> 00:20:48,330
Speaker 1: of cake around there. Carefully created video, the Moon Disaster

283
00:20:48,450 --> 00:20:53,170
Speaker 1: team created an entire art installation a nineteen sixties living

284
00:20:53,210 --> 00:20:56,890
Speaker 1: room with a fake vintage newspaper sharing the fake tragic

285
00:20:56,930 --> 00:21:01,730
Speaker 1: news while a fake Nixon speaks solemnly on a vintage

286
00:21:01,770 --> 00:21:05,530
Speaker 1: black and white television. Some people, when they were watching

287
00:21:05,570 --> 00:21:08,570
Speaker 1: the installation, they watched a number of times. You'd see them,

288
00:21:08,610 --> 00:21:10,610
Speaker 1: they'd watch at once, then they would watch it again,

289
00:21:11,410 --> 00:21:13,450
Speaker 1: staring at the lips to see if they could see

290
00:21:13,490 --> 00:21:17,770
Speaker 1: any lack of synchronicity. We had some people who thought

291
00:21:17,810 --> 00:21:21,850
Speaker 1: that perhaps Nixon had actually recorded this speech as a

292
00:21:21,890 --> 00:21:24,810
Speaker 1: contingency speech for it to go onto television. Lots of

293
00:21:24,850 --> 00:21:28,610
Speaker 1: folks who were listening, viewing, and even press folks just

294
00:21:28,690 --> 00:21:30,690
Speaker 1: immediately said, oh, the voice is real or whatever you

295
00:21:30,810 --> 00:21:34,770
Speaker 1: said these things that weren't accurate because they just felt

296
00:21:34,770 --> 00:21:37,330
Speaker 1: like there wasn't even a question. I suppose that is

297
00:21:37,330 --> 00:21:39,490
Speaker 1: what we wanted to achieve, But at the same time,

298
00:21:39,530 --> 00:21:42,530
Speaker 1: it was a little bit eye opening and like a

299
00:21:42,530 --> 00:21:50,890
Speaker 1: little scary. You know that that could happen. Chapter three,

300
00:21:51,610 --> 00:21:55,930
Speaker 1: Everybody dance. What do you see as just the wonderful

301
00:21:57,130 --> 00:22:01,050
Speaker 1: upside of having technologies like this? Yeah, I mean a

302
00:22:01,210 --> 00:22:06,850
Speaker 1: aion art is becoming a whole field in itself, so creatively,

303
00:22:07,170 --> 00:22:11,010
Speaker 1: there is enormous potential. One of the potential positive educational

304
00:22:11,090 --> 00:22:14,730
Speaker 1: uses of deep fake technology would be to bring historical

305
00:22:14,770 --> 00:22:17,930
Speaker 1: figures back to life to make learning more durable. I

306
00:22:17,970 --> 00:22:21,010
Speaker 1: think one could do that with bringing Abraham Lincoln back

307
00:22:21,010 --> 00:22:23,770
Speaker 1: to life and having him deliver speeches. Film companies are

308
00:22:23,770 --> 00:22:26,930
Speaker 1: really excited about re enactments. We're already beginning to see

309
00:22:26,970 --> 00:22:31,010
Speaker 1: this in films like Star Wars, when we're bringing people

310
00:22:31,010 --> 00:22:33,770
Speaker 1: like Carrie Fisher back to life. I mean that is

311
00:22:33,970 --> 00:22:37,090
Speaker 1: at the moment not being done through deep fake technologies.

312
00:22:37,130 --> 00:22:40,650
Speaker 1: This is using fatty traditional techniques of CGI at the moment,

313
00:22:40,890 --> 00:22:43,570
Speaker 1: So we still have to see our first deep fake

314
00:22:43,770 --> 00:22:47,330
Speaker 1: big cinema screen release. But this is just to come

315
00:22:47,450 --> 00:22:50,450
Speaker 1: like the technology is getting better and better. Not only

316
00:22:50,490 --> 00:22:53,530
Speaker 1: will we be able to potentially bring back actors and

317
00:22:53,570 --> 00:22:56,690
Speaker 1: actresses who are no longer alive and have them star

318
00:22:56,770 --> 00:22:59,610
Speaker 1: in movies, but an actor could make a model of

319
00:22:59,650 --> 00:23:02,210
Speaker 1: their own voice and then sell the use of that

320
00:23:02,330 --> 00:23:06,290
Speaker 1: voice to anybody to do a voiceover of whatever is wanted,

321
00:23:06,410 --> 00:23:09,410
Speaker 1: and so they could have twenty of the is going

322
00:23:09,410 --> 00:23:11,410
Speaker 1: on at the same time, and the sort of restriction

323
00:23:11,490 --> 00:23:15,170
Speaker 1: of their physical presence is no longer there. And that

324
00:23:15,250 --> 00:23:17,410
Speaker 1: might mean that, you know, Brad Pitt is in everything,

325
00:23:18,370 --> 00:23:21,530
Speaker 1: or it might just mean that lower budget films can

326
00:23:21,570 --> 00:23:24,050
Speaker 1: afford to have some of the higher cost talent. At

327
00:23:24,090 --> 00:23:26,370
Speaker 1: that point, you know, the top twenty actors could just

328
00:23:26,450 --> 00:23:29,210
Speaker 1: do everything. Yes, there's no doubt that there will be

329
00:23:29,210 --> 00:23:33,050
Speaker 1: winners and losers from these technologies, but the potential of

330
00:23:33,090 --> 00:23:36,530
Speaker 1: synthetic media goes way beyond the arts. There are possible

331
00:23:36,610 --> 00:23:40,730
Speaker 1: medical and therapeutic applications. There are companies that are working

332
00:23:40,810 --> 00:23:44,210
Speaker 1: very hard to allow people who have either lost their

333
00:23:44,290 --> 00:23:46,330
Speaker 1: voice or who never had a voice, to be able

334
00:23:46,370 --> 00:23:49,890
Speaker 1: to speak in a way that is either how they

335
00:23:49,970 --> 00:23:52,690
Speaker 1: used to speak or in a way that isn't a

336
00:23:52,690 --> 00:23:56,570
Speaker 1: canned voice that everybody has. Alexei ePROs and his students

337
00:23:56,690 --> 00:24:01,410
Speaker 1: discovered potential uses of synthetic media and medicine quite unintentionally

338
00:24:01,810 --> 00:24:06,050
Speaker 1: while working on their Everybody Dance Now project that could

339
00:24:06,050 --> 00:24:10,970
Speaker 1: turn anyone into a ballerina. Were kind of surprised for

340
00:24:11,210 --> 00:24:14,650
Speaker 1: all the positive feedback we got. We've got emails from

341
00:24:14,770 --> 00:24:17,650
Speaker 1: people who were quadriplegic and they asked us if we

342
00:24:17,690 --> 00:24:21,450
Speaker 1: could make them dance, and it was very unexpected. So

343
00:24:21,490 --> 00:24:24,570
Speaker 1: now we are trying to get the software to be

344
00:24:24,650 --> 00:24:27,890
Speaker 1: in a state where people can use it, because yeah,

345
00:24:27,930 --> 00:24:33,530
Speaker 1: it's somehow it did hit a nerve with folks. Chapter

346
00:24:33,690 --> 00:24:39,250
Speaker 1: four Unicorns in the Andes. The past few years have

347
00:24:39,370 --> 00:24:43,250
Speaker 1: seen amazing advances in the creation of synthetic media through

348
00:24:43,330 --> 00:24:47,690
Speaker 1: artificial intelligence. The technology now goes far beyond fitting one

349
00:24:47,730 --> 00:24:51,610
Speaker 1: face over another face in a video. A recent breakthrough

350
00:24:51,690 --> 00:24:56,130
Speaker 1: has made it possible to create entirely new and very

351
00:24:56,250 --> 00:25:01,370
Speaker 1: convincing content out of thin air. The breakthrough called generative

352
00:25:01,530 --> 00:25:06,690
Speaker 1: adversarial networks or GAMS, came from a machine learning researcher

353
00:25:06,730 --> 00:25:11,530
Speaker 1: at Google named Ian Goodfellow. Like auto encoders, the basic

354
00:25:11,610 --> 00:25:15,930
Speaker 1: idea is simple but brilliant. Suppose you want to create

355
00:25:16,050 --> 00:25:20,570
Speaker 1: amazingly realistic photos of people who don't exist. While you

356
00:25:20,610 --> 00:25:24,810
Speaker 1: build a GAN consisting of two computer programs, a photo

357
00:25:25,010 --> 00:25:29,170
Speaker 1: generator that learns to generate fake photos and a photo

358
00:25:29,330 --> 00:25:35,170
Speaker 1: discriminator that learns to discriminate or identify fake photos from

359
00:25:35,170 --> 00:25:38,850
Speaker 1: a vast collection of real photos. You then let the

360
00:25:38,850 --> 00:25:43,690
Speaker 1: two programs compete, continually tweaking their code to outsmart each other.

361
00:25:44,490 --> 00:25:48,330
Speaker 1: By the time they're done, the GAN can generate amazingly

362
00:25:48,330 --> 00:25:51,730
Speaker 1: convincing fakes. You can see for yourself if you go

363
00:25:51,770 --> 00:25:55,810
Speaker 1: to the website this Person does Not Exist dot com.

364
00:25:56,650 --> 00:25:59,890
Speaker 1: Every time you refresh the page, you're shown a new

365
00:26:00,130 --> 00:26:04,410
Speaker 1: uncanny image of a person who, as the website says,

366
00:26:04,890 --> 00:26:09,450
Speaker 1: does not and never did exist. Francescan I actually tried

367
00:26:09,530 --> 00:26:17,650
Speaker 1: out the website. This young Asian woman. She's got great complexion.

368
00:26:17,890 --> 00:26:20,850
Speaker 1: Envious of that neat black hair with a fringe pink

369
00:26:20,850 --> 00:26:24,130
Speaker 1: lipstick and a slightly dreamy look as she's kind of

370
00:26:24,130 --> 00:26:30,330
Speaker 1: gazing off to her left. Oh, here's a woman who

371
00:26:30,330 --> 00:26:32,770
Speaker 1: looks like she could be a neighbor of mine in Cambridge,

372
00:26:33,410 --> 00:26:38,090
Speaker 1: probably about sixty five. She's got nice wire framed glasses,

373
00:26:38,370 --> 00:26:43,610
Speaker 1: layered hair. Her earrings don't actually match, but that could

374
00:26:43,610 --> 00:26:46,690
Speaker 1: just be her distinctive style. I mean, of course, she

375
00:26:47,050 --> 00:26:52,330
Speaker 1: doesn't really exist. It's hard to argue that gams aren't

376
00:26:52,330 --> 00:26:57,970
Speaker 1: creating original art. In fact, an artist collective recently used

377
00:26:58,010 --> 00:27:03,050
Speaker 1: a GAM to create a French Impressionist style portrait. When

378
00:27:03,170 --> 00:27:07,210
Speaker 1: Christie's sold it at auction, it fetched an eye popping

379
00:27:07,290 --> 00:27:11,970
Speaker 1: four hundred and thirty two thousand dollars. Alexei Efros, the

380
00:27:12,010 --> 00:27:17,130
Speaker 1: Berkeley professor, recently pushed gans a step further, creating something

381
00:27:17,210 --> 00:27:21,770
Speaker 1: called cycle gans. By connecting two gans together in a

382
00:27:21,770 --> 00:27:27,290
Speaker 1: clever way, cycle gans can transform a monet painting into

383
00:27:27,290 --> 00:27:31,210
Speaker 1: what's seemingly a photograph of the same scene, or turn

384
00:27:31,250 --> 00:27:34,930
Speaker 1: a summer landscape into a winter landscape of the same view.

