1
00:00:01,600 --> 00:00:04,080
Speaker 1: We've talked quite a lot on this show lately about

2
00:00:04,200 --> 00:00:08,520
Speaker 1: artificial intelligence because of how quickly it's inserted itself into

3
00:00:08,560 --> 00:00:12,080
Speaker 1: so many parts of our lives, and how it's becoming

4
00:00:12,160 --> 00:00:16,759
Speaker 1: more and more difficult to tell what's real and what's fabricated.

5
00:00:17,800 --> 00:00:22,360
Speaker 1: One place where that's increasingly becoming a problem politics, especially

6
00:00:22,400 --> 00:00:26,360
Speaker 1: now as the twenty twenty four US presidential campaign heats up.

7
00:00:26,720 --> 00:00:32,280
Speaker 2: EI artificial intelligence is now hitting the US presidential campaign trail.

8
00:00:33,080 --> 00:00:36,760
Speaker 2: Donald Trump has been got in a series of aidep

9
00:00:36,920 --> 00:00:41,520
Speaker 2: figs this time. Ron DeSantis is also in the spotlight.

10
00:00:41,600 --> 00:00:45,040
Speaker 2: Reports say Desamptus published fake images of Donald Trump and

11
00:00:45,120 --> 00:00:46,040
Speaker 2: Anthony Fauci.

12
00:00:46,440 --> 00:00:51,880
Speaker 1: Already, candidates and their surrogates and even bad actors overseas

13
00:00:51,960 --> 00:00:54,280
Speaker 1: are reaching for every tool they can to try to

14
00:00:54,320 --> 00:00:58,640
Speaker 1: shape voters' perceptions and ultimately influence the outcome.

15
00:00:59,000 --> 00:01:02,160
Speaker 3: There is steph an arms race when it comes to AI,

16
00:01:02,480 --> 00:01:05,679
Speaker 3: but there is an element of mutually assured destruction.

17
00:01:05,880 --> 00:01:08,560
Speaker 1: And the government is struggling to figure out how to

18
00:01:08,640 --> 00:01:11,240
Speaker 1: police the use of AI in campaigns.

19
00:01:11,680 --> 00:01:14,560
Speaker 4: What makes the FEC really interesting in this is that

20
00:01:14,600 --> 00:01:17,640
Speaker 4: they are by designing a partisan agency with a mix

21
00:01:17,680 --> 00:01:20,480
Speaker 4: of Republicans and Democrats and so across the board, you

22
00:01:20,520 --> 00:01:29,319
Speaker 4: really see them deadlocking on almost every issue.

23
00:01:30,720 --> 00:01:33,720
Speaker 1: I'm west Caasova today on The Big Take, Bloomberg's Emily

24
00:01:33,760 --> 00:01:38,039
Speaker 1: Burnbaum and Laura Davison tell us how AI is transforming

25
00:01:38,400 --> 00:01:46,640
Speaker 1: and distorting our political reality. Emily, it seems like it

26
00:01:46,680 --> 00:01:51,240
Speaker 1: took about five minutes after AI really came forward as

27
00:01:51,240 --> 00:01:54,640
Speaker 1: a thing that was easily accessible for the political world

28
00:01:54,720 --> 00:01:57,600
Speaker 1: to say, hey, we can use this for all kinds

29
00:01:57,640 --> 00:02:00,920
Speaker 1: of reasons, some good and some real not so good.

30
00:02:02,320 --> 00:02:05,720
Speaker 3: Yeah, we're still really in the early stages of this.

31
00:02:05,880 --> 00:02:09,720
Speaker 3: We've yet to see the most dramatic examples of political

32
00:02:09,760 --> 00:02:13,480
Speaker 3: misinformation thanks to generative AI, but there are a couple

33
00:02:13,600 --> 00:02:17,880
Speaker 3: really big and well known examples that have been circulating.

34
00:02:18,120 --> 00:02:22,160
Speaker 3: So some far right politicians in Germany used AI to

35
00:02:22,440 --> 00:02:26,040
Speaker 3: generate images of immigrants that look very angry and violent.

36
00:02:26,720 --> 00:02:30,480
Speaker 3: No one was real, just you know, meant to incite

37
00:02:30,760 --> 00:02:34,839
Speaker 3: sort of nationalist fervor. Here back in the US, we're

38
00:02:34,880 --> 00:02:39,000
Speaker 3: seeing a lot more of that. So presidential candidate Ron

39
00:02:39,080 --> 00:02:44,280
Speaker 3: DeSantis his campaign arm put out a video that included

40
00:02:44,400 --> 00:02:49,680
Speaker 3: images of Anthony Fauci hugging former President Donald Trump, and

41
00:02:49,960 --> 00:02:53,960
Speaker 3: they were not real. So Desantas's campaign didn't say that

42
00:02:54,080 --> 00:02:58,600
Speaker 3: they were AI generated, but experts kind of circled and

43
00:02:58,720 --> 00:03:02,680
Speaker 3: identified them as fake, and then within a couple hours

44
00:03:02,720 --> 00:03:05,120
Speaker 3: there was a Twitter note appended to it saying these

45
00:03:05,160 --> 00:03:08,400
Speaker 3: are not real. And the other big example is the

46
00:03:08,440 --> 00:03:12,760
Speaker 3: Republican National Committee put together an AI generated video that

47
00:03:12,880 --> 00:03:16,040
Speaker 3: purported to show, you know, here's the future that Biden wants,

48
00:03:16,080 --> 00:03:19,919
Speaker 3: and it was very dystopian and disturbing. They actually said

49
00:03:19,960 --> 00:03:22,720
Speaker 3: it was AI generated.

