The TED AI Show: Humanity’s first AI election w/ WIRED's Vittoria Elliott

[00:00:00] [SPEAKER_02]: Audio Collective.

[00:00:38] [SPEAKER_04]: Following on X1 Day that I realized what the next few months might look like.

[00:00:42] [SPEAKER_04]: X had just updated grok, their AI chatbot so that it could generate images using a largely

[00:00:48] [SPEAKER_04]: un-censored open source model called flux.

[00:00:51] [SPEAKER_04]: And immediately, the results were far more unhinged than anything we've seen.

[00:00:56] [SPEAKER_04]: Now we've been able to generate images in the past, but this is the first time we've

[00:01:00] [SPEAKER_04]: had these capabilities put into a social media app that 250 million people use everyday.

[00:01:06] [SPEAKER_04]: I saw photo realistic pictures of Kamala Harris and Donald Trump in bizarre situations.

[00:01:12] [SPEAKER_04]: That range from easily-click-badable images of them lovingly holding hands to skin-crawling

[00:01:17] [SPEAKER_04]: images of the two of them celebrating 9-11.

[00:01:21] [SPEAKER_04]: Suddenly, my feed was full of wild, harrous and Trump memes, and I started to realize

[00:01:26] [SPEAKER_04]: It's kind of impossible to imagine a future where AI won't have an impact on elections moving

[00:01:31] [SPEAKER_04]: forward.

[00:01:34] [SPEAKER_04]: It feels like uncharted territory, but actually in many ways the US is late to this

[00:01:41] [SPEAKER_04]: specific party.

[00:01:42] [SPEAKER_04]: Other countries around the world have been adapting their elections alongside the rise

[00:01:46] [SPEAKER_04]: of AI and it goes beyond just means, deep fakes and chatbots have been deployed in surprising

[00:01:52] [SPEAKER_04]: ways in some of the world's biggest elections, like in India and Pakistan.

[00:01:57] [SPEAKER_04]: And in Europe, there's even AI candidates that are run for office.

[00:02:03] [SPEAKER_04]: So what can we learn from the rest of the world to prepare for our first AI election?

[00:02:08] [SPEAKER_04]: And is it really just doom and gloom or can AI actually be good for democracy?

[00:02:18] [SPEAKER_04]: I'm below the sedu, and this is the Tediasha, where we figure out how to live and thrive in a

[00:02:23] [SPEAKER_04]: world where AI is changing everything.

[00:02:32] [SPEAKER_03]: Imagine this. In 2030, the CFO of a Fortune 100 company is a bot.

[00:02:38] [SPEAKER_03]: I'm Paul Michaelman, and on Imagine This will be exploring possible futures

[00:02:42] [SPEAKER_03]: and the implications they hold for organizations. Joining me will be BCG's top experts

[00:02:47] [SPEAKER_02]: as well as my co-host gene, BCG's conversational Gen AI agent.

[00:02:52] [SPEAKER_02]: Blending human creativity with AI innovation, this podcast promises an unmasked listening journey.

[00:02:59] [SPEAKER_02]: Join us on Imagine This from BCG.

[00:03:02] [SPEAKER_00]: Hey Tediah listeners, if you enjoy this show, we think you'll enjoy a new show about how tech

[00:03:07] [SPEAKER_00]: is changing our politics. It's called Wired Politics Lab, hosted by Wired Senior Politics Editor

[00:03:13] [SPEAKER_00]: Leah Fyger, and a rotating cast of Wired's political reporters. Each week, you'll be guided

[00:03:19] [SPEAKER_00]: through the exciting, challenging and sometimes entertaining vortex of internet extremism,

[00:03:24] [SPEAKER_00]: conspiracies, and disinformation. From AI chat bots running for political office to deep-fix

[00:03:29] [SPEAKER_00]: influencing voters, you'll hear in-depth analysis and conversations based on facts and research.

[00:03:34] [SPEAKER_00]: Find and follow Wired Politics Lab wherever you listen.