385
00:27:35,810 --> 00:27:39,290
Speaker 1: Alexei's cycle gans seem like magic. If you were to

386
00:27:39,290 --> 00:27:44,810
Speaker 1: add in virtual reality, the possibilities become mind blowing. You

387
00:27:45,090 --> 00:27:51,490
Speaker 1: may be reminiscing about walking down Saint German and Paris

388
00:27:51,570 --> 00:27:54,450
Speaker 1: and with a few clicks. You are there, and you're

389
00:27:54,530 --> 00:27:57,410
Speaker 1: walking down the boulevard, and you're looking at all the buildings,

390
00:27:57,450 --> 00:28:00,970
Speaker 1: and maybe you can even switch to a different year.

391
00:28:01,130 --> 00:28:06,330
Speaker 1: And I think that is I think very exciting as

392
00:28:06,370 --> 00:28:10,490
Speaker 1: a way to mentally travel to different places. So if

393
00:28:10,490 --> 00:28:12,650
Speaker 1: you do this in VR, I mean, can you imagine

394
00:28:13,010 --> 00:28:16,810
Speaker 1: classes going on a class visit to ancient Rome. That's right,

395
00:28:17,170 --> 00:28:21,890
Speaker 1: you could imagine from how a particular city like Chrome.

396
00:28:22,290 --> 00:28:24,770
Speaker 1: Luke's now trying to extrapolate to how it looked in

397
00:28:24,810 --> 00:28:29,210
Speaker 1: the past. It turns out that gans aren't just transforming images.

398
00:28:29,890 --> 00:28:32,690
Speaker 1: I spoke with a friend who's very familiar with another

399
00:28:32,810 --> 00:28:37,250
Speaker 1: remarkable application of the technology. My name is Reid Hoffman.

400
00:28:37,370 --> 00:28:40,650
Speaker 1: I'm a podcaster of Master's Scale. I'm a partner at Greylock,

401
00:28:40,690 --> 00:28:43,570
Speaker 1: which is where we're sitting right now, co founder of LinkedIn,

402
00:28:43,770 --> 00:28:47,810
Speaker 1: and then a variety of other eccentric hobbies. Reid is

403
00:28:47,850 --> 00:28:52,170
Speaker 1: a board member of an unusual organization called open AI.

404
00:28:52,650 --> 00:28:55,850
Speaker 1: Open a Eyes is highly concerned with artificial general intelligence

405
00:28:55,930 --> 00:28:59,890
Speaker 1: human level intelligence. I helped Sam Altman and Elon Musk

406
00:29:00,130 --> 00:29:05,410
Speaker 1: standing up. The basic concern was that if one company

407
00:29:05,490 --> 00:29:09,850
Speaker 1: created and deployed that that could be is balancing in

408
00:29:09,890 --> 00:29:12,970
Speaker 1: all kinds of ways. And so the thought is, if

409
00:29:12,970 --> 00:29:15,770
Speaker 1: it could be created, we should make sure that there

410
00:29:15,850 --> 00:29:19,090
Speaker 1: is essentially a nonprofit that is creating this and that

411
00:29:19,170 --> 00:29:23,530
Speaker 1: can make that technology available at selective time, slices to

412
00:29:24,250 --> 00:29:28,690
Speaker 1: industry as a whole, government, etc. Last year, open ai

413
00:29:28,890 --> 00:29:33,490
Speaker 1: released a program that uses gams to write language from

414
00:29:33,490 --> 00:29:38,450
Speaker 1: a short opening prompt. The system, called GPT two, can

415
00:29:38,490 --> 00:29:42,130
Speaker 1: spin a convincing article or story instead of a deep

416
00:29:42,170 --> 00:29:47,010
Speaker 1: fake video. It's deep fake text. It's pretty amazing actually.

417
00:29:47,450 --> 00:29:52,770
Speaker 1: For example, open ai researchers gave the program the following prompt.

418
00:29:53,650 --> 00:29:56,970
Speaker 1: In a shocking finding, scientists discovered a herd of unicorns

419
00:29:57,050 --> 00:30:00,610
Speaker 1: living in a remote, previously unexplored valley in the Andes Mountains.

420
00:30:01,290 --> 00:30:04,250
Speaker 1: Even more surprising to the researches was the fact that

421
00:30:04,290 --> 00:30:08,290
Speaker 1: the unicorns spoke perfect to English. GPT two took it

422
00:30:08,330 --> 00:30:13,250
Speaker 1: from there, the livering nine crisp paragraphs on the landmark discovery.

423
00:30:13,850 --> 00:30:16,050
Speaker 1: I asked Franz to read a bit from the story.

424
00:30:16,610 --> 00:30:20,490
Speaker 1: Doctor Jorge Perez, an evolutionary biologists from the University of

425
00:30:20,570 --> 00:30:25,130
Speaker 1: Lapez and several companions, were exploring the Andes Mountains when

426
00:30:25,130 --> 00:30:28,530
Speaker 1: they found a small valley with no other animals or humans.

427
00:30:29,210 --> 00:30:31,810
Speaker 1: Perez noticed that the valley had what appeared to be

428
00:30:31,970 --> 00:30:35,370
Speaker 1: a natural fountains surrounded by two peaks of rock and

429
00:30:35,450 --> 00:30:39,570
Speaker 1: silver snow. Perez and the others then ventured further into

430
00:30:39,610 --> 00:30:41,930
Speaker 1: the valley. By the time we reached the top of

431
00:30:41,930 --> 00:30:44,690
Speaker 1: one peak, the water looked blue with some crystals on top,

432
00:30:44,770 --> 00:30:48,130
Speaker 1: said Perez. Perez and his friends were astonished to see

433
00:30:48,130 --> 00:30:55,250
Speaker 1: the unicorn. Heard. Tell me some of the great things

434
00:30:55,290 --> 00:30:59,410
Speaker 1: you can do with language generation, well, say, for example, entertainment,

435
00:30:59,850 --> 00:31:03,530
Speaker 1: generate stories that could be fresh and interesting and new

436
00:31:03,610 --> 00:31:08,130
Speaker 1: and personal for every child. Embed educational things in those

437
00:31:08,170 --> 00:31:11,130
Speaker 1: stories of the on into the fact that the story

438
00:31:11,250 --> 00:31:14,690
Speaker 1: is involving them and their friends, but also now brings

439
00:31:14,730 --> 00:31:19,170
Speaker 1: in grammar and math and other kinds of things as

440
00:31:19,490 --> 00:31:24,010
Speaker 1: the doing it generate explanatory material of this kind of

441
00:31:24,170 --> 00:31:27,890
Speaker 1: education that works best for this audience, for this kind

442
00:31:27,890 --> 00:31:29,370
Speaker 1: of people, like we want to have this kind of

443
00:31:29,410 --> 00:31:31,130
Speaker 1: math or this kind of physics, or this kind of

444
00:31:31,170 --> 00:31:34,730
Speaker 1: history or this kind of poetry explained in the right way,

445
00:31:34,850 --> 00:31:37,530
Speaker 1: and also the style of language right like you know

446
00:31:37,730 --> 00:31:41,890
Speaker 1: native city x language. When open ai announced its breakthrough

447
00:31:41,930 --> 00:31:45,970
Speaker 1: program for text generation, it took the unusual step of

448
00:31:46,090 --> 00:31:49,130
Speaker 1: not releasing the full powered version because it was worried

449
00:31:49,170 --> 00:31:52,810
Speaker 1: about the possible consequences. Now, part of the open AI

450
00:31:52,890 --> 00:31:56,330
Speaker 1: decision to say we're going to release a smaller model

451
00:31:56,570 --> 00:31:59,410
Speaker 1: than the one we did is because we think that

452
00:31:59,450 --> 00:32:02,250
Speaker 1: the deep fake problem hasn't been solved. And by the way,

453
00:32:02,450 --> 00:32:04,570
Speaker 1: some people complained about that, because they said, well, you're

454
00:32:04,570 --> 00:32:07,250
Speaker 1: slowing down our ability to do progress. And so for

455
00:32:07,410 --> 00:32:09,490
Speaker 1: the answer and say, look, when these are at least

456
00:32:09,650 --> 00:32:13,130
Speaker 1: to the entire public, we cannot control the downside as

457
00:32:13,130 --> 00:32:20,650
Speaker 1: well as upsides. Downsides from art to therapy to virtual

458
00:32:20,730 --> 00:32:26,850
Speaker 1: time travel, personalized stories and education, synthetic media has amazing upsides.

459
00:32:27,530 --> 00:32:34,690
Speaker 1: What could possibly go wrong? Chapter five? What could possibly

460
00:32:34,770 --> 00:32:39,330
Speaker 1: go wrong? The downsides are actually not hard to find.

461
00:32:39,890 --> 00:32:44,570
Speaker 1: The ability to reshape reality brings extraordinary power, and people

462
00:32:44,690 --> 00:32:49,210
Speaker 1: inevitably use power to control other people. It should be

463
00:32:49,250 --> 00:32:52,650
Speaker 1: no surprise, therefore, that ninety six percent of fake videos

464
00:32:52,690 --> 00:32:58,650
Speaker 1: posted online are non consensual pornography videos, almost always of

465
00:32:58,650 --> 00:33:03,450
Speaker 1: women manipulated to depict sex acts that never actually occurred.

466
00:33:04,210 --> 00:33:07,450
Speaker 1: I spoke with a professor who studies deep fakes, including

467
00:33:07,450 --> 00:33:11,930
Speaker 1: digital attempts to control women's bodies. I'm Danielle Citron and

468
00:33:12,170 --> 00:33:15,410
Speaker 1: I am a law professor at Boston University School of Law.

469
00:33:15,610 --> 00:33:20,610
Speaker 1: I write about privacy, technology, automation. My newest work and

470
00:33:20,690 --> 00:33:23,130
Speaker 1: my next book is going to be about sexual privacy.