50
00:03:23,520 --> 00:03:25,600
Speaker 4: Yeah, this video was really quite disturbing. You know. It

51
00:03:25,680 --> 00:03:29,040
Speaker 4: showed images of China attacking Taiwan. It showed the streets

52
00:03:29,080 --> 00:03:31,760
Speaker 4: of San Francisco with troops marching in as martial law

53
00:03:31,840 --> 00:03:34,680
Speaker 4: was taking over, and those looked very very real. The

54
00:03:34,680 --> 00:03:37,080
Speaker 4: one tailtale sign at the end of this video that

55
00:03:37,120 --> 00:03:40,040
Speaker 4: it was fake was it showed Joe Biden just slumped

56
00:03:40,080 --> 00:03:42,960
Speaker 4: over his desk, looking really dejected, and you could see

57
00:03:42,960 --> 00:03:45,040
Speaker 4: that his elbows weren't quite connecting with the edge of

58
00:03:45,040 --> 00:03:47,680
Speaker 4: the desk. So it also had a little disclaimer, very

59
00:03:47,760 --> 00:03:50,360
Speaker 4: very tiny in the corner saying this was AI generated.

60
00:03:50,400 --> 00:03:52,200
Speaker 4: But one of the things that consumers are gonna have

61
00:03:52,280 --> 00:03:54,920
Speaker 4: to start doing is be really good about determining what

62
00:03:55,080 --> 00:03:57,760
Speaker 4: video is fake and what is real? And there's always

63
00:03:57,800 --> 00:04:00,400
Speaker 4: with AI, the technology is pretty good, but you know,

64
00:04:00,440 --> 00:04:03,800
Speaker 4: sometimes people have six fingers on the Fauci and Trump

65
00:04:03,920 --> 00:04:07,320
Speaker 4: hugging images. You know, it looked very airbrushed. Trump's hair

66
00:04:07,360 --> 00:04:10,000
Speaker 4: looked very weird in certain angles. So that's kind of

67
00:04:10,160 --> 00:04:12,520
Speaker 4: a telltale sign if you think something looks weird and

68
00:04:12,520 --> 00:04:13,960
Speaker 4: you're like, look at the hands, look at the feet,

69
00:04:13,960 --> 00:04:15,960
Speaker 4: look at the limbs, look at some of these features.

70
00:04:15,960 --> 00:04:18,520
Speaker 4: They look a little bit distorted, and that's usually a

71
00:04:18,560 --> 00:04:20,080
Speaker 4: sign that something is a little off.

72
00:04:20,680 --> 00:04:23,240
Speaker 1: I remember in the last election we started thinking about

73
00:04:23,320 --> 00:04:27,679
Speaker 1: deep face, these videos, these pictures which are indistinguishable from

74
00:04:27,920 --> 00:04:31,360
Speaker 1: what's real, and the technology wasn't quite there yet, And now,

75
00:04:31,400 --> 00:04:33,799
Speaker 1: of course we see every day that it is getting

76
00:04:33,880 --> 00:04:36,080
Speaker 1: much better. And I guess that raises this question that

77
00:04:36,200 --> 00:04:39,479
Speaker 1: as the campaign heats up, we're going to see so

78
00:04:39,720 --> 00:04:41,839
Speaker 1: many of these images that it's going to be really

79
00:04:41,839 --> 00:04:45,279
Speaker 1: hard to tell what's real and what's not correct.

80
00:04:45,320 --> 00:04:49,720
Speaker 4: And there's no regulation that requires any person, any advertiser,

81
00:04:49,800 --> 00:04:53,839
Speaker 4: any candidate to declare something as fake. Google has said

82
00:04:54,120 --> 00:04:55,920
Speaker 4: that in their tools that they are going to have

83
00:04:56,080 --> 00:04:58,320
Speaker 4: not a physical marker on it, but sort of a

84
00:04:58,360 --> 00:05:01,200
Speaker 4: digital marker, So if people go and reverse image search

85
00:05:01,240 --> 00:05:03,040
Speaker 4: it, it will flag it as something that's fake. But that's

86
00:05:03,240 --> 00:05:05,800
Speaker 4: how often do you go and actually reverse image search something.

87
00:05:06,360 --> 00:05:08,599
Speaker 4: Not that often, So it's going to really take a

88
00:05:08,600 --> 00:05:11,920
Speaker 4: lot of discerning, both from kind of the truth squatting

89
00:05:12,040 --> 00:05:14,760
Speaker 4: faction on Twitter as well as from individual consumers to

90
00:05:14,800 --> 00:05:16,560
Speaker 4: know if something's real or fake. And as we know,

91
00:05:16,640 --> 00:05:18,040
Speaker 4: a lot of people don't even know who's running for

92
00:05:18,080 --> 00:05:18,839
Speaker 4: president right now.

93
00:05:19,320 --> 00:05:23,839
Speaker 3: Yeah, just on the topic of regulation not being in place,

94
00:05:24,120 --> 00:05:27,359
Speaker 3: the Federal Election Commission this month tried to vote on

95
00:05:27,600 --> 00:05:31,039
Speaker 3: a measure that would have enabled them to create rules

96
00:05:31,480 --> 00:05:36,200
Speaker 3: around AI generated political ads. Basically, the FEC would say

97
00:05:36,320 --> 00:05:39,800
Speaker 3: you have to declare when AI has generated your ad.

98
00:05:40,160 --> 00:05:43,119
Speaker 3: But they found themselves deadlocked. So there was a deep

99
00:05:43,160 --> 00:05:46,760
Speaker 3: disagreement among the commissioners, not even just along partisan lines,

100
00:05:46,960 --> 00:05:49,600
Speaker 3: about if they even have the authority to do that.

101
00:05:50,040 --> 00:05:55,640
Speaker 3: So the agencies that we have are currently deadlocked, unsure

102
00:05:55,640 --> 00:05:58,599
Speaker 3: how to proceed, and on Capitol Hill they're only beginning

103
00:05:58,600 --> 00:06:01,200
Speaker 3: to have these conversations.