[00:03:42] [SPEAKER_04]: Today I'm talking with Victoria Elliott, a reporter at Wired who specializes in disinformation

[00:03:47] [SPEAKER_04]: and social media platforms, and she's in the middle of a massive endeavor to collect stories

[00:03:52] [SPEAKER_04]: about AI and elections from around the world for Wired's AI Elections Project.

[00:03:58] [SPEAKER_04]: If you go to the website, you'll see they've got a geospatial map that's basically tracking

[00:04:02] [SPEAKER_04]: all the ways this tech is being used. When I checked it out, I was struck by how important a tool

[00:04:07] [SPEAKER_04]: like this can be. It's a perfect example of what I'd call open source intelligence,

[00:04:12] [SPEAKER_04]: a compilation of public sources that can reflect relevant insights and build awareness.

[00:04:17] [SPEAKER_04]: And who knows, maybe those insights could be the key to updating our democracies firmware

[00:04:21] [SPEAKER_04]: ahead of our first AI election. I was really interested in asking Victoria to walk me through

[00:04:27] [SPEAKER_04]: some of the most important stories she'd come across and how AI's reshaping elections around the globe.

[00:04:38] [SPEAKER_01]: Hi, Victoria. Welcome to the show. Hi, Billov. Well, thank you so much for having me.

[00:04:44] [SPEAKER_04]: So let's get right into it. Can you start by walking us through the Wired AI Elections Project?

[00:04:49] [SPEAKER_01]: Yeah. So this year is the biggest election year. Pretty much ever. There's about 65 elections

[00:04:56] [SPEAKER_01]: that are happening around the world. This is a massive year for democracy and it's also not only

[00:05:01] [SPEAKER_01]: the biggest election we've probably ever seen, but it is certainly the biggest election

[00:05:05] [SPEAKER_01]: since the advent of the internet and social media. And I think we saw it as a moment to really take

[00:05:12] [SPEAKER_01]: stock of how technology and democracy are interacting, particularly in this moment of generative

[00:05:19] [SPEAKER_01]: AI. And so part of it was to see how is this actually going to impact democracy, but also, you know,

[00:05:25] [SPEAKER_01]: are the things that people were afraid of? Are those the trends we're going to see or is it

[00:05:29] [SPEAKER_01]: going to be something else? And how are those trends going to look differently in different parts

[00:05:33] [SPEAKER_01]: of the world? So much of technology discourse is really US-centric because a lot of big

[00:05:38] [SPEAKER_01]: technology companies are US companies. But that doesn't mean that's what that technology or

[00:05:44] [SPEAKER_01]: that innovation is going to look like in other places, that doesn't mean that's how it's going

[00:05:47] [SPEAKER_01]: to be used in other contexts. We have a big world map that shows every country that's having

[00:05:52] [SPEAKER_01]: election this year and then whether or not there's been instances of generative AI related to

[00:05:58] [SPEAKER_01]: election showing up. So this could look like the Joe Biden Robo call that happened right before

[00:06:04] [SPEAKER_01]: the New Hampshire primaries, where a voice clone of Joe Biden called all these voters in New

[00:06:10] [SPEAKER_01]: Hampshire discouraging them from voting in the Democratic primary. That's an instance of generative

[00:06:15] [SPEAKER_01]: AI, but then we're also seeing things like in Pakistan, the former Prime Minister in Ron Khan,

[00:06:22] [SPEAKER_01]: a deep fake of him that he authorized, that he, his party authorized declaring victory,

[00:06:29] [SPEAKER_01]: speaking to people because he was in prison. And so we see both these like really interesting examples

[00:06:34] [SPEAKER_01]: of something being like sanctioned and unsanctioned. So we have these types of examples in the project

[00:06:40] [SPEAKER_01]: and they're broken down into regional categories like North America, Africa, Europe, etc.