471
00:33:24,130 --> 00:33:27,450
Speaker 1: So I've worked in and around consumer privacy, individual rights,

472
00:33:27,490 --> 00:33:30,490
Speaker 1: civil rights. I write a lot about free speech and

473
00:33:30,530 --> 00:33:34,890
Speaker 1: then automated systems. When do you first become aware of

474
00:33:35,010 --> 00:33:38,050
Speaker 1: deep fakes? Do you remember when this cross rit I did? So,

475
00:33:38,570 --> 00:33:41,730
Speaker 1: there was a Reddit thread devoted to, you know, fake

476
00:33:42,090 --> 00:33:46,130
Speaker 1: pornography movies of Gal Jadot Emma Watson. But the reddit

477
00:33:46,170 --> 00:33:50,690
Speaker 1: thread sort of spooled not just from celebrities but ordinary people,

478
00:33:51,170 --> 00:33:53,570
Speaker 1: and so you had rereditors asking each other, how do

479
00:33:53,650 --> 00:33:55,890
Speaker 1: I make a deep fake sex video of max girlfriend?

480
00:33:55,930 --> 00:33:58,770
Speaker 1: I have thirty pictures? And then other redditors would chime

481
00:33:58,810 --> 00:34:01,850
Speaker 1: in and say, look at this YouTube tutorial. You can

482
00:34:01,930 --> 00:34:04,850
Speaker 1: absolutely make a deep fake sex video of your ex

483
00:34:05,330 --> 00:34:08,250
Speaker 1: with thirty pictures. I've done it with twenty. In November

484
00:34:08,370 --> 00:34:13,690
Speaker 1: two thousand seventeen, an anonymous reditor began posting synthesized porn

485
00:34:13,810 --> 00:34:18,130
Speaker 1: videos under the pseudonym deep fakes, perhaps a nod to

486
00:34:18,210 --> 00:34:21,450
Speaker 1: the deep learning technology used to create them as well

487
00:34:21,490 --> 00:34:26,250
Speaker 1: as the nineteen seventies porn film deep Throat. The Internet

488
00:34:26,530 --> 00:34:30,570
Speaker 1: quickly adopted the term deep fakes and broadened its meanings

489
00:34:30,570 --> 00:34:34,970
Speaker 1: beyond pornography. To create the videos, he used celebrity faces

490
00:34:35,050 --> 00:34:39,410
Speaker 1: from Google image search and YouTube videos and then trains

491
00:34:39,450 --> 00:34:44,130
Speaker 1: an algorithm on that content together with pornographic videos. Have

492
00:34:44,250 --> 00:34:49,690
Speaker 1: you seen deep fake pornography videos? Yes, so still pretty crude,

493
00:34:49,730 --> 00:34:53,050
Speaker 1: so you probably can tell that it's a fake, but

494
00:34:53,370 --> 00:34:57,810
Speaker 1: for the person who's inserted into pornography, it's devastating. You

495
00:34:57,970 --> 00:35:03,170
Speaker 1: use the neural network technology, the artificial intelligence technology to

496
00:35:03,210 --> 00:35:08,930
Speaker 1: create out of digital whole cloth pornography videos using proba

497
00:35:09,370 --> 00:35:13,050
Speaker 1: real pornography and then inserting the person in the pornography

498
00:35:13,250 --> 00:35:15,970
Speaker 1: so they become the female actress. If it's a female,

499
00:35:16,010 --> 00:35:20,690
Speaker 1: it's usually a female in that video. My name is

500
00:35:20,810 --> 00:35:26,690
Speaker 1: Noel Martin and I am an activist and Laura Form

501
00:35:26,810 --> 00:35:31,650
Speaker 1: campaigner in Australia. Noel is twenty six years old and

502
00:35:31,770 --> 00:35:36,330
Speaker 1: she lives in Perth, Australia. So the first time that

503
00:35:36,490 --> 00:35:43,050
Speaker 1: I discovered myself on pornographic sites was when I was

504
00:35:43,290 --> 00:35:48,890
Speaker 1: eighteen and out of curiosity, decided to Google image reverse

505
00:35:48,890 --> 00:35:53,210
Speaker 1: search myself in an instant, like in a less than

506
00:35:53,250 --> 00:35:58,170
Speaker 1: a millisecond, my life completely changed. At first, it started

507
00:35:58,170 --> 00:36:02,610
Speaker 1: with photos still images stolen from Noel's social media accounts.

508
00:36:03,090 --> 00:36:08,730
Speaker 1: They were then doctoring my face from ordinary images and

509
00:36:09,450 --> 00:36:14,650
Speaker 1: superimposing those onto the bodies of women depicting me having

510
00:36:14,690 --> 00:36:19,370
Speaker 1: sexual intercourse. It proved impossible to identify who was manipulating

511
00:36:19,450 --> 00:36:23,050
Speaker 1: Nowell's image in this way. It's still unclear today, which

512
00:36:23,050 --> 00:36:25,850
Speaker 1: made it difficult for her to seek legal action. I

513
00:36:25,890 --> 00:36:31,730
Speaker 1: went to the police soon after, I contacted government agencies,

514
00:36:32,650 --> 00:36:36,770
Speaker 1: tried getting a private investigator. Essentially, there's nothing that they

515
00:36:36,770 --> 00:36:40,850
Speaker 1: could do. The sites are hosted overseas, the perpetrators are

516
00:36:40,890 --> 00:36:44,530
Speaker 1: probably overseas. The reaction was at the end of the day,

517
00:36:44,570 --> 00:36:48,050
Speaker 1: I think you can contact the webmasters to try and

518
00:36:48,090 --> 00:36:51,370
Speaker 1: get things deleted. You know, you can adjust your privacy

519
00:36:51,410 --> 00:36:56,210
Speaker 1: setting so that nothing is available to anyone publicly. It

520
00:36:56,250 --> 00:37:01,370
Speaker 1: was an unwinnable situation. Then things started to escalate. In

521
00:37:01,410 --> 00:37:05,570
Speaker 1: twenty eighteen, who Well saw a synthesized pornographic video of

522
00:37:05,570 --> 00:37:09,570
Speaker 1: herself and I believe that it was done for the

523
00:37:09,610 --> 00:37:16,050
Speaker 1: purposes of silencing me because I've been very public about

524
00:37:16,130 --> 00:37:19,970
Speaker 1: my story and advocating for change. So I had actually

525
00:37:20,050 --> 00:37:25,170
Speaker 1: gotten email from a fake email address, and you know,

526
00:37:25,210 --> 00:37:28,570
Speaker 1: I clicked the link. I was actually at work. It

527
00:37:28,650 --> 00:37:33,330
Speaker 1: was a video of me having sexual intercourse. The title

528
00:37:33,410 --> 00:37:36,770
Speaker 1: had my name, the face of the woman in it

529
00:37:37,330 --> 00:37:41,290
Speaker 1: was edited so that it was my face, and you know,

530
00:37:41,330 --> 00:37:46,050
Speaker 1: all the tags were like Noel Martin Australia, feminist, and

531
00:37:46,970 --> 00:37:51,930
Speaker 1: it didn't look real, but the context of everything with

532
00:37:52,010 --> 00:37:56,730
Speaker 1: the title my face, with the tags all points to

533
00:37:57,330 --> 00:38:00,730
Speaker 1: me being depicted in this video. The fakes were of

534
00:38:00,810 --> 00:38:05,290
Speaker 1: poor quality, but poor and consumers are in a discriminating lot,

535
00:38:05,730 --> 00:38:08,210
Speaker 1: and many people reacted to them as if they were real.

536
00:38:08,410 --> 00:38:12,330
Speaker 1: The public reaction was horrifying to me. I was a victim,

537
00:38:12,370 --> 00:38:16,050
Speaker 1: blamed and slut shamed, and it's definitely limited the course

538
00:38:16,090 --> 00:38:20,570
Speaker 1: of where I can go in terms of career and employment.

539
00:38:21,090 --> 00:38:24,810
Speaker 1: Noel finished a degree in law and began campaigning to

540
00:38:24,810 --> 00:38:29,650
Speaker 1: criminalize this sort of content. My advocacy and my activism

541
00:38:29,850 --> 00:38:32,730
Speaker 1: started off because I had a lived experience of this,

542
00:38:32,890 --> 00:38:36,570
Speaker 1: and I experienced it at a time where it wasn't

543
00:38:36,610 --> 00:38:43,450
Speaker 1: criminalized in Australia. The distribution of altered intimate images or

544
00:38:43,610 --> 00:38:49,090
Speaker 1: altered intimate videos and so I had to petition, meet

545
00:38:49,130 --> 00:38:53,210
Speaker 1: with my politicians in my area. I wrote a number

546
00:38:53,210 --> 00:38:55,970
Speaker 1: of articles, I spoke to the media, and I was

547
00:38:56,450 --> 00:39:00,130
Speaker 1: involved in the law reform in Australia in a number

548
00:39:00,130 --> 00:39:04,050
Speaker 1: of jurisdictions in Western Australia and New South Wales, and

549
00:39:04,490 --> 00:39:08,290
Speaker 1: I ended up being involved in two press conferences with

550
00:39:08,370 --> 00:39:12,570
Speaker 1: the Attorney generals of each state at the announcement of

551
00:39:12,610 --> 00:39:18,530
Speaker 1: the law that was criminalizing this abuse. Today, in part

552
00:39:18,570 --> 00:39:22,090
Speaker 1: because of Noel's activism, it is illegal in Australia to

553
00:39:22,170 --> 00:39:27,210
Speaker 1: distribute intimate images without consent, including intimate images and videos

554
00:39:27,410 --> 00:39:31,530
Speaker 1: that have been altered. Although it doesn't encompass all malicious

555
00:39:31,570 --> 00:39:39,650
Speaker 1: synthetic media, Noel has made a solid start. Chapter six,

556
00:39:40,170 --> 00:39:45,810
Speaker 1: Scissors and Glue. The videos depicting Noel Martin were nowhere

557
00:39:45,930 --> 00:39:49,850
Speaker 1: near as sophisticated as those made by the Moon Disastered team.

558
00:39:50,290 --> 00:39:54,410
Speaker 1: They were more cheap fakes than deep fakes, and yet

559
00:39:54,490 --> 00:39:56,930
Speaker 1: the point didn't have to be perfect to be devastating.

560
00:39:57,690 --> 00:40:00,690
Speaker 1: The same turns out to be true in politics. To

561
00:40:00,770 --> 00:40:05,490
Speaker 1: understand the power of fakes, you have to understand human psychology.

562
00:40:05,570 --> 00:40:08,250
Speaker 1: It turns out that people are pretty easy to fool.

563
00:40:09,250 --> 00:40:12,250
Speaker 1: Carry I was running for President of the US. His

564
00:40:12,650 --> 00:40:16,450
Speaker 1: stance on the Vietnam War was controversial. Jane Fonda, of course,

565
00:40:16,530 --> 00:40:18,770
Speaker 1: was a very controversial figure back then because of her

566
00:40:18,810 --> 00:40:21,890
Speaker 1: anti war stand. What have we become as a nation

567
00:40:21,930 --> 00:40:23,890
Speaker 1: if we call the men heroes that were used by

568
00:40:23,890 --> 00:40:26,410
Speaker 1: the Pentagon to try to exterminate an entire people? What

569
00:40:26,490 --> 00:40:28,610
Speaker 1: business have we to try to exterminate a people? And

570
00:40:28,690 --> 00:40:30,810
Speaker 1: somebody had created a photo of the two of them

571
00:40:30,890 --> 00:40:33,770
Speaker 1: sharing a stage and an anti war rally with the

572
00:40:33,810 --> 00:40:37,090
Speaker 1: hopes of damaging the Carry campaign. The photo was fake.