104
00:06:02,480 --> 00:06:06,320
Speaker 1: Emily, why would the Federal Election Commission have any doubts

105
00:06:06,360 --> 00:06:08,400
Speaker 1: about what to do? Is there anyone in favor of

106
00:06:08,440 --> 00:06:11,640
Speaker 1: allowing fake stuff to be spread all over the place

107
00:06:11,680 --> 00:06:12,839
Speaker 1: without having to be labeled.

108
00:06:13,360 --> 00:06:16,599
Speaker 3: It's less about the actual measure and more about if

109
00:06:16,720 --> 00:06:19,599
Speaker 3: the FEC has the authority to do this. So we

110
00:06:19,640 --> 00:06:22,800
Speaker 3: actually run into this with tech related issues and our

111
00:06:22,839 --> 00:06:26,240
Speaker 3: government all the time. Most of these agencies weren't set

112
00:06:26,320 --> 00:06:30,880
Speaker 3: up to deal with issues like deep fakes, misinformation scams online,

113
00:06:31,000 --> 00:06:33,440
Speaker 3: and so they're having to adapt in real time, and

114
00:06:33,480 --> 00:06:37,240
Speaker 3: the FEC, some of the commissioners at least say that

115
00:06:37,279 --> 00:06:40,480
Speaker 3: they just don't have power over this part of the lot.

116
00:06:40,920 --> 00:06:44,720
Speaker 3: There are a lot of protections for political speech that

117
00:06:44,960 --> 00:06:48,880
Speaker 3: is an important bedrock of communication in America, so that

118
00:06:49,000 --> 00:06:51,400
Speaker 3: makes this issue even more difficult to deal with.

119
00:06:51,920 --> 00:06:54,800
Speaker 4: What makes the FEC really interesting in this is that

120
00:06:54,839 --> 00:06:57,880
Speaker 4: they are by designing a partisan agency with a mix

121
00:06:57,920 --> 00:07:00,920
Speaker 4: of Republicans and Democrats, and across the board you really

122
00:07:00,920 --> 00:07:04,440
Speaker 4: see them deadlocking on almost every issue. The FEC has

123
00:07:04,480 --> 00:07:07,480
Speaker 4: had very little success doing anything, which just points to

124
00:07:07,520 --> 00:07:10,320
Speaker 4: that the FEC realistically isn't going to be able to

125
00:07:10,320 --> 00:07:12,720
Speaker 4: take this one on based on all the other issues

126
00:07:12,720 --> 00:07:14,920
Speaker 4: facing the agency, and what about.

127
00:07:14,720 --> 00:07:19,000
Speaker 1: The candidates themselves. It's easy to put something out that's

128
00:07:19,040 --> 00:07:21,720
Speaker 1: fake about your opponent and maybe it gets you all

129
00:07:21,800 --> 00:07:23,560
Speaker 1: kinds of attention, but then you know they're going to

130
00:07:23,560 --> 00:07:26,400
Speaker 1: be doing it about you too, And so there's this

131
00:07:26,520 --> 00:07:29,520
Speaker 1: kind of AI fakery arms race that we seem to

132
00:07:29,560 --> 00:07:31,920
Speaker 1: be setting up. Is there any recognition on the part

133
00:07:31,960 --> 00:07:34,320
Speaker 1: of the campaigns that this is a really bad line

134
00:07:34,360 --> 00:07:34,880
Speaker 1: to cross?

135
00:07:36,280 --> 00:07:39,480
Speaker 3: So there is definitely an arms race when it comes

136
00:07:39,480 --> 00:07:42,480
Speaker 3: to AI, but there is an element of mutually assured

137
00:07:42,480 --> 00:07:46,400
Speaker 3: destruction where as soon as one campaign really goes out,

138
00:07:46,720 --> 00:07:49,200
Speaker 3: puts some words in their opponent's mouth that makes them

139
00:07:49,200 --> 00:07:53,960
Speaker 3: look bad, distributes images videos widely, then that sort of

140
00:07:54,160 --> 00:07:57,160
Speaker 3: opens this whole can of worms, and so there's going

141
00:07:57,240 --> 00:07:59,880
Speaker 3: to be tons of these kinds of ads, and actually

142
00:07:59,880 --> 00:08:03,480
Speaker 3: no one really wants that. So there's some internal reckoning

143
00:08:03,560 --> 00:08:07,800
Speaker 3: that's happening among political consultants and campaigns right now.

144
00:08:08,000 --> 00:08:10,160
Speaker 1: I guess they've been coming from these outside groups that

145
00:08:10,280 --> 00:08:12,680
Speaker 1: support campaigns so that they're able to have a certain

146
00:08:12,680 --> 00:08:17,040
Speaker 1: amount of deniability even though it's very clear which candidate

147
00:08:17,080 --> 00:08:17,760
Speaker 1: they're supporting.

148
00:08:18,200 --> 00:08:21,560
Speaker 4: But it creates for what we call like dark money groups.

149
00:08:21,560 --> 00:08:24,320
Speaker 4: So five H one C four political advacy groups. This

150
00:08:24,360 --> 00:08:26,160
Speaker 4: is a real wide opening for you know, groups that

151
00:08:26,200 --> 00:08:29,360
Speaker 4: aren't necessarily representing a specific person, but an interest area

152
00:08:29,600 --> 00:08:32,320
Speaker 4: or a larger party. The dark money groups, they can

153
00:08:32,360 --> 00:08:34,280
Speaker 4: go out and take all these images and they're super

154
00:08:34,360 --> 00:08:36,920
Speaker 4: cheap and easy to produce, so they can react in

155
00:08:36,960 --> 00:08:39,360
Speaker 4: real time. If something is said in a debate about

156
00:08:39,360 --> 00:08:41,160
Speaker 4: a candidate they don't like, they can have an image

157
00:08:41,240 --> 00:08:43,440
Speaker 4: or a video out there online by the end of

158
00:08:43,440 --> 00:08:45,320
Speaker 4: the debate. So that's what we're going to start to

159
00:08:45,320 --> 00:08:47,320
Speaker 4: see if this really rapid response. You know, what we

160
00:08:47,400 --> 00:08:49,400
Speaker 4: used to just see with with tweets or emails can

161
00:08:49,440 --> 00:08:52,000
Speaker 4: now be done with images and videos.