[00:06:46] [SPEAKER_01]: And so a user can look at the map and see how far this technology is reaching in terms of

[00:06:52] [SPEAKER_01]: where it's showing up in elections and description of what happened when it happened, where it happened

[00:06:58] [SPEAKER_01]: and then a link out to like other sourcing that we used to sort of confirm that this was real.

[00:07:04] [SPEAKER_04]: I think it's super exciting to have this all collated in one place. It's also slightly disconcerting.

[00:07:09] [SPEAKER_04]: And on that note, you know, I think when people think about generative AI and elections,

[00:07:14] [SPEAKER_04]: people immediately jump to deepfakes. How are deepfakes affecting campaigns in the U.S. so far?

[00:07:21] [SPEAKER_04]: This national level you gave the Biden example, but also the local level.

[00:07:25] [SPEAKER_01]: The deepfakes I think are the most visible form and sometimes the easiest to detect.

[00:07:30] [SPEAKER_01]: So I think that's why we see so many instances of that. And I think even in the project,

[00:07:34] [SPEAKER_01]: you'll see that there is a bias towards visual media because that's just easier in many cases to confirm

[00:07:40] [SPEAKER_01]: that something's fake. And a lot of times just because we know something's fake,

[00:07:46] [SPEAKER_01]: doesn't necessarily mean it's still not emotionally resonant. So I think a really great example of

[00:07:50] [SPEAKER_01]: is a couple weeks ago Elon Musk, owner of formerly Twitter currently X, tweeted a parody video

[00:07:57] [SPEAKER_01]: of Vice President and Democratic presidential nominee, Kamala Harris saying she's the ultimate

[00:08:03] [SPEAKER_01]: DI higher, which is one of the things that the right really likes to say about her, but it was

[00:08:07] [SPEAKER_01]: using her voice saying that in the style of a campaign ad. And when he initially tweeted it out,

[00:08:14] [SPEAKER_01]: it didn't have a disclaimer saying that it was parody, it didn't have labels saying that it was

[00:08:20] [SPEAKER_01]: AI generated, which many platforms require. But I think even people who really do believe that

[00:08:26] [SPEAKER_01]: about Kamala Harris might know that that video's not real, but it's still really emotionally

[00:08:32] [SPEAKER_01]: resonant because it's the voice of this politician saying something that people kind of already

[00:08:38] [SPEAKER_01]: believed to be true. It's less about necessarily that people are getting tricked than the fact that

[00:08:44] [SPEAKER_01]: they are seeing out in the world represented something that up until this point, they've sort of

[00:08:50] [SPEAKER_01]: only personally believed to be true. And then I think when we're looking at the local level,

[00:08:54] [SPEAKER_01]: a really interesting example is Peter Dixon who's a congressional candidate in California.

[00:09:00] [SPEAKER_01]: He used AI in a campaign ad to sort of like look like he was jumping through these various

[00:09:05] [SPEAKER_01]: points in his life in different locations to sort of illustrate his background. And then we had

[00:09:11] [SPEAKER_01]: the example of Jason Palmer, who is an investor and businessman who ran in the Democratic

[00:09:17] [SPEAKER_01]: primaries in American Smoa. And he used very openly AI to conduct that campaign. He used AI

[00:09:24] [SPEAKER_01]: to have a deep fake avatar of himself. And it answered questions about his public policy,

[00:09:30] [SPEAKER_01]: people could ask it, so there's this sort of broad swath of how it can be used in a really

[00:09:36] [SPEAKER_01]: sort of legitimate way to be like, hey, this man is not going to assert fly to American Smoa,

[00:09:41] [SPEAKER_01]: but how can people in American Smoa feel connected to this candidate? Maybe it would be good for

[00:09:45] [SPEAKER_01]: him to have an AI avatar that can answer questions that can state his policy positions.

[00:09:51] [SPEAKER_01]: And then you know, on a more hyperlocal level, there is someone who's actually running for

[00:09:56] [SPEAKER_01]: mayor of Cheyenne while you're being as an AI candidate as in he has created an AI bot and the

[00:10:06] [SPEAKER_01]: right at citizen. And Vic is ingesting thousands and thousands of city council documents to make

[00:10:14] [SPEAKER_01]: policy decisions. And that's what we're seeing on a local level. And that's not a deep fake.