573
00:40:37,330 --> 00:40:39,650
Speaker 1: They had never shared a stage together. They just took

574
00:40:39,690 --> 00:40:42,770
Speaker 1: two images, probably put it into some standard photo editing

575
00:40:42,770 --> 00:40:46,090
Speaker 1: software like a Photoshop, and just put a headline around it,

576
00:40:46,170 --> 00:40:48,690
Speaker 1: and out to the world it went. And I will

577
00:40:48,730 --> 00:40:51,810
Speaker 1: tell you I remember the most fascinating interview I've heard

578
00:40:51,850 --> 00:40:55,650
Speaker 1: in a long time was right after the election, Kerry

579
00:40:55,690 --> 00:40:58,930
Speaker 1: of course lost, and a voter was being interviewed and

580
00:40:59,010 --> 00:41:01,570
Speaker 1: asked how they voted, and he said he couldn't vote

581
00:41:01,610 --> 00:41:04,050
Speaker 1: for Carry, and the interview said, well why not? And

582
00:41:04,170 --> 00:41:06,610
Speaker 1: the gentleman said, I couldn't get that photo of John

583
00:41:06,690 --> 00:41:09,810
Speaker 1: Carry and Jane Fonda out of my head. And the interviews, well,

584
00:41:09,970 --> 00:41:12,090
Speaker 1: you know, that photo is fake, and the guy said,

585
00:41:12,170 --> 00:41:15,010
Speaker 1: much to my surprise, yes, but I couldn't get it

586
00:41:15,050 --> 00:41:18,330
Speaker 1: out of my mind. And this shows you the power

587
00:41:18,370 --> 00:41:21,450
Speaker 1: of visual imagery, Like even after I tell you something

588
00:41:21,530 --> 00:41:24,330
Speaker 1: is fake, it still had an impact on somebody, and

589
00:41:24,450 --> 00:41:27,650
Speaker 1: I thought, Wow, we're in a lot of trouble because

590
00:41:27,890 --> 00:41:30,050
Speaker 1: it's very very hard to put the cat back into

591
00:41:30,050 --> 00:41:32,810
Speaker 1: the bag. Once that content is out there, you can't

592
00:41:32,890 --> 00:41:37,650
Speaker 1: undo it. So seeing is believing, even above thinking Yeah,

593
00:41:37,690 --> 00:41:40,050
Speaker 1: that seems to be the rule. There is very good

594
00:41:40,050 --> 00:41:43,330
Speaker 1: evidence from the social science literature that it's very very

595
00:41:43,370 --> 00:41:46,570
Speaker 1: difficult to correct the record after the mistakes are out there.

596
00:41:46,930 --> 00:41:50,650
Speaker 1: Law professor Danielle Citram also notes that humans tend to

597
00:41:50,690 --> 00:41:55,530
Speaker 1: pass on information without thinking, which triggers what she calls

598
00:41:55,930 --> 00:42:00,570
Speaker 1: information cascades. Information cascades is a phenomenon where we have

599
00:42:00,650 --> 00:42:03,850
Speaker 1: so much information overload that when someone sends us something,

600
00:42:03,970 --> 00:42:06,570
Speaker 1: some information, and we trust that person, we pass it on.

601
00:42:06,890 --> 00:42:10,610
Speaker 1: We don't even check it's veracity, and so information can

602
00:42:10,650 --> 00:42:16,250
Speaker 1: go viral fairly quickly because we're not terribly reflective, because

603
00:42:16,250 --> 00:42:21,130
Speaker 1: we act on impulse. Danielle says that information cascades have

604
00:42:21,210 --> 00:42:24,010
Speaker 1: been given new life in the twenty first century through

605
00:42:24,090 --> 00:42:27,770
Speaker 1: social media. Think about the twentieth century phenomenon, where do

606
00:42:27,770 --> 00:42:33,090
Speaker 1: we get most of our information from trusted sources, trusted newspapers,

607
00:42:33,530 --> 00:42:37,130
Speaker 1: trusted major couple of TV channels. Growing up, we only

608
00:42:37,170 --> 00:42:40,210
Speaker 1: had you know, we didn't have a million, and they

609
00:42:41,130 --> 00:42:44,930
Speaker 1: were adhering to journalistic ethics and commitments to truth and

610
00:42:44,970 --> 00:42:48,570
Speaker 1: neutrality and notion that you can't publish something without checking it.

611
00:42:49,090 --> 00:42:52,810
Speaker 1: Now we are publishing information that most people say. We're

612
00:42:52,850 --> 00:42:56,330
Speaker 1: lying on our peers and our friends. Social media platforms

613
00:42:56,370 --> 00:43:00,130
Speaker 1: are designed to tailor our information diet to what we

614
00:43:00,210 --> 00:43:03,610
Speaker 1: want and to our pre existing views, so we're locked

615
00:43:03,610 --> 00:43:07,490
Speaker 1: in a digital echo chamber. We think everybody agrees with us.

616
00:43:07,930 --> 00:43:12,010
Speaker 1: We pass on that information. We haven't checked the veracity.

617
00:43:12,090 --> 00:43:15,250
Speaker 1: It goes wild and we're especially likely to pass it

618
00:43:15,290 --> 00:43:18,850
Speaker 1: on if it's negative and novel. Why's that? It's just

619
00:43:19,010 --> 00:43:22,530
Speaker 1: like it's one of our weaknesses. We know how gossip

620
00:43:22,570 --> 00:43:26,850
Speaker 1: goes like wildfire online. So like Hillary Clinton as running

621
00:43:27,010 --> 00:43:31,530
Speaker 1: a sex ring. That's crazy. Oh my god, Eric, did

622
00:43:31,570 --> 00:43:34,170
Speaker 1: you hear about that. I'll post it on Facebook. Eric,

623
00:43:34,210 --> 00:43:37,410
Speaker 1: you pass it on. We just can't help ourselves, and

624
00:43:37,530 --> 00:43:40,650
Speaker 1: it is much in the way that we love suits

625
00:43:40,650 --> 00:43:44,970
Speaker 1: and fats and pizza. You know, we indulge. We don't think.

626
00:43:45,890 --> 00:43:49,570
Speaker 1: On some sense, this phenomenon is an old phenomenon. Right

627
00:43:49,770 --> 00:43:54,250
Speaker 1: is the famous observation by Mark Twain about how a

628
00:43:54,330 --> 00:43:57,170
Speaker 1: lie gets halfway around the world before the truth gets

629
00:43:57,170 --> 00:43:59,010
Speaker 1: its pants. Hall. Yeah, the truth still in the bedroom

630
00:43:59,010 --> 00:44:03,130
Speaker 1: getting dressed, and we often will see the lie, but

631
00:44:03,210 --> 00:44:08,650
Speaker 1: the rebuttal is not seen. It's often lost in the

632
00:44:08,770 --> 00:44:12,650
Speaker 1: noise ways of the defamatory statements. That is not new.

633
00:44:12,730 --> 00:44:15,930
Speaker 1: But what is new is a number of things about

634
00:44:15,970 --> 00:44:28,610
Speaker 1: our information ecosystem are our force multipliers Chapter seven, Truth Decay.

635
00:44:30,090 --> 00:44:34,050
Speaker 1: Many experts are worried that the rapid advances in making fakes,

636
00:44:34,090 --> 00:44:38,650
Speaker 1: combined with a catalyst of information cascades, will undermine democracy.

637
00:44:39,330 --> 00:44:44,250
Speaker 1: The biggest concerns have focused on elections Globally, we are

638
00:44:44,250 --> 00:44:50,610
Speaker 1: looking at highly polarized situations where this kind of manipulated

639
00:44:50,690 --> 00:44:52,770
Speaker 1: media can be used as a weapon. One of the

640
00:44:52,770 --> 00:44:56,690
Speaker 1: main reasons Francesca and Halsey made their Nixon deep fake

641
00:44:57,290 --> 00:45:01,050
Speaker 1: was to spread awareness about the risks of misinformation campaigns

642
00:45:01,530 --> 00:45:05,690
Speaker 1: before the twenty twenty US presidential election. Similarly, a group

643
00:45:05,770 --> 00:45:09,170
Speaker 1: showcased the power of deep fakes by making videos the

644
00:45:09,290 --> 00:45:12,690
Speaker 1: run up to the UK parliamentary election showing the two

645
00:45:12,730 --> 00:45:17,730
Speaker 1: bitter rivals, Boris Johnson and Jeremy Corman, each endorsing the other.

646
00:45:18,330 --> 00:45:21,330
Speaker 1: I wish to rise above this divide and indorse my

647
00:45:21,410 --> 00:45:25,530
Speaker 1: worthy opponent, the right Honorable Jeremy Corbyn. SIPI Prime Minister

648
00:45:25,770 --> 00:45:29,810
Speaker 1: of our United Kingdom, back Boris Johnson to continue as

649
00:45:29,810 --> 00:45:33,610
Speaker 1: our Prime Minister. But you know what, don't listen to me.

650
00:45:33,930 --> 00:45:36,210
Speaker 1: I think I may be one of the thousands of

651
00:45:36,290 --> 00:45:40,530
Speaker 1: deep fakes on the Internet, using powerful technologies to tell

652
00:45:40,650 --> 00:45:44,850
Speaker 1: stories that aren't so. This just kind of indicates how

653
00:45:45,250 --> 00:45:50,130
Speaker 1: candidates and political figures can be misrepresented, and you just

654
00:45:50,170 --> 00:45:54,770
Speaker 1: need to feed them into people's social media feeds for

655
00:45:54,810 --> 00:45:57,370
Speaker 1: them to be seeing this at times when the stakes

656
00:45:57,370 --> 00:46:01,570
Speaker 1: are pretty high. So far, we haven't yet seen sophisticated

657
00:46:01,650 --> 00:46:05,770
Speaker 1: deep fakes in US or UK politics. That might be

658
00:46:05,770 --> 00:46:08,770
Speaker 1: because fakes will be most effective if they're tim for

659
00:46:09,170 --> 00:46:13,130
Speaker 1: XM chaos, say close to election day, when newsrooms won't

660
00:46:13,130 --> 00:46:16,570
Speaker 1: have the time to investigate and debunk them. But another

661
00:46:16,610 --> 00:46:20,810
Speaker 1: reason might be the cheap fakes made with basic video

662
00:46:20,970 --> 00:46:25,210
Speaker 1: editing software are actually pretty effective. Remember the video that

663
00:46:25,330 --> 00:46:28,850
Speaker 1: surfaced of how speaker Nancy Pelosi, in which she appeared

664
00:46:28,890 --> 00:46:34,050
Speaker 1: intoxicated and confused. We want to give this president the

665
00:46:34,250 --> 00:46:42,010
Speaker 1: opportunity to something historic for our country. Both President Trump

666
00:46:42,050 --> 00:46:45,290
Speaker 1: and Rudy Giuliani shared the video as fact on Twitter.