162
00:08:54,559 --> 00:08:57,280
Speaker 1: Laurie, we've been talking about how we're seeing this happening

163
00:08:57,320 --> 00:09:01,079
Speaker 1: at the national level, at the presidential level, but it's

164
00:09:01,120 --> 00:09:04,800
Speaker 1: also filtering down throughout politics in the US.

165
00:09:05,440 --> 00:09:05,680
Speaker 2: Yeah.

166
00:09:05,679 --> 00:09:08,000
Speaker 4: So AI is really cheap, and so that makes it

167
00:09:08,040 --> 00:09:11,040
Speaker 4: a really great tool for these smaller campaigns that don't

168
00:09:11,040 --> 00:09:12,400
Speaker 4: have a lot of money, that don't have a lot

169
00:09:12,400 --> 00:09:15,720
Speaker 4: of staff. So there's been in some mayoral elections recently

170
00:09:15,920 --> 00:09:19,120
Speaker 4: AI used both by campaigns to sort of support themselves

171
00:09:19,280 --> 00:09:21,880
Speaker 4: as well as attack ads. There was a really great

172
00:09:21,920 --> 00:09:24,560
Speaker 4: example in Shreveport, Louisiana, where there was an attack ad

173
00:09:24,640 --> 00:09:28,240
Speaker 4: against the incumbent Mayor Adrian Perkins, and it showed this

174
00:09:28,400 --> 00:09:30,640
Speaker 4: scene where it was you know, it was his face

175
00:09:30,720 --> 00:09:33,200
Speaker 4: on a body and his voice was talking, his mouth

176
00:09:33,280 --> 00:09:34,840
Speaker 4: was moving. It wasn't him, it was an actor and

177
00:09:34,840 --> 00:09:36,920
Speaker 4: then they had used AI to create it to make

178
00:09:36,960 --> 00:09:39,120
Speaker 4: it look like him. He ended up losing this race,

179
00:09:39,200 --> 00:09:41,640
Speaker 4: not necessarily because of this ad, but just a really

180
00:09:41,640 --> 00:09:44,600
Speaker 4: great example of how someone a developer in town who

181
00:09:44,640 --> 00:09:46,920
Speaker 4: didn't particularly care for the mayor could come in and

182
00:09:46,960 --> 00:09:50,160
Speaker 4: create this ad really quickly, easily and cheaply, and get

183
00:09:50,200 --> 00:09:52,160
Speaker 4: it on the airwaves and really make a splash with it.

184
00:09:52,800 --> 00:09:55,000
Speaker 4: In Toronto, we also saw the same thing in their

185
00:09:55,040 --> 00:09:59,840
Speaker 4: mayoral race, where candidate Anthony Fury used artificial intelligence images

186
00:10:00,000 --> 00:10:02,520
Speaker 4: in sort of his campaign materials, you know, showing packages

187
00:10:02,600 --> 00:10:05,040
Speaker 4: of crack pipe kits that were sponsored by the City

188
00:10:05,080 --> 00:10:07,600
Speaker 4: of Toronto, you know, using this to criticize policies of

189
00:10:07,600 --> 00:10:08,280
Speaker 4: his opponents.

190
00:10:08,679 --> 00:10:12,520
Speaker 1: Emily, you talked about how difficult it is for the

191
00:10:12,520 --> 00:10:14,760
Speaker 1: government to try and figure out how and whether they

192
00:10:14,760 --> 00:10:18,920
Speaker 1: can regulate this. Is there anything illegal about a candidate

193
00:10:19,679 --> 00:10:24,360
Speaker 1: making another candidate appear to be doing something completely false?

194
00:10:25,200 --> 00:10:29,400
Speaker 3: One use of AI that could be illegal, And we

195
00:10:29,440 --> 00:10:32,520
Speaker 3: can remember this is still an untested area of the

196
00:10:32,640 --> 00:10:35,559
Speaker 3: law in many ways, so we're not positive. But one

197
00:10:35,880 --> 00:10:38,680
Speaker 3: thing that could be illegal is if you use a

198
00:10:38,720 --> 00:10:44,560
Speaker 3: candidate's likeness to get money that the candidate doesn't actually receive.

199
00:10:45,080 --> 00:10:48,319
Speaker 3: So that is one thing that the FBC was discussing

200
00:10:48,600 --> 00:10:51,839
Speaker 3: that seems like a clear violation of our campaign laws

201
00:10:51,920 --> 00:10:54,960
Speaker 3: when you are not allowed to impersonate anyone else and

202
00:10:55,000 --> 00:10:58,000
Speaker 3: not allowed to lie about where money is going. There

203
00:10:58,080 --> 00:11:01,600
Speaker 3: are also laws dropping up in the states across the

204
00:11:01,640 --> 00:11:06,360
Speaker 3: country that ban certain kinds of quote unquote deep fakes.

205
00:11:06,400 --> 00:11:10,160
Speaker 3: So a lot of those laws pertained to pornography. So

206
00:11:10,200 --> 00:11:13,440
Speaker 3: there have been some reports that show over ninety percent

207
00:11:13,640 --> 00:11:17,120
Speaker 3: of the deep fakes on the Internet are pornographic images,

208
00:11:17,480 --> 00:11:22,600
Speaker 3: mostly featuring women, and so a bunch of states have

209
00:11:23,480 --> 00:11:26,160
Speaker 3: rushed in to fill this gap in the law, basically

210
00:11:26,280 --> 00:11:30,240
Speaker 3: just improving victims' rights, so people who were featured in

211
00:11:30,280 --> 00:11:34,440
Speaker 3: pornographic images on the internet can sue and have the

212
00:11:34,520 --> 00:11:37,360
Speaker 3: right to stop the proliferation of those. But there are

213
00:11:37,440 --> 00:11:41,640
Speaker 3: far fewer rules about election related deep fakes. There are

214
00:11:41,679 --> 00:11:47,559
Speaker 3: some they probably won't withstand legal challenges, so it's still

215
00:11:47,640 --> 00:11:50,920
Speaker 3: kind of complicated. Whenever you try to touch political speech,

216
00:11:51,320 --> 00:11:55,400
Speaker 3: but there are clearer violations after the break.