[00:10:18] [SPEAKER_01]: So I think we're seeing them using all these really innovative ways that go beyond just trying

[00:10:22] [SPEAKER_01]: to trick people, but more around how can they be useful to campaigns? And very specifically,

[00:10:28] [SPEAKER_01]: how can they make people feel connected to these issues, even if people know that thing they're

[00:10:34] [SPEAKER_04]: dealing with is not a real human. You know, it strikes me that there is this sort of sliding scale

[00:10:39] [SPEAKER_04]: of like pure utility to let's say like deceptiveness on the other end and it is a blurry line.

[00:10:45] [SPEAKER_04]: The example that you gave of, you know, a politician basically making themselves more accessible,

[00:10:50] [SPEAKER_04]: almost like a async virtual town hall where you can go ask this politician questions and maybe

[00:10:55] [SPEAKER_04]: learn about their views a little bit better and a more intuitive fashion. It's also interesting

[00:11:00] [SPEAKER_04]: to just see folks like bolster, they're like the media that they're going to push out there to like

[00:11:05] [SPEAKER_04]: illustrating the various stages of a person's journey and being able to like bring people

[00:11:09] [SPEAKER_04]: along for the ride. And it just make that feel a little bit more

[00:11:12] [SPEAKER_04]: transported and authentic. I think is very exciting. And then the last piece, which is, you know,

[00:11:18] [SPEAKER_04]: the meme one, you can create content that is overtly, hey, this is intended to be fake,

[00:11:24] [SPEAKER_04]: but it can still have a visceral impact on you both in a good sense. And maybe in a negative sense

[00:11:30] [SPEAKER_04]: to or kind of bypasses your firewall because it has this emotional reaction in you. And even if

[00:11:36] [SPEAKER_04]: you know, hey, this is factually incorrect, you're still going to be influenced by that, right?

[00:11:42] [SPEAKER_01]: Totally. I mean, I think when we think about the use of AI and satire and parody in situations

[00:11:49] [SPEAKER_01]: where again, people know its fake. That does just because people know its fake doesn't mean it's

[00:11:55] [SPEAKER_04]: resonant to them. On that note, the lines do get blurry at times. You know, we're getting to this

[00:12:00] [SPEAKER_04]: point where events and photos are being called in quotes AI. Of course, the recent example is

[00:12:06] [SPEAKER_04]: you know, Donald Trump claimed that, you know, Kamala's crowd size was AI in quotes. How do you

[00:12:12] [SPEAKER_04]: think this dynamic alter is how to run a campaign when you can like call out like factual things

[00:12:17] [SPEAKER_04]: as being fake? Even though I'd assume these are exceedingly photographed events? So there's actually

[00:12:23] [SPEAKER_01]: term for this. It's called the Liars dividend and it's sort of the idea that when everything

[00:12:30] [SPEAKER_01]: could possibly be fake, nothing is real as someone who used to cover specifically tech in the global

[00:12:36] [SPEAKER_01]: south, everything comes home to roost. We see it elsewhere before it comes here. We saw that with the

[00:12:41] [SPEAKER_01]: abusive social media for disinformation campaigns in places like India or in the Philippines before

[00:12:47] [SPEAKER_01]: it became a problem in the US. And I think when we're talking about the Liars dividend, same thing,

[00:12:53] [SPEAKER_01]: last year in India, we had politicians who really, to recordings of them were being shared, saying

[00:13:00] [SPEAKER_01]: bad stuff and their immediate response was, that's fake. That's not real. It's AI. And you know,

[00:13:08] [SPEAKER_01]: I think back on the Donald Trump access Hollywood tape. I think that happened now. He would just say

[00:13:15] [SPEAKER_01]: that was fake audio. He would just say that was AI generated. And so I think what we're really

[00:13:19] [SPEAKER_01]: going to see is this further blurring of a shared sense of reality and a shared sense of truth,

[00:13:26] [SPEAKER_04]: because if nothing is real, anything is possible. There's something so dare I say earth shattering

[00:13:34] [SPEAKER_04]: about that because it's like at the core of weight, I can't trust my eyes anymore because you know,

[00:13:39] [SPEAKER_04]: there's real evidence of wrongdoing. Now more and more people are deploying it and employing it.