667
00:46:45,930 --> 00:46:48,930
Speaker 1: The video is just a cheap fake, just slowed down

668
00:46:49,010 --> 00:46:53,810
Speaker 1: Pelosi's speech to make her seem incompetent. But maybe elections

669
00:46:54,450 --> 00:46:58,010
Speaker 1: won't be the biggest targets. Some people worry that deep

670
00:46:58,090 --> 00:47:03,690
Speaker 1: fakes could be weaponized to foment international conflict. Berkeley professor

671
00:47:03,730 --> 00:47:06,610
Speaker 1: Honey f Reed has been working with US government's Media

672
00:47:06,690 --> 00:47:11,330
Speaker 1: Forensics program to address this issue. DARPA, the Defense Department's

673
00:47:11,570 --> 00:47:13,730
Speaker 1: research arm, has been pouring a lot of money over

674
00:47:13,770 --> 00:47:16,970
Speaker 1: the last five years into this program. They are very

675
00:47:17,050 --> 00:47:21,290
Speaker 1: concerned about how this technology can be a threat to

676
00:47:21,370 --> 00:47:24,490
Speaker 1: national security and also how when we get images and

677
00:47:24,570 --> 00:47:26,770
Speaker 1: videos from around the world in areas of conflict, do

678
00:47:26,850 --> 00:47:28,850
Speaker 1: we know if they're real or not? Is this really

679
00:47:28,850 --> 00:47:32,290
Speaker 1: an image of a US soldier who has been taken hostage?

680
00:47:32,570 --> 00:47:34,770
Speaker 1: How do we know? So? What do you see as

681
00:47:34,850 --> 00:47:38,090
Speaker 1: some of the worst case scenarios. Here's the things that

682
00:47:38,170 --> 00:47:41,170
Speaker 1: keep me up at night. Right. A video of Donald

683
00:47:41,250 --> 00:47:45,010
Speaker 1: Trump saying I've launched nuclear weapons against Iran, and before

684
00:47:45,010 --> 00:47:46,970
Speaker 1: anybody gets around to figuring out whether this is real

685
00:47:47,050 --> 00:47:49,250
Speaker 1: or not, where we have global nuclear moutdown. And here's

686
00:47:49,250 --> 00:47:52,570
Speaker 1: the thing. I don't think that that's likely, but I

687
00:47:52,610 --> 00:47:55,730
Speaker 1: also don't think that the probability of that is zero.

688
00:47:56,330 --> 00:48:00,730
Speaker 1: And that should worry us because while it's not likely,

689
00:48:00,810 --> 00:48:06,010
Speaker 1: the consequences are spectacularly bad. Lawyer Danielle Citram worries about

690
00:48:06,010 --> 00:48:10,290
Speaker 1: an even more plausible scenario. And imagine a deep fake

691
00:48:10,690 --> 00:48:15,050
Speaker 1: of a well known American general burning a koran and

692
00:48:15,170 --> 00:48:19,410
Speaker 1: it is timed at a very tense moment in a

693
00:48:19,450 --> 00:48:25,330
Speaker 1: particular most you know country, whether it's Afghanistan. It could

694
00:48:25,370 --> 00:48:28,650
Speaker 1: then lead to physical violence. And you think this could

695
00:48:28,730 --> 00:48:32,370
Speaker 1: be made. No general, no qoran actually used in the video,

696
00:48:32,490 --> 00:48:37,890
Speaker 1: just programmed. You can use the technology to mine existing photographs.

697
00:48:37,930 --> 00:48:40,410
Speaker 1: Kind of easy, especially with someone like take Jim Mattis

698
00:48:40,810 --> 00:48:44,490
Speaker 1: when he was our defense secretary. Of Jim Mattis, you know,

699
00:48:44,770 --> 00:48:47,130
Speaker 1: actually taking a koran and ripping it in half and

700
00:48:47,130 --> 00:48:52,810
Speaker 1: say all Muslims should die. Imagine the chaos in diplomacy,

701
00:48:53,210 --> 00:48:58,050
Speaker 1: the chaos of our soldiers abroad in Muslim countries. It

702
00:48:58,130 --> 00:49:01,810
Speaker 1: would be inciting violence without question. Well, we haven't yet

703
00:49:01,810 --> 00:49:06,250
Speaker 1: seen spectacular fake videos used to disrupt elections or create

704
00:49:06,290 --> 00:49:12,970
Speaker 1: international chaos. We have seen recingly sophisticated attacks on public policymaking.

705
00:49:13,530 --> 00:49:16,690
Speaker 1: So we've got an example in twenty seventeen where the

706
00:49:16,810 --> 00:49:21,970
Speaker 1: FEC solicited public comment on the proposal to repeal net neutrality.

707
00:49:22,450 --> 00:49:26,570
Speaker 1: Net neutrality is the principle that internet service providers should

708
00:49:26,610 --> 00:49:31,370
Speaker 1: be a neutral public utility. They shouldn't discriminate between websites,

709
00:49:31,650 --> 00:49:35,530
Speaker 1: say slowing down Netflix streaming to encourage you to purchase

710
00:49:35,570 --> 00:49:40,410
Speaker 1: a different online video service. As President Barack Obama described

711
00:49:40,450 --> 00:49:44,210
Speaker 1: in twenty fourteen, there are no gatekeepers deciding which sites

712
00:49:44,250 --> 00:49:47,210
Speaker 1: you get to access. There are no toll roads on

713
00:49:47,250 --> 00:49:51,770
Speaker 1: the information super Highway. Federal Communications Policy had long supported

714
00:49:51,850 --> 00:49:56,730
Speaker 1: net neutrality, but in twenty seventeen, the Trump administration favored

715
00:49:56,810 --> 00:50:01,130
Speaker 1: repealing the policy. There were twenty two million comments that

716
00:50:01,290 --> 00:50:05,890
Speaker 1: the FEC received, but ninety six percent of those were

717
00:50:05,930 --> 00:50:10,930
Speaker 1: actually fake. The interesting thing is the real comments were

718
00:50:10,930 --> 00:50:15,330
Speaker 1: opposed to repeal, whereas the fake comments were in favor.

719
00:50:15,690 --> 00:50:19,810
Speaker 1: A Wall Street Journal investigation exposed that the fake public

720
00:50:19,850 --> 00:50:24,210
Speaker 1: comments were generated by bots. It found similar problems with

721
00:50:24,250 --> 00:50:28,250
Speaker 1: public comments about pay data lending. The bots varied their

722
00:50:28,290 --> 00:50:33,050
Speaker 1: comments in a combinatorial fashion so that the content wasn't identical.

723
00:50:33,690 --> 00:50:36,090
Speaker 1: With a little sleuthing, though, you could see that they

724
00:50:36,090 --> 00:50:40,570
Speaker 1: were generated by computers. But with the technology increasingly able

725
00:50:40,610 --> 00:50:45,210
Speaker 1: to generate completely original writing, like open aiyes program that

726
00:50:45,250 --> 00:50:48,570
Speaker 1: wrote the story about unicorns in the Andes, it's going

727
00:50:48,610 --> 00:50:51,930
Speaker 1: to become hard to spot the fakes. So there was

728
00:50:51,970 --> 00:50:55,850
Speaker 1: this Harvest student, Max Weiss, who used GPT two to

729
00:50:55,930 --> 00:50:58,330
Speaker 1: kind of demonstrate this, And I went on his site

730
00:50:58,410 --> 00:51:01,970
Speaker 1: yesterday and he's got this little test where you need

731
00:51:02,130 --> 00:51:07,130
Speaker 1: to decide whether a comment is real or fake. So

732
00:51:07,290 --> 00:51:09,330
Speaker 1: you go on and you read it and you decide

733
00:51:09,370 --> 00:51:11,570
Speaker 1: whether it's been written by a bot or by a human.

734
00:51:12,450 --> 00:51:15,570
Speaker 1: So I did this, and the ones that seemed to

735
00:51:15,610 --> 00:51:19,370
Speaker 1: be really well written and quite narrative discussive, generally I

736
00:51:19,410 --> 00:51:21,970
Speaker 1: was picking them as human. I was wrong almost all

737
00:51:21,970 --> 00:51:26,170
Speaker 1: the time. It was amazing and alarming. In our democracy,

738
00:51:26,250 --> 00:51:29,370
Speaker 1: public comments have been an important way in which citizens

739
00:51:29,410 --> 00:51:33,370
Speaker 1: can make their voices heard, but now it's becoming easy

740
00:51:33,490 --> 00:51:38,210
Speaker 1: to drown out those voices with millions of fake opinions. Now,

741
00:51:38,250 --> 00:51:41,610
Speaker 1: the downfall of truth likely won't come with a bang,

742
00:51:41,970 --> 00:51:46,810
Speaker 1: but a whimper, a slow, steady erosion that some call

743
00:51:47,410 --> 00:51:50,090
Speaker 1: truth decay. If you can't believe anything you read, or

744
00:51:50,130 --> 00:51:52,250
Speaker 1: hear or see anymore, I don't know how you have

745
00:51:52,290 --> 00:51:54,770
Speaker 1: a democracy a I don't know, frankly, how we have

746
00:51:54,850 --> 00:51:57,650
Speaker 1: civilized society if everybody's going to live in an echo

747
00:51:57,730 --> 00:52:00,850
Speaker 1: chamber believing their own version of events. How do we

748
00:52:00,890 --> 00:52:03,370
Speaker 1: have a dialogue if we can't agree on basic facts.

749
00:52:03,850 --> 00:52:07,490
Speaker 1: In the end, the most insidious impact of deep fakes

750
00:52:07,930 --> 00:52:11,170
Speaker 1: may not be the deep fake content itself, but the

751
00:52:11,170 --> 00:52:15,450
Speaker 1: ability to claim that real content is fake. It's something

752
00:52:15,450 --> 00:52:19,730
Speaker 1: that Danielle Citron refers to as the liar's dividend. The

753
00:52:19,850 --> 00:52:23,050
Speaker 1: liar's dividend is that the more you educate people about

754
00:52:23,090 --> 00:52:26,570
Speaker 1: the phenomenon of deep fags, the more the wrongdoer can

755
00:52:26,610 --> 00:52:31,210
Speaker 1: disclaim reality. Think about what President Trump did with the

756
00:52:31,290 --> 00:52:34,770
Speaker 1: Access Hollywood tape. You know, I'm automatically attracted to beautiful

757
00:52:34,810 --> 00:52:37,330
Speaker 1: I just started kissing them. It's like a magnet, kid.

758
00:52:38,370 --> 00:52:40,090
Speaker 1: I don't even know it. And when you're started, they

759
00:52:40,210 --> 00:52:42,530
Speaker 1: let you do it. You can do anything whatever you want.