217
00:11:55,440 --> 00:11:59,920
Speaker 1: How AI is changing the way campaigns get their message.

218
00:11:59,520 --> 00:12:11,200
Speaker 5: Out, Laurie, you hinted a little bit before about how

219
00:12:11,240 --> 00:12:15,040
Speaker 5: campaigns are using AI for non nefarious reasons, that it's

220
00:12:15,040 --> 00:12:17,880
Speaker 5: just like become a useful tool being used by campaigns

221
00:12:17,920 --> 00:12:20,200
Speaker 5: the way it's used by a lot of different industries.

222
00:12:20,600 --> 00:12:22,640
Speaker 4: Yeah, so there's a bunch of different things that campaigns

223
00:12:22,640 --> 00:12:24,679
Speaker 4: are using and some that people would kind of clearly

224
00:12:24,679 --> 00:12:28,040
Speaker 4: put into a non nefarious camp. So things like helping

225
00:12:28,280 --> 00:12:31,080
Speaker 4: use AI to analyze voter rolls of people that they

226
00:12:31,080 --> 00:12:34,760
Speaker 4: should be targeting, organizing some of their data, figuring out

227
00:12:34,760 --> 00:12:38,360
Speaker 4: what kind of messages would resonate with which voters. Another

228
00:12:38,480 --> 00:12:41,080
Speaker 4: thing that they're doing is using tools like chat GBT,

229
00:12:41,240 --> 00:12:44,040
Speaker 4: not necessarily chat GPT because that tool has some restrictions

230
00:12:44,040 --> 00:12:46,680
Speaker 4: on political speech, but to write the first drafts of

231
00:12:46,800 --> 00:12:50,520
Speaker 4: press releases of fundraising emails. And they found a really

232
00:12:50,559 --> 00:12:54,480
Speaker 4: great cost savings instead of having five people who are

233
00:12:54,559 --> 00:12:57,599
Speaker 4: copywriters on a campaign working on all these messages, that

234
00:12:57,640 --> 00:13:00,400
Speaker 4: you'd have one person that is just telling the tool, okay,

235
00:13:00,440 --> 00:13:04,160
Speaker 4: write a fundraising email with a focus on democracy, right,

236
00:13:04,200 --> 00:13:06,000
Speaker 4: one with a focus on the environment right one with

237
00:13:06,040 --> 00:13:10,160
Speaker 4: a focus on defeeding Republicans, for example, and then they

238
00:13:10,200 --> 00:13:12,400
Speaker 4: can get something on the page and then just edit

239
00:13:12,440 --> 00:13:15,680
Speaker 4: it and it goes much much quicker. The one campaign

240
00:13:15,720 --> 00:13:18,200
Speaker 4: consultant described this as solving the blank page problem. Anyone

241
00:13:18,200 --> 00:13:20,640
Speaker 4: who's ever written anything can probably sympathize with that if

242
00:13:20,640 --> 00:13:22,319
Speaker 4: it's much easier to get something on the paper and

243
00:13:22,360 --> 00:13:25,160
Speaker 4: then edit it versus just starting from scratch. So they're

244
00:13:25,200 --> 00:13:27,400
Speaker 4: finding really big cost savings in that of they're able

245
00:13:27,440 --> 00:13:30,760
Speaker 4: to both raise more money because they're more effectively targeting

246
00:13:30,760 --> 00:13:33,160
Speaker 4: people as well as cut down on staff cost for

247
00:13:33,400 --> 00:13:36,640
Speaker 4: that particular function of writing. That's been really widespread and

248
00:13:36,880 --> 00:13:40,320
Speaker 4: one consultant estimated about fifty percent of campaigns are doing this,

249
00:13:40,440 --> 00:13:42,840
Speaker 4: but expect that number to be somewhere closer to one

250
00:13:42,920 --> 00:13:44,640
Speaker 4: hundred by the end of the twenty twenty four cycle.

251
00:13:45,120 --> 00:13:48,560
Speaker 3: But Congress just set new rules on how staffers can

252
00:13:48,720 --> 00:13:52,280
Speaker 3: use chat GPT, and the main concern is that a

253
00:13:52,320 --> 00:13:56,800
Speaker 3: lot of what you put into chat GPT can become

254
00:13:57,080 --> 00:13:59,720
Speaker 3: available in other ways. You know, there are a lot

255
00:13:59,720 --> 00:14:03,520
Speaker 3: of privacy concerns around chat GPT. So the new rules say,

256
00:14:04,040 --> 00:14:06,400
Speaker 3: first of all, you have to use the paid for

257
00:14:06,720 --> 00:14:11,439
Speaker 3: version of chatchept, which is better protective of privacy. Second

258
00:14:11,480 --> 00:14:14,840
Speaker 3: of all, you can't put confidential information into chat GPT,

259
00:14:15,240 --> 00:14:18,480
Speaker 3: which congressional staffers are often working with, so they're kind

260
00:14:18,559 --> 00:14:21,080
Speaker 3: of like making it up as they go along. But

261
00:14:21,360 --> 00:14:24,680
Speaker 3: congressional staffers work all day and night. They are never

262
00:14:24,760 --> 00:14:26,760
Speaker 3: off the clock, and so anything that cuts back on

263
00:14:26,960 --> 00:14:28,320
Speaker 3: time is really valuable.