[00:13:45] [SPEAKER_04]: I have to ask you when you see this happening and now you're tracking all of these examples. Good and bad,

[00:13:51] [SPEAKER_04]: where do you net out on your like zero to 10 scale of like excited to extremely scared spectrum?

[00:13:58] [SPEAKER_01]: I think I net in the middle, mostly because what this to me shows is the incredible

[00:14:10] [SPEAKER_01]: mispriorization of what we're innovating for on a sort of grander scale. Rather than like thinking

[00:14:15] [SPEAKER_01]: about the social impacts, Silicon Valley has always been move fast break things and I think

[00:14:21] [SPEAKER_01]: the AI thing is always very interesting to me because it repeats so many of the mistakes of

[00:14:26] [SPEAKER_01]: web 2.0. The idea of deploying technologies before we know their social impact,

[00:14:31] [SPEAKER_01]: the idea of not thinking about how something's going to be used outside of the context that you thought

[00:14:36] [SPEAKER_01]: you were designing for, not necessarily being able to differentiate like what's real and good

[00:14:41] [SPEAKER_01]: and what's not not being able to control who uses your product. Like all of these sort of baseline

[00:14:47] [SPEAKER_01]: things that we're still trying to figure out like we don't even understand on a policy level

[00:14:52] [SPEAKER_01]: and even I think on a company level how we deal with content moderation. That has been the

[00:14:57] [SPEAKER_01]: conversation from basically 2010 onwards. It's been how do we deal with content moderation on web 2.0?

[00:15:02] [SPEAKER_01]: And instead of thinking about like how can we deal with this, it's all the same mistakes,

[00:15:08] [SPEAKER_01]: all the same behaviors all over again in different iterations for the AI revolution and it doesn't

[00:15:14] [SPEAKER_04]: really feel like we've learned very much. Now it strikes me even web 2.0 and let's say social

[00:15:19] [SPEAKER_04]: being an exemplification of that. It ended up being quite valuable in the election context as well.

[00:15:27] [SPEAKER_04]: Right? I remember reading this article where it's like when the nerds go marching into the White

[00:15:30] [SPEAKER_04]: House, this is like Obama's reelection campaign circa 2012. And so I'm kind of curious,

[00:15:36] [SPEAKER_04]: given the tools at our disposal, do you think that AI could be an effective policy messaging tool?

[00:15:43] [SPEAKER_04]: If it can be factual and it seems like there's certain approaches to make it more factual than not.

[00:15:48] [SPEAKER_01]: Yeah, I mean the biggest issue is transparency. So I think a really good example for instance is

[00:15:53] [SPEAKER_01]: like Indonesia during their election this year, the man who was formerly the head of the military

[00:15:59] [SPEAKER_01]: under the country's former dictator won that election. And there were people who worked for

[00:16:06] [SPEAKER_01]: his party that openly admitted to using AI and specifically to write campaign speeches,

[00:16:11] [SPEAKER_01]: and that's great except apparently they had built that on top of chatGPT, which open AI

[00:16:18] [SPEAKER_01]: has said we don't want our product use for politics. How are they tracking that? How are they

[00:16:23] [SPEAKER_01]: ensuring that's not happening? So I think like yeah, there could be really valuable ways for this

[00:16:28] [SPEAKER_01]: to be used for, say, campaign messaging or like the example that we saw in American Smoa,

[00:16:34] [SPEAKER_01]: where someone can sort of as you describe like an asynchronous town hall.

[00:16:37] [SPEAKER_04]: You know this kind of brings me back to how you see social media factoring into all of this.