760
00:52:42,610 --> 00:52:47,490
Speaker 1: Grab them by the pro I can do anything. Initially,

761
00:52:47,690 --> 00:52:51,730
Speaker 1: Trump apologized for the remarks. Anyone who knows me knows

762
00:52:51,810 --> 00:52:55,690
Speaker 1: these words don't reflect who I am. I said it,

763
00:52:56,130 --> 00:53:00,250
Speaker 1: I was wrong, and I apologize. But in twenty seventeen,

764
00:53:00,770 --> 00:53:04,330
Speaker 1: a year after his initial apology and with the idea

765
00:53:04,330 --> 00:53:08,570
Speaker 1: of deep fake content starting to gain attention, Trump changed

766
00:53:08,650 --> 00:53:11,770
Speaker 1: his tomb upon reflection, he said, they're not real. That

767
00:53:11,890 --> 00:53:14,330
Speaker 1: wasn't me. I don't think that was my voice. That's

768
00:53:14,370 --> 00:53:18,810
Speaker 1: the liar's dividend. In practice, the Trump comments about excess

769
00:53:18,810 --> 00:53:24,050
Speaker 1: Hollywood was remarkable. Slightly more subtle than that, he said,

770
00:53:24,490 --> 00:53:27,290
Speaker 1: I'm not sure that was me. Right, Well, that's the

771
00:53:27,370 --> 00:53:42,370
Speaker 1: corrosive gas lighting. Chapter eight, A Life Stored in the Cloud.

772
00:53:44,290 --> 00:53:48,650
Speaker 1: Deep fakes have the potential to devastate individuals and harms society.

773
00:53:49,250 --> 00:53:53,050
Speaker 1: The question is can we stop them from spreading before

774
00:53:53,090 --> 00:53:56,330
Speaker 1: they get out of control. To do so, we'd need

775
00:53:56,410 --> 00:54:00,570
Speaker 1: reliable ways to spot deep fakes. So the good news

776
00:54:00,650 --> 00:54:03,530
Speaker 1: is there are still artifacts in the synthesized content, whether

777
00:54:03,570 --> 00:54:06,170
Speaker 1: those are images, audio, or a video, that we as

778
00:54:06,210 --> 00:54:09,170
Speaker 1: the experts, can tell apart. So when, for example, The

779
00:54:09,170 --> 00:54:11,370
Speaker 1: New York Times wants to run a story with a video,

780
00:54:11,930 --> 00:54:14,610
Speaker 1: we can help them validate it. What are the real

781
00:54:14,730 --> 00:54:18,970
Speaker 1: sophisticated experts looking. Yeah, so the eyes are really wonderful

782
00:54:19,130 --> 00:54:22,850
Speaker 1: forensically because they reflect back to you what is in

783
00:54:22,850 --> 00:54:26,290
Speaker 1: the scene. So I'm sitting now right now in a studio,

784
00:54:26,450 --> 00:54:28,770
Speaker 1: there's maybe about a dozen or so lights around me,

785
00:54:28,810 --> 00:54:31,170
Speaker 1: and you can see this very complex set of reflections

786
00:54:31,170 --> 00:54:35,210
Speaker 1: in my eyes. So we can analyze fairly complex lighting patterns,

787
00:54:35,210 --> 00:54:38,170
Speaker 1: for example, to determine if this is one person's head

788
00:54:38,250 --> 00:54:41,170
Speaker 1: spliced onto another person's body, or if the two people

789
00:54:41,210 --> 00:54:45,530
Speaker 1: standing next to each other were digitally inserted from another photograph.

790
00:54:45,650 --> 00:54:48,250
Speaker 1: I could spend another hour telling you about the many

791
00:54:48,250 --> 00:54:52,130
Speaker 1: different forensic techniques that we've developed. There's no silver bullet here.

792
00:54:52,130 --> 00:54:55,050
Speaker 1: Really is a sort of a time consuming and deliberate

793
00:54:55,090 --> 00:54:58,450
Speaker 1: and thoughtful and it requires many many tools, and it

794
00:54:58,490 --> 00:55:00,530
Speaker 1: requires people with a fair amount of skill to do this.

795
00:55:01,090 --> 00:55:04,370
Speaker 1: Honey Freed also has quite a few detection techniques that

796
00:55:04,450 --> 00:55:07,050
Speaker 1: he won't speak about publicly for fear of the deep

797
00:55:07,050 --> 00:55:10,450
Speaker 1: fake creators will learn how beat his tests. I don't

798
00:55:10,490 --> 00:55:12,770
Speaker 1: create a GitHub repository and give my code to all

799
00:55:12,810 --> 00:55:16,050
Speaker 1: my adversaries. I don't have just one forensic techniques. I

800
00:55:16,090 --> 00:55:18,730
Speaker 1: have a couple dozen of them. So that means you,

801
00:55:18,970 --> 00:55:21,250
Speaker 1: as the person creating this now have to go back

802
00:55:21,250 --> 00:55:24,410
Speaker 1: and implement twenty different techniques and you have to do

803
00:55:24,450 --> 00:55:27,330
Speaker 1: it just perfectly, and that makes the landscape a little

804
00:55:27,330 --> 00:55:30,570
Speaker 1: bit more tricky for you to manage. As technology makes

805
00:55:30,570 --> 00:55:34,050
Speaker 1: it easier to create deep fakes, a big problem will

806
00:55:34,090 --> 00:55:37,450
Speaker 1: be the sheer amounts of content to review. So the

807
00:55:37,490 --> 00:55:41,330
Speaker 1: average person can download software repositories, and so it's getting

808
00:55:41,330 --> 00:55:44,370
Speaker 1: to the point now where the average person can just

809
00:55:44,490 --> 00:55:47,170
Speaker 1: run these as if they're running any standard piece of software.

810
00:55:47,210 --> 00:55:50,050
Speaker 1: There's also websites that have propped up where you can

811
00:55:50,090 --> 00:55:52,530
Speaker 1: pay them twenty bucks and you tell them, please put

812
00:55:52,530 --> 00:55:54,890
Speaker 1: this person's face into this person's video, and they will

813
00:55:54,930 --> 00:55:57,330
Speaker 1: do that for you. And so it doesn't take a

814
00:55:57,370 --> 00:55:59,850
Speaker 1: lot to get access to these tools. Now, I will

815
00:55:59,890 --> 00:56:02,210
Speaker 1: say that the output of those are not quite as

816
00:56:02,250 --> 00:56:04,690
Speaker 1: good as what we can create inside the lab. And

817
00:56:04,770 --> 00:56:06,490
Speaker 1: you just know what the trend is. You just know

818
00:56:06,570 --> 00:56:08,770
Speaker 1: it's going to get better and cheaper and faster and

819
00:56:08,810 --> 00:56:12,130
Speaker 1: easy to use. Detecting dow fakes will be a never

820
00:56:12,370 --> 00:56:17,210
Speaker 1: ending cat and mouse game. Remember how generative adversarial networks

821
00:56:17,290 --> 00:56:21,650
Speaker 1: or gams are built by training a fake generator to

822
00:56:21,730 --> 00:56:27,250
Speaker 1: outsmart a detector. Well. As detectors get better, fake generators

823
00:56:27,330 --> 00:56:31,850
Speaker 1: will be trained to keep pays still. Detectives like Honey

824
00:56:31,970 --> 00:56:36,090
Speaker 1: and platforms like Facebook are working to develop automated ways

825
00:56:36,130 --> 00:56:40,810
Speaker 1: to spot deep fakes rapidly and reliably. That's important because

826
00:56:40,850 --> 00:56:44,970
Speaker 1: more than five hundred additional hours of video are being

827
00:56:45,090 --> 00:56:48,690
Speaker 1: uploaded to YouTube every minute. I don't mean to sound

828
00:56:48,730 --> 00:56:51,730
Speaker 1: defeatist about this, but I'm going to lose this war.

829
00:56:51,970 --> 00:56:54,850
Speaker 1: I know this because it's always going to be easier

830
00:56:54,890 --> 00:56:57,130
Speaker 1: to create content than it is to detect it. But

831
00:56:58,130 --> 00:57:00,290
Speaker 1: here's where I will win. I will take it out

832
00:57:00,290 --> 00:57:03,050
Speaker 1: of the hands of the average person. So think about,

833
00:57:03,090 --> 00:57:06,690
Speaker 1: for example, the creation of counterfeit currency. With the latest

834
00:57:06,810 --> 00:57:10,090
Speaker 1: innovations brought on by the Treasure Department, it is hard

835
00:57:10,130 --> 00:57:12,290
Speaker 1: for the average person to take their inkjet printer and

836
00:57:12,370 --> 00:57:15,730
Speaker 1: create compelling fake currency. And I think that's going to

837
00:57:15,770 --> 00:57:18,130
Speaker 1: be the same trend here is that if you're using

838
00:57:18,130 --> 00:57:20,330
Speaker 1: some off the shelf tool, if you're paying somebody on

839
00:57:20,370 --> 00:57:22,090
Speaker 1: the website, we're going to find you, and we're going

840
00:57:22,130 --> 00:57:24,490
Speaker 1: to find you quickly. But if you are a dedicated,

841
00:57:24,690 --> 00:57:27,410
Speaker 1: highly skilled of the time and the effort to create it,

842
00:57:27,690 --> 00:57:29,530
Speaker 1: we are going to have to work really hard to

843
00:57:29,570 --> 00:57:34,210
Speaker 1: detect those. Given the challenges of detecting fake content, some

844
00:57:34,290 --> 00:57:38,130
Speaker 1: people envision a different kind of techno fix. They propose

845
00:57:38,250 --> 00:57:42,930
Speaker 1: developing airtight ways for content creators to mark their own

846
00:57:42,930 --> 00:57:48,370
Speaker 1: original video as real. That way, we get instantly recognize

847
00:57:48,410 --> 00:57:51,850
Speaker 1: an altered version if it wasn't identical. Now there's ways

848
00:57:51,850 --> 00:57:54,410
Speaker 1: of authenticating at the point of recording, and these are

849
00:57:54,450 --> 00:57:58,010
Speaker 1: what it's called control capture system. So here's the idea.

850
00:57:58,250 --> 00:58:01,330
Speaker 1: You use a special app on your mobile device that,

851
00:58:01,410 --> 00:58:05,250
Speaker 1: at the point of capture, a cryptographically signs the image

852
00:58:05,250 --> 00:58:07,970
Speaker 1: of the video or the audio. It puts that signature

853
00:58:07,970 --> 00:58:09,930
Speaker 1: onto the blockchain. The only thing you have to know

854
00:58:09,930 --> 00:58:13,370
Speaker 1: about the blockchain is that it is an immutable distributed ledger,

855
00:58:13,450 --> 00:58:16,930
Speaker 1: which means that that signature is essentially impossible to manipulate.

856
00:58:17,490 --> 00:58:20,850
Speaker 1: And now all of that happened at the point of recording.