264
00:14:31,400 --> 00:14:35,360
Speaker 1: Open AI. The makers of chat GPT, they are kind

265
00:14:35,400 --> 00:14:39,080
Speaker 1: of alarmed at the way campaigns could be using their

266
00:14:39,120 --> 00:14:41,280
Speaker 1: technology and or trying to figure out a way to

267
00:14:41,360 --> 00:14:41,840
Speaker 1: limit it.

268
00:14:42,320 --> 00:14:44,600
Speaker 4: If you put in really anything related to politics and

269
00:14:44,680 --> 00:14:48,040
Speaker 4: chat gept, there's sort of this standard disclaimer language coming

270
00:14:48,120 --> 00:14:49,640
Speaker 4: up of saying, you know, look, I don't have an

271
00:14:49,680 --> 00:14:52,360
Speaker 4: opinion on this. There's nuance to this, and so chat

272
00:14:52,440 --> 00:14:56,160
Speaker 4: GPT is being very careful with anything related to politics.

273
00:14:56,200 --> 00:14:57,880
Speaker 4: But if you push it a little bit, like if

274
00:14:57,920 --> 00:15:00,520
Speaker 4: you type who are the wokest Democrats, it will name

275
00:15:00,720 --> 00:15:05,120
Speaker 4: Alexandria Kazi Kort, says Rashida talib Aana Presley Bill on Omar.

276
00:15:05,440 --> 00:15:07,280
Speaker 4: Interesting that all like the top four or five names

277
00:15:07,280 --> 00:15:09,280
Speaker 4: that came up were all women. Sort of treading this

278
00:15:09,320 --> 00:15:12,080
Speaker 4: careful line, it's not completely staying out of the political arena.

279
00:15:12,600 --> 00:15:15,800
Speaker 4: So for a lot of political things, chat gibt has

280
00:15:15,840 --> 00:15:19,840
Speaker 4: basically stopped cataloging language from after September twenty twenty one,

281
00:15:20,080 --> 00:15:22,760
Speaker 4: So when you ask chat gibt a political question will

282
00:15:22,800 --> 00:15:25,400
Speaker 4: be like, here's my information as of September twenty twenty one.

283
00:15:25,600 --> 00:15:27,280
Speaker 4: You may need to do more research if you want

284
00:15:27,280 --> 00:15:29,520
Speaker 4: something more recent. So it's kind of just clamping down

285
00:15:29,560 --> 00:15:31,800
Speaker 4: and not feeding the beast, so to speak, with more

286
00:15:31,800 --> 00:15:34,480
Speaker 4: information to make it sort of an obsolete tool when

287
00:15:34,520 --> 00:15:35,840
Speaker 4: it comes to political speech.

288
00:15:36,040 --> 00:15:40,360
Speaker 3: They're also putting together something called classifiers, so open ai

289
00:15:40,600 --> 00:15:44,000
Speaker 3: is able to see what people put into chat GPT,

290
00:15:44,520 --> 00:15:47,240
Speaker 3: and so they're working hard now to figure out what

291
00:15:47,320 --> 00:15:52,200
Speaker 3: are the classifiers or identifying freezes that people use, where

292
00:15:52,280 --> 00:15:54,120
Speaker 3: when they're using those, you know that it's for a

293
00:15:54,120 --> 00:15:57,440
Speaker 3: political purpose, And then if that's happening, how can you

294
00:15:57,640 --> 00:16:01,400
Speaker 3: control what chat cept says in response? They really don't

295
00:16:01,400 --> 00:16:04,600
Speaker 3: want their product to be used to create lies and

296
00:16:04,640 --> 00:16:06,200
Speaker 3: misinformation that they're blamed for.

297
00:16:07,040 --> 00:16:09,120
Speaker 4: To sort of add though, CHATGBT is sort of the

298
00:16:09,160 --> 00:16:12,920
Speaker 4: big juggernaut and the industry of generative text AI, but

299
00:16:13,000 --> 00:16:15,280
Speaker 4: there's so many other tools that are being developed and

300
00:16:15,440 --> 00:16:18,680
Speaker 4: tools that are being developed specifically for political uses, so

301
00:16:18,760 --> 00:16:21,360
Speaker 4: usually these are paid tools that campaigns would subscribe to.

302
00:16:21,680 --> 00:16:24,000
Speaker 4: But just because chat GBT is sort of clamping down

303
00:16:24,000 --> 00:16:26,480
Speaker 4: on this, it doesn't mean that other actors can go

304
00:16:26,520 --> 00:16:29,440
Speaker 4: buy these tools that are specifically designed for this to

305
00:16:30,080 --> 00:16:31,840
Speaker 4: create some political misinformation.

306
00:16:32,680 --> 00:16:36,880
Speaker 1: Laura earlier, Emily painted this picture of mutual assured destruction

307
00:16:37,040 --> 00:16:39,520
Speaker 1: where if one candidate attacks the other and the other

308
00:16:39,600 --> 00:16:42,400
Speaker 1: attacks back, then everybody gets hurt. And in you're reporting,

309
00:16:42,440 --> 00:16:46,400
Speaker 1: you found that both Democrats and Republicans are using this technology,

310
00:16:46,400 --> 00:16:47,760
Speaker 1: but in slightly different ways.

311
00:16:48,880 --> 00:16:51,160
Speaker 4: So what we've seen in terms of the visual and

312
00:16:51,280 --> 00:16:54,120
Speaker 4: video side as Republicans using this more. We saw big

313
00:16:54,160 --> 00:16:56,920
Speaker 4: first video ad was the RNC attacking Biden. We saw

314
00:16:57,000 --> 00:17:00,400
Speaker 4: DeSantis attacking Trump with this. But both parties are definitely

315
00:17:00,480 --> 00:17:03,040
Speaker 4: using it, and probably Democrats are using it more. When

316
00:17:03,080 --> 00:17:05,600
Speaker 4: it comes to the text and generative AI, they have

317
00:17:05,760 --> 00:17:09,200
Speaker 4: firms that are using the text side a lot more.