[00:16:43] [SPEAKER_04]: I mean since the last presidential election, we've seen how important it is in campaigning,

[00:16:49] [SPEAKER_04]: but also in spreading misinformation. The FEC made a decision earlier this year that they

[00:16:55] [SPEAKER_04]: wouldn't put new restrictions on AI and political ads, which means it'll be up on platforms.

[00:17:00] [SPEAKER_04]: And you gave this example of somebody like illustrated the jury and stages of their life

[00:17:05] [SPEAKER_04]: in a campaign ad, what are platforms doing right now to like prepare for this like very pivotal

[00:17:11] [SPEAKER_04]: year to curb perhaps AI-generated content related to elections? Are they trying to get ahead of it?

[00:17:17] [SPEAKER_04]: Or do you think it'll again be sort of more responsive and retroactive? So first off,

[00:17:22] [SPEAKER_01]: platforms, you know, everything they're doing right now is voluntary. Number two, you know,

[00:17:27] [SPEAKER_01]: a lot of them have really leaned into labeling, or they're saying, you know, it has to be

[00:17:32] [SPEAKER_01]: labeled if there's generated an AI content on our platform. But so often,

[00:17:39] [SPEAKER_01]: it is very difficult for them to detect AI-generated stuff on their platform. We don't really

[00:17:46] [SPEAKER_01]: have a ton of transparency into how they're detecting this stuff. What systems they have in place?

[00:17:53] [SPEAKER_01]: Companies are not necessarily sharing data back and forth about what's being generated on their

[00:17:58] [SPEAKER_01]: platform, so something can be labeled consistently across platforms. And we saw this also with

[00:18:03] [SPEAKER_01]: disinformation issues where like, you know, maybe something would get taken down on Twitter, but it

[00:18:07] [SPEAKER_01]: would live on Facebook or maybe it would get taken off a Facebook, but it still be platformed on YouTube.

[00:18:11] [SPEAKER_01]: Like, you know, there's all these sort of gaps because these companies are not sharing information

[00:18:17] [SPEAKER_01]: or coordinating with each other except on things like child sexual abuse material or like

[00:18:21] [SPEAKER_01]: terrorism where there's actual legal implications, right? A lot of companies that have AI models have

[00:18:26] [SPEAKER_01]: said, you know, we're going to start water marking, meaning that anything that comes off our platform.

[00:18:30] [SPEAKER_01]: It'll have some sort of signature on it whether or not that's visible to a human or not,

[00:18:35] [SPEAKER_01]: but that a machine can read to say, hey, this is AI-generated. Okay, well, that's great.

[00:18:40] [SPEAKER_01]: But that implies that everyone has to be a good actor. Yeah, exactly.

[00:18:44] [SPEAKER_01]: Everyone has to agree to watermark and everyone has to be able to read everybody else's watermarks.

[00:18:48] [SPEAKER_01]: Right? If you're a bad actor, you have no incentive to use technology that's going to watermark

[00:18:54] [SPEAKER_01]: your content. You know, we're talking about something that's detectable by a machine,

[00:18:59] [SPEAKER_01]: but will that be detectable by a human who knows? Yeah, this is going to be so challenging.

[00:19:04] [SPEAKER_04]: You're totally right. It's like, yeah, the good actors, you know, oh, yes, you know,

[00:19:07] [SPEAKER_04]: there happened to be some watermark in the image I uploaded from a commercial image generator.

[00:19:11] [SPEAKER_04]: But are people as they scroll through their feet, even actually going to care about that?

[00:19:15] [SPEAKER_04]: And certainly bad actors are actively going to try and avoid, you know, either removing the

[00:19:21] [SPEAKER_04]: watermark or using tech that doesn't have it. And that's on the visual side. I mean,

[00:19:25] [SPEAKER_04]: I'm even curious sort of about like fake engagements and bots have been a thing for,

[00:19:30] [SPEAKER_04]: for at least a half decade plus now. This is a well-publicized strategy from 2016, right?