857
00:58:21,330 --> 00:58:23,370
Speaker 1: If I was running a campaign today and I was

858
00:58:23,450 --> 00:58:27,610
Speaker 1: worried about my candidates likeness being misused, absolutely every public

859
00:58:27,610 --> 00:58:29,730
Speaker 1: event that they were at, I would record with a

860
00:58:29,770 --> 00:58:31,770
Speaker 1: control capture system and I'd be able to prove what

861
00:58:31,810 --> 00:58:35,250
Speaker 1: they actually said or did at any point in the future.

862
00:58:35,690 --> 00:58:39,370
Speaker 1: So this approach would shift the burden of authentication so

863
00:58:39,410 --> 00:58:43,650
Speaker 1: the people creating the videos rather than publishers or consumers.

864
00:58:44,530 --> 00:58:48,170
Speaker 1: Law professor Danielle Citron has explored how this solution could

865
00:58:48,290 --> 00:58:52,410
Speaker 1: quickly become dystopium. We might see the emergence of an

866
00:58:52,530 --> 00:58:54,810
Speaker 1: essentially an audit trail of everything you do and say

867
00:58:54,850 --> 00:58:57,810
Speaker 1: all of the time. Danielle refers to the business model

868
00:58:57,930 --> 00:59:02,010
Speaker 1: as immutable lifelogs in the cloud. In a way we

869
00:59:02,090 --> 00:59:05,410
Speaker 1: sort of already seen it. There are health plans that

870
00:59:05,490 --> 00:59:07,570
Speaker 1: if you wear a fitbit all the time and you

871
00:59:07,650 --> 00:59:10,290
Speaker 1: let yourself be monitored as your insurance, you know your

872
00:59:10,330 --> 00:59:13,370
Speaker 1: health insurance rates. But you can see how if the

873
00:59:13,450 --> 00:59:17,530
Speaker 1: incentives are there in the market to self surveil, whether

874
00:59:17,570 --> 00:59:22,050
Speaker 1: it's for health insurance, life insurance, car insurance, we're going

875
00:59:22,090 --> 00:59:25,650
Speaker 1: to see the unraveling of bribe to say by ourselves.

876
00:59:26,050 --> 00:59:30,850
Speaker 1: You know, corporations may very well, because the CEO is

877
00:59:30,930 --> 00:59:35,050
Speaker 1: so valuable, they may say you've got to have a log,

878
00:59:35,370 --> 00:59:37,530
Speaker 1: an immutable audit trail of everything you do and say.

879
00:59:37,570 --> 00:59:39,770
Speaker 1: So when that deep fake comes up the night before

880
00:59:39,770 --> 00:59:43,970
Speaker 1: the IPO, you can say, look, the CEO wasn't taking

881
00:59:44,010 --> 00:59:47,290
Speaker 1: the bribe wasn't having sex with a prostitute, and so

882
00:59:47,330 --> 00:59:50,930
Speaker 1: we have proof because we have an auto trail, we

883
00:59:50,930 --> 00:59:53,690
Speaker 1: have a log. So when we were imagining, we were

884
00:59:53,690 --> 00:59:58,410
Speaker 1: imagining a business model that hasn't quite come up, but

885
00:59:58,450 --> 01:00:02,570
Speaker 1: we have gotten a number of requests from insurance companies

886
01:00:03,010 --> 01:00:06,730
Speaker 1: as well as companies to say we're interested in this idea.

887
01:00:06,850 --> 01:00:08,490
Speaker 1: So how much has to be in that log? Does

888
01:00:08,490 --> 01:00:10,610
Speaker 1: this have to be a whole video of your life?

889
01:00:10,650 --> 01:00:13,570
Speaker 1: That is a great question, one that terrifies us. So

890
01:00:13,610 --> 01:00:19,130
Speaker 1: it may be that you're logging locate geolocation, you're logging videos,

891
01:00:19,170 --> 01:00:22,050
Speaker 1: you see people talking and who they're interacting with, and

892
01:00:22,130 --> 01:00:25,170
Speaker 1: that might be good enough to prevent the mischief that

893
01:00:25,450 --> 01:00:30,970
Speaker 1: would hijack the IP. Your whole life online, yes, stored securely,

894
01:00:31,250 --> 01:00:35,170
Speaker 1: our clock down, protected in the cloud. It is, at

895
01:00:35,250 --> 01:00:38,290
Speaker 1: least for a privacy scholar. There are so many reasons

896
01:00:38,290 --> 01:00:42,210
Speaker 1: why we ought to have privacy that aren't about hiding things.

897
01:00:42,850 --> 01:00:47,530
Speaker 1: It's about creating spaces and managing boundaries around ourselves and

898
01:00:47,650 --> 01:00:51,650
Speaker 1: our intimates and our loved ones. So I worry that

899
01:00:51,690 --> 01:00:55,570
Speaker 1: if we entirely unravel privacy A in the wrong hands

900
01:00:56,050 --> 01:01:00,970
Speaker 1: is very dangerous. Right B It changes how we think

901
01:01:00,970 --> 01:01:10,770
Speaker 1: about ourselves and humanity? Chapter nine, Section two thirty. So

902
01:01:10,890 --> 01:01:15,730
Speaker 1: technofixes are complicated. What about passing laws to ban deep

903
01:01:15,810 --> 01:01:19,090
Speaker 1: fakes or at least deep fakes that don't disclose their

904
01:01:19,090 --> 01:01:22,850
Speaker 1: fake So the video and audio is speech in our

905
01:01:22,850 --> 01:01:26,730
Speaker 1: First Amendment doctrine is very much a protective of free speech,

906
01:01:27,010 --> 01:01:30,610
Speaker 1: and the Supreme Court has explained that lies just lies

907
01:01:30,650 --> 01:01:34,930
Speaker 1: themselves without harm is protected speech. When lies cause certain

908
01:01:34,970 --> 01:01:42,250
Speaker 1: kinds of harm, we can regulate it. Defamation of private people, threats, incitement, fraud,

909
01:01:42,890 --> 01:01:47,410
Speaker 1: impersonation of government officials. What about lies concerning public figures

910
01:01:47,450 --> 01:01:53,010
Speaker 1: like politicians? California and Texas, for instance, recently pass laws

911
01:01:53,090 --> 01:01:56,490
Speaker 1: making it illegal to publish deep fakes of a candidate

912
01:01:56,770 --> 01:01:59,370
Speaker 1: in the weeks leading up to an election. It's not

913
01:01:59,490 --> 01:02:03,650
Speaker 1: clear yet whether the laws will pass constitutional muster. As

914
01:02:03,690 --> 01:02:07,810
Speaker 1: you're saying in an American content, we are just not

915
01:02:07,930 --> 01:02:10,730
Speaker 1: going to be able to law great fakes. Yeah, we

916
01:02:10,770 --> 01:02:12,810
Speaker 1: can't have a flat van, and I don't think we should.

917
01:02:12,970 --> 01:02:16,970
Speaker 1: It would fail on doctrinal grounds, but ultimately it would

918
01:02:17,010 --> 01:02:23,850
Speaker 1: prevent the positive uses. Interestingly, in January twenty twenty, China,

919
01:02:24,050 --> 01:02:28,970
Speaker 1: which has no First Amendment protecting free speech promulgated regulations

920
01:02:29,370 --> 01:02:33,450
Speaker 1: banning deep fakes. The use of AI or virtuality now

921
01:02:33,570 --> 01:02:36,650
Speaker 1: needs to be clearly marked in a prominent manner, and

922
01:02:36,770 --> 01:02:40,090
Speaker 1: the failure to do so is considered a criminal offense.

923
01:02:40,930 --> 01:02:43,970
Speaker 1: To explore other options for the US, I went to

924
01:02:44,010 --> 01:02:47,650
Speaker 1: speak with a public policy expert. My name is Joan Donovan,

925
01:02:47,770 --> 01:02:51,370
Speaker 1: and I work at Harvard Kennedy Shorenstein Center, where I

926
01:02:51,490 --> 01:02:54,730
Speaker 1: lead a team of researchers looking at medium manipulation and

927
01:02:54,770 --> 01:02:58,530
Speaker 1: disinformation campaigns. Joan is head of the Technology and Social

928
01:02:58,610 --> 01:03:03,210
Speaker 1: Change Research Project, and her staff studies how social media

929
01:03:03,530 --> 01:03:07,610
Speaker 1: gives rise to hoaxes and scams. Her team is particularly

930
01:03:07,610 --> 01:03:13,450
Speaker 1: interested and precisely how misinformation spreads across the Internet. Ultimately,

931
01:03:13,570 --> 01:03:16,730
Speaker 1: underneath all of this is the distribution mechanism, which is

932
01:03:16,970 --> 01:03:23,690
Speaker 1: social media and platforms. And platforms have to rethink the

933
01:03:23,730 --> 01:03:27,530
Speaker 1: openness of their design because that has now become a

934
01:03:27,650 --> 01:03:32,450
Speaker 1: territory for information warfare. In early twenty twenty, Facebook announced

935
01:03:32,490 --> 01:03:38,330
Speaker 1: a major policy change about synthesized content. Facebook preissued policies

936
01:03:38,450 --> 01:03:41,530
Speaker 1: now on deep fake saying that if it is an

937
01:03:41,530 --> 01:03:46,650
Speaker 1: AI generated video and it's misleading in some other contextual way,

938
01:03:47,410 --> 01:03:52,970
Speaker 1: then they will remove it. Interestingly, Facebook ban the Moon

939
01:03:53,050 --> 01:03:56,410
Speaker 1: Disaster Team's Nixon video even though it was made for

940
01:03:56,530 --> 01:04:00,810
Speaker 1: educational purposes, but didn't remove the slowed down version of

941
01:04:00,890 --> 01:04:05,690
Speaker 1: Nancy Pelosi, which was made to mislead the public. Why

942
01:04:05,810 --> 01:04:10,730
Speaker 1: because the Pelosi video wasn't created with artificial elligience. For now,

943
01:04:11,170 --> 01:04:14,970
Speaker 1: Facebook is choosing to target deep fakes, but not cheap fakes.

944
01:04:15,410 --> 01:04:18,170
Speaker 1: One way to push platforms to take a stronger stance

945
01:04:18,290 --> 01:04:21,170
Speaker 1: might be to remove some of the legal protections that

946
01:04:21,250 --> 01:04:25,570
Speaker 1: they currently enjoy. Under Section two thirty of the Communication's

947
01:04:25,610 --> 01:04:30,490
Speaker 1: Decency Act past in nineteen ninety six, platforms aren't legally

948
01:04:30,610 --> 01:04:34,890
Speaker 1: liable for content posted by its users. The fact that

949
01:04:34,970 --> 01:04:38,890
Speaker 1: platforms have no responsibility for the content they host has

950
01:04:38,930 --> 01:04:42,530
Speaker 1: an upside. It's led to the massive diversity of online

951
01:04:42,610 --> 01:04:47,050
Speaker 1: content we enjoyed today, but it also allows a dangerous

952
01:04:47,210 --> 01:04:51,330
Speaker 1: escalation of fake news. Is it time to change section

953
01:04:51,370 --> 01:04:56,290
Speaker 1: two thirty to create incentives for platforms to police false content?