318
00:17:09,400 --> 00:17:12,320
Speaker 4: That's really widespread among campaigns. This is a really hard

319
00:17:12,320 --> 00:17:14,919
Speaker 4: thing that's hard to track you know exactly how widespread

320
00:17:14,920 --> 00:17:18,480
Speaker 4: it is within either party. Democrats are definitely more cautious

321
00:17:18,680 --> 00:17:21,760
Speaker 4: on using the visual and image side. They kind of

322
00:17:21,800 --> 00:17:24,119
Speaker 4: like to say, look, we're being responsible here, We're not

323
00:17:24,200 --> 00:17:28,680
Speaker 4: going to create misinformation or deep fakes. Republicans, particularly because

324
00:17:28,680 --> 00:17:31,000
Speaker 4: they're going into a contested primary so they have more

325
00:17:31,040 --> 00:17:32,919
Speaker 4: going on right now, but they're a little bit more

326
00:17:32,960 --> 00:17:35,359
Speaker 4: willing to try out this tool and see how it works,

327
00:17:35,400 --> 00:17:37,960
Speaker 4: while Democrats are a little bit hanging back on the sidelines.

328
00:17:38,400 --> 00:17:41,439
Speaker 3: I spoke to a spokesperson for the Democratic National Committee

329
00:17:41,720 --> 00:17:47,160
Speaker 3: who said Democrats are wary of misinformation. There's a history

330
00:17:47,200 --> 00:17:50,800
Speaker 3: there that goes back to the twenty sixteen election. There

331
00:17:50,880 --> 00:17:55,840
Speaker 3: is a whole team dedicated to stopping misinformation at the DNC,

332
00:17:56,400 --> 00:17:59,359
Speaker 3: so that makes them extra cautious as they begin to

333
00:17:59,440 --> 00:18:03,160
Speaker 3: implement AI slowly into things like fundraising.

334
00:18:05,400 --> 00:18:08,480
Speaker 1: Emily, that's an interesting point that the Democrats are saying

335
00:18:08,480 --> 00:18:10,320
Speaker 1: that they're going to really need to be monitoring this.

336
00:18:10,400 --> 00:18:11,919
Speaker 1: Do you think that we're going to have now this

337
00:18:11,960 --> 00:18:16,920
Speaker 1: whole campaign industry where groups are doing AI checks similar

338
00:18:16,960 --> 00:18:18,560
Speaker 1: to the way we have fact checkers.

339
00:18:19,280 --> 00:18:24,880
Speaker 3: Yeah, there is this anxiety and wariness among political consultants

340
00:18:25,080 --> 00:18:28,880
Speaker 3: about how AI is going to be deployed. So the

341
00:18:29,000 --> 00:18:34,400
Speaker 3: trade Association for Political Consultants, which exists, shockingly came out

342
00:18:34,400 --> 00:18:37,639
Speaker 3: with this statement saying we are against the use of

343
00:18:37,680 --> 00:18:41,080
Speaker 3: deep fakes in political campaigns, and we really shouldn't be

344
00:18:41,160 --> 00:18:45,000
Speaker 3: messing too much with this technology. It's quite dangerous. So

345
00:18:45,200 --> 00:18:49,160
Speaker 3: I do think that increasingly there is desire to call

346
00:18:49,240 --> 00:18:53,719
Speaker 3: out misinformation, to call out when generative AI is being used.

347
00:18:54,040 --> 00:18:56,919
Speaker 3: I also think that a lot of technology companies are

348
00:18:56,920 --> 00:18:59,480
Speaker 3: about to make a lot of money off of basically

349
00:18:59,560 --> 00:19:03,000
Speaker 3: tools that allow you to identify what's real. Those are

350
00:19:03,080 --> 00:19:07,119
Speaker 3: going to be commonly used, at least among political campaigns,

351
00:19:07,119 --> 00:19:08,560
Speaker 3: if not the general population.

352
00:19:09,640 --> 00:19:12,720
Speaker 1: When we come back, what can we do to avoid

353
00:19:12,760 --> 00:19:25,920
Speaker 1: getting duped? So what are we supposed to do about this?

354
00:19:26,200 --> 00:19:31,080
Speaker 1: Just as citizens watching a campaign, trying to make good decisions,

355
00:19:31,240 --> 00:19:35,879
Speaker 1: trying to figure out what's real and what's not. I mean, Laura,

356
00:19:36,000 --> 00:19:39,120
Speaker 1: you kind of jokingly at the beginning said, oh, look

357
00:19:39,119 --> 00:19:41,520
Speaker 1: at the hands, let's eve it as six fingers. But

358
00:19:41,960 --> 00:19:44,920
Speaker 1: this is a pretty serious thing. How are we going

359
00:19:44,960 --> 00:19:47,399
Speaker 1: to be able to discern what's real and what's not.

360
00:19:48,600 --> 00:19:50,560
Speaker 4: This is probably when tech companies are going to have

361
00:19:50,640 --> 00:19:53,520
Speaker 4: to start stepping in. We've seen Twitter and Facebook come

362
00:19:53,560 --> 00:19:57,880
Speaker 4: in with labels on misinformation regarding vaccines or other things.

363
00:19:57,960 --> 00:20:01,000
Speaker 4: Regarding the twenty twenty election, we're probably going to see

364
00:20:01,040 --> 00:20:02,560
Speaker 4: those start to come in. I don't know that the

365
00:20:02,560 --> 00:20:05,720
Speaker 4: tech companies, they don't have fully fledged policies yet on

366
00:20:05,760 --> 00:20:07,639
Speaker 4: how to do that, and part of it is that

367
00:20:07,680 --> 00:20:10,840
Speaker 4: it is really difficult to check. We see companies like

368
00:20:10,880 --> 00:20:13,240
Speaker 4: Google saying, Okay, we're going to put these markers in

369
00:20:13,320 --> 00:20:16,119
Speaker 4: these images. That's probably the next step that we're going

370
00:20:16,160 --> 00:20:18,399
Speaker 4: to see of each company, because they want to be

371
00:20:18,440 --> 00:20:20,879
Speaker 4: seen as responsible and because they don't want an image

372
00:20:20,880 --> 00:20:23,160
Speaker 4: created on their platform to be the thing that blows

373
00:20:23,240 --> 00:20:26,080
Speaker 4: up an election, for example. You're going to see pressure

374
00:20:26,280 --> 00:20:28,720
Speaker 4: in the industry to do more, but it's right now,

375
00:20:28,800 --> 00:20:31,400
Speaker 4: it's really incumbent upon the viewer. So it is kind

376
00:20:31,400 --> 00:20:32,920
Speaker 4: of looking at the image, look to see if there's

377
00:20:32,920 --> 00:20:36,000
Speaker 4: a little tag in the corner that mentions it's AI generated,

378
00:20:36,119 --> 00:20:37,960
Speaker 4: or a screen at the very end of an ad

379
00:20:38,000 --> 00:20:39,760
Speaker 4: that says this was created by AI.

380
00:20:40,200 --> 00:20:43,120
Speaker 1: You know, in previous campaigns, we've seen Facebook, we've seen

381
00:20:43,160 --> 00:20:46,560
Speaker 1: Twitter make big public announcements, They've hired all of these

382
00:20:46,560 --> 00:20:48,480
Speaker 1: people who are going to be monitoring, they're going to

383
00:20:48,480 --> 00:20:50,520
Speaker 1: try to crack down on fakery, and yet they just

384
00:20:50,720 --> 00:20:53,280
Speaker 1: could not possibly do it in the case of Twitter,

385
00:20:53,400 --> 00:20:56,840
Speaker 1: we're seeing all kinds of fakery all over the platform,

386
00:20:56,960 --> 00:21:00,840
Speaker 1: especially since Elon Musk took over the company and kind

387
00:21:00,840 --> 00:21:03,920
Speaker 1: of opened it up in a way that previously people

388
00:21:03,960 --> 00:21:05,119
Speaker 1: couldn't post stuff on it.

389
00:21:05,960 --> 00:21:09,520
Speaker 3: Yes, but we're also seeing tools like Twitter's community notes,

390
00:21:09,560 --> 00:21:13,840
Speaker 3: which allows everyday users to weigh in on whether something

391
00:21:13,960 --> 00:21:18,240
Speaker 3: is true or false, that was already under construction before

392
00:21:18,280 --> 00:21:21,359
Speaker 3: Elon Musk came in. It's actually proven to be a

393
00:21:21,400 --> 00:21:27,119
Speaker 3: really effective obstacle for the spread of generative AI misinformation.

394
00:21:27,320 --> 00:21:29,440
Speaker 3: So if we think about the ad Byron de Sampas's

395
00:21:29,560 --> 00:21:32,760
Speaker 3: campaign within hours that had a Twitter note on it

396
00:21:33,000 --> 00:21:36,200
Speaker 3: saying tech experts have looked at this and these images

397
00:21:36,240 --> 00:21:39,399
Speaker 3: are not real. So if you think about what you

398
00:21:39,480 --> 00:21:42,800
Speaker 3: can do as an individual, it's a mitzvah to just

399
00:21:43,040 --> 00:21:48,400
Speaker 3: get out the word that something is fake, that maybe

400
00:21:48,720 --> 00:21:54,720
Speaker 3: tech experts are weighing in and just spreading that information together.

401
00:21:55,200 --> 00:21:57,280
Speaker 3: People create media literacy that way.

402
00:21:58,000 --> 00:22:02,240
Speaker 4: And the thing too is ethicists are less concerned about

403
00:22:02,240 --> 00:22:04,639
Speaker 4: people like Biden and Donald Trump that have swarms of

404
00:22:04,640 --> 00:22:07,320
Speaker 4: press following everywhere they go and everything they do. But

405
00:22:07,440 --> 00:22:10,800
Speaker 4: on the local level, where local news isn't as strong,

406
00:22:10,840 --> 00:22:13,760
Speaker 4: where there are people and candidates and politicians who aren't

407
00:22:13,760 --> 00:22:17,440
Speaker 4: necessarily followed as closely. It's way easier to spread misinformation

408
00:22:17,520 --> 00:22:20,080
Speaker 4: in these areas where there is more of a news vacuum.

409
00:22:20,440 --> 00:22:22,920
Speaker 1: Emily Laura, thanks so much for coming on the show.

410
00:22:23,280 --> 00:22:25,080
Speaker 3: Thanks so much for having us.

411
00:22:25,119 --> 00:22:28,040
Speaker 1: Thank you, thanks for listening to us here at The

412
00:22:28,040 --> 00:22:31,360
Speaker 1: Big Take. It's a daily podcast from Bloomberg and iHeartRadio.

413
00:22:31,680 --> 00:22:36,120
Speaker 1: For more shows from iHeartRadio, visit the iHeartRadio app, Apple Podcasts,

414
00:22:36,200 --> 00:22:38,920
Speaker 1: or wherever you listen, and we'd love to hear from you.

415
00:22:39,200 --> 00:22:42,400
Speaker 1: Email us questions or comments to Big Take at Bloomberg

416
00:22:42,440 --> 00:22:45,760
Speaker 1: dot net. The supervising producer of The Big Take is

417
00:22:45,880 --> 00:22:50,560
Speaker 1: Vicky Virgalina. Our senior producer is Catherine Fink. Federica Romanello

418
00:22:50,720 --> 00:22:55,440
Speaker 1: is our producer. Our associate producer is Zeneb Sidiki. Rafael

419
00:22:55,600 --> 00:22:59,000
Speaker 1: m Seely is our engineer. Our original music was composed

420
00:22:59,000 --> 00:23:02,600
Speaker 1: by Leo Sidrin. I'm Westkasova. We'll be back tomorrow with

421
00:23:02,720 --> 00:23:03,680
Speaker 1: another Big Take.