954
01:04:57,090 --> 01:05:00,250
Speaker 1: I ask the former head of a major platform, LinkedIn

955
01:05:00,410 --> 01:05:04,170
Speaker 1: co founder Reid Hoffman. For example, let's take my view

956
01:05:04,250 --> 01:05:07,570
Speaker 1: of what the response to the Christchurch shooting should be

957
01:05:07,730 --> 01:05:10,970
Speaker 1: as to say, well, we want you to solve, not

958
01:05:11,130 --> 01:05:16,970
Speaker 1: having terrorism, murderer or murderers displayed to people. So we're

959
01:05:17,010 --> 01:05:19,570
Speaker 1: simply going to do a fine of ten thousand dollars

960
01:05:19,650 --> 01:05:23,850
Speaker 1: per view. Two shootings occurred at mosques in Christchurch, New

961
01:05:23,890 --> 01:05:28,050
Speaker 1: Zealand in March twenty nineteen. Graphic videos of the event

962
01:05:28,490 --> 01:05:32,210
Speaker 1: were soon posted online. Five people saw it, that's fifty

963
01:05:32,250 --> 01:05:35,050
Speaker 1: thousand dollars. But if he becomes a meme and a

964
01:05:35,130 --> 01:05:39,730
Speaker 1: million people see it, that's ten billion dollars. Yes, right,

965
01:05:39,890 --> 01:05:42,490
Speaker 1: So what's really trying to do is get you to say,

966
01:05:43,010 --> 01:05:45,930
Speaker 1: let's make sure that the meme never happens. Okay, So

967
01:05:45,970 --> 01:05:50,730
Speaker 1: that's a governance mechanism there. Yes, you find the channel

968
01:05:50,770 --> 01:05:54,130
Speaker 1: the platform based on number of views would be a

969
01:05:54,410 --> 01:05:57,290
Speaker 1: very general way to say. Now you guys have to solve.

970
01:05:57,450 --> 01:06:01,330
Speaker 1: Now you solve, you figure it out. What about other solutions?

971
01:06:01,610 --> 01:06:05,050
Speaker 1: If we are to make regulation, it should be about

972
01:06:05,090 --> 01:06:09,450
Speaker 1: the amount of staff in proportion to the out of

973
01:06:09,530 --> 01:06:13,050
Speaker 1: users so that they can get a handle on the content.

974
01:06:13,530 --> 01:06:17,010
Speaker 1: But can they be fast enough. Maybe the viral spread

975
01:06:17,050 --> 01:06:21,010
Speaker 1: should be slowed down enough to allow them to moderate.

976
01:06:21,090 --> 01:06:25,210
Speaker 1: Let's put it this way. The stock market has certain

977
01:06:25,850 --> 01:06:29,770
Speaker 1: governors built in when there's massive changes in a stock price.

978
01:06:30,250 --> 01:06:33,370
Speaker 1: There are decelerators that kick in, breaks that kick in

979
01:06:33,770 --> 01:06:37,010
Speaker 1: should the platforms have breaks that kick in before something

980
01:06:37,010 --> 01:06:41,970
Speaker 1: can go fully viral. So in terms of deceleration, there

981
01:06:41,970 --> 01:06:44,930
Speaker 1: are things that they do already that accelerate the process

982
01:06:45,010 --> 01:06:48,210
Speaker 1: that they need to think differently about, especially when it

983
01:06:48,250 --> 01:06:52,810
Speaker 1: comes to something turning into a trending topic. So there

984
01:06:52,850 --> 01:06:56,930
Speaker 1: needs to be an intervening moment before things get to

985
01:06:56,970 --> 01:07:00,130
Speaker 1: the homepage and get to trending, where there is a

986
01:07:00,170 --> 01:07:04,250
Speaker 1: content review. So much to say here, but I want

987
01:07:04,290 --> 01:07:08,170
Speaker 1: to think particularly about listeners who are in their twenties

988
01:07:08,170 --> 01:07:12,010
Speaker 1: and thirties, are very tech savvy. They're going to be

989
01:07:12,090 --> 01:07:15,530
Speaker 1: part of the solution here. What would you say to

990
01:07:15,650 --> 01:07:21,090
Speaker 1: them about what they can do? I think it's important

991
01:07:21,650 --> 01:07:26,690
Speaker 1: that younger people advocate for the Internet that they want.

992
01:07:27,010 --> 01:07:29,050
Speaker 1: We have to fight for it, We have to ask

993
01:07:29,130 --> 01:07:33,850
Speaker 1: for different things, and that kind of agitation can come

994
01:07:33,890 --> 01:07:38,570
Speaker 1: in the form of posting on the platform, writing letters,

995
01:07:39,370 --> 01:07:42,770
Speaker 1: joining groups like Fight for the Future, and trying to

996
01:07:44,170 --> 01:07:48,570
Speaker 1: work on getting platforms to do better and to advocate

997
01:07:48,570 --> 01:07:51,210
Speaker 1: for the kind of content that you want to see

998
01:07:51,570 --> 01:07:56,770
Speaker 1: more of. The important thing is that our society is

999
01:07:56,810 --> 01:08:00,650
Speaker 1: shaped by these platforms and so we're not going to

1000
01:08:00,770 --> 01:08:03,730
Speaker 1: do away with them, but we don't have to make

1001
01:08:03,890 --> 01:08:17,050
Speaker 1: do with them either. Conclusion, choose your planet. So there

1002
01:08:17,090 --> 01:08:20,650
Speaker 1: you have it. Stewards of the Brave New Planet. Synthetic

1003
01:08:20,730 --> 01:08:25,850
Speaker 1: media or deep fakes. People have been manipulating content for

1004
01:08:26,010 --> 01:08:29,770
Speaker 1: more than a hundred years, but recent advances in AI

1005
01:08:29,890 --> 01:08:33,250
Speaker 1: have taken it to a whole new level of verisimilitude.

1006
01:08:33,810 --> 01:08:38,530
Speaker 1: The technology could transform movies and television, favored actors from

1007
01:08:38,610 --> 01:08:42,290
Speaker 1: years past starring in new narratives, along with actors who

1008
01:08:42,330 --> 01:08:46,490
Speaker 1: never existed, patients regaining the ability to speak in their

1009
01:08:46,490 --> 01:08:52,290
Speaker 1: own voices, personalized stories created on demand for any child

1010
01:08:52,330 --> 01:08:56,370
Speaker 1: around the globe, matching their interests, written in their dialect,

1011
01:08:56,690 --> 01:09:02,170
Speaker 1: representing their communities. But there's also great potential for harm,

1012
01:09:02,330 --> 01:09:07,210
Speaker 1: the ability to cast anyone in a pornographic video, weaponized

1013
01:09:07,290 --> 01:09:11,810
Speaker 1: media dropping days before an election, or provoking international conflicts.

1014
01:09:12,650 --> 01:09:15,610
Speaker 1: Are we going to be able to tell fact from fiction?

1015
01:09:16,170 --> 01:09:21,890
Speaker 1: Will truth survive? And what does it mean for our democracy? Better?

1016
01:09:21,930 --> 01:09:25,010
Speaker 1: Fake detection may help, but it'll be hard for it

1017
01:09:25,050 --> 01:09:28,570
Speaker 1: to keep up, and logging our lives in blockchain to

1018
01:09:28,650 --> 01:09:34,970
Speaker 1: protect against misrepresentation doesn't sound like an attractive idea. Outright

1019
01:09:35,090 --> 01:09:37,850
Speaker 1: bands on deep fakes are being tried in some countries,

1020
01:09:38,210 --> 01:09:41,530
Speaker 1: but they're tricky in the US given our constitutional protections

1021
01:09:41,570 --> 01:09:45,450
Speaker 1: for free speech. Maybe the best solution is to put

1022
01:09:45,490 --> 01:09:50,090
Speaker 1: the liability on platforms like Facebook and YouTube. If we

1023
01:09:50,210 --> 01:09:54,090
Speaker 1: can joan Donovan's right to get the future you want,

1024
01:09:54,490 --> 01:09:57,010
Speaker 1: you're going to have to fight for it. You don't

1025
01:09:57,050 --> 01:09:59,410
Speaker 1: have to be an expert, and you don't have to

1026
01:09:59,450 --> 01:10:02,850
Speaker 1: do it alone. When enough people get engaged, we make

1027
01:10:02,890 --> 01:10:07,170
Speaker 1: wise choices. Deep fakes are a problem that everyone can

1028
01:10:07,210 --> 01:10:11,050
Speaker 1: engage with. Brainstorm with your friends about what should be done.

1029
01:10:11,490 --> 01:10:15,330
Speaker 1: Use social media. Tweet at your elected representatives to ask

1030
01:10:15,370 --> 01:10:18,810
Speaker 1: if they're working on laws, like in California and Texas.

1031
01:10:19,490 --> 01:10:23,210
Speaker 1: And if you work for a tech company, ask yourself

1032
01:10:23,490 --> 01:10:27,250
Speaker 1: and your colleagues if you're doing enough. You can find

1033
01:10:27,410 --> 01:10:31,090
Speaker 1: lots of resources and ideas at our website Brave New

1034
01:10:31,130 --> 01:10:35,850
Speaker 1: Planet dot org. It's time to choose our planet. The

1035
01:10:35,970 --> 01:10:49,890
Speaker 1: future is up to us. Brave New Planet is a

1036
01:10:49,930 --> 01:10:53,050
Speaker 1: coproduction of the Broad Institute of MT and Harvard Pushkin

1037
01:10:53,130 --> 01:10:56,530
Speaker 1: Industries in the Boston Globe, with support from the Alfred P.

1038
01:10:56,690 --> 01:11:00,490
Speaker 1: Sloan Foundation. Our show is produced by Rebecca Lee Douglas

1039
01:11:00,610 --> 01:11:05,290
Speaker 1: with Mary Doo theme song composed by Ned Porter, mastering

1040
01:11:05,330 --> 01:11:09,210
Speaker 1: and sound designed by James Garver, fact checking by as

1041
01:11:09,210 --> 01:11:13,290
Speaker 1: If Fridman and a Stitt and Enchant. Special thanks to

1042
01:11:13,370 --> 01:11:17,970
Speaker 1: Christine Heenan and Rachel Roberts at Clarendon Communications, to Lee McGuire,

1043
01:11:18,170 --> 01:11:21,530
Speaker 1: Kristen Zarelli and Justine Levin Allerhans at the Broad, to

1044
01:11:21,730 --> 01:11:25,930
Speaker 1: mil Lobell and Heather Faine at Pushkin, and to Eliah

1045
01:11:26,090 --> 01:11:30,170
Speaker 1: Edie Brode who made the Broad Institute possible. This is

1046
01:11:30,210 --> 01:11:32,650
Speaker 1: brave new planet. I'm Eric Lander.