[00:19:37] [SPEAKER_04]: Russia, so discord on social media. And their strategy was almost like, hey, let's just start

[00:19:43] [SPEAKER_04]: starring, you know, these sort of very polarizing issues. And, you know, you emulate the kind of

[00:19:48] [SPEAKER_04]: rage bait that then drives engagement and it's a total mess. And do you think there's a strategy for

[00:19:54] [SPEAKER_04]: addressing sort of AI-generated bot forms that can start impersonating voters and sort of add

[00:19:59] [SPEAKER_01]: themselves to this public discourse? Yeah, I mean, that is a massive problem. It's actually very

[00:20:03] [SPEAKER_01]: interesting. Open AI released its first threat reported the end of May. And what it showed actually

[00:20:09] [SPEAKER_01]: was that for an actors, still are trying to figure out this technology like they still are not

[00:20:14] [SPEAKER_01]: sure how useful it is, you know what I mean? Which I thought was very funny. I was like, oh,

[00:20:18] [SPEAKER_01]: I guess we're all confused. But we definitely do see instances of them sort of really big strategy

[00:20:24] [SPEAKER_01]: that we see in foreign influence operations is they will link to websites that are meant to look like

[00:20:32] [SPEAKER_01]: a legitimate information source. And then they will use generative AI to populate articles that

[00:20:38] [SPEAKER_01]: reinforce certain political views, those articles then get shared on social platforms. So it's

[00:20:43] [SPEAKER_01]: not necessarily the bots or the content themselves, those sometimes are also AI-generated.

[00:20:49] [SPEAKER_01]: It's creating these websites that are populated with chatchip-t-style AI bullshit articles

[00:20:56] [SPEAKER_01]: and then sharing those on social platforms. But we're definitely seeing foreign influence

[00:21:00] [SPEAKER_01]: operations experiment with this. And again, I think you know, it's one of those things like

[00:21:06] [SPEAKER_01]: they're still testing it out too and seeing what's most effective. After the break, I asked

[00:21:14] [SPEAKER_04]: Victoria about how AI's reshaping elections around the world. In ways you might not expect.

[00:21:29] [SPEAKER_04]: So I want to transition to global elections because as you mentioned, the global angle here is very

[00:21:34] [SPEAKER_04]: interesting because oftentimes we're seeing the initial instantiation of these technologies being used

[00:21:40] [SPEAKER_04]: for good and bad overseas before it makes its way back here. It feels new to us in the United

[00:21:46] [SPEAKER_04]: States and maybe even in Europe, but the rest of the world's like, yeah, we've been dealing with this.

[00:21:52] [SPEAKER_04]: So I've got friends in Pakistan who've been what's happening, they're like, yeah,

[00:21:55] [SPEAKER_04]: defect politicians, you know, new voice recording, so common. There's like another one

[00:22:00] [SPEAKER_04]: floats around on what'sapp groups almost every week. So how has AI normalized this kind of content

[00:22:06] [SPEAKER_01]: in Pakistan? South Asia, particularly India, Pakistan places with really high concentration of like

[00:22:14] [SPEAKER_01]: highly educated people and tech talent. That's where we're seeing a lot of this. And I think

[00:22:20] [SPEAKER_01]: you know, in Pakistan there's, you know, the sort of legitimate use of this as we mentioned with

[00:22:25] [SPEAKER_01]: Immoron Khan. And then during the elections, there were also deepfakes of local politicians telling

[00:22:31] [SPEAKER_01]: people not to vote, telling people that they were dropping out of the race. And I think one of

[00:22:37] [SPEAKER_01]: the big things, particularly from the global south is audio messages are particularly common,

[00:22:44] [SPEAKER_01]: especially when people use WhatsApp. So there's a ton of instances of audio deepfakes and they

[00:22:51] [SPEAKER_01]: are very difficult to detect because you're not going to have the same signals that you would have with

[00:23:01] [SPEAKER_01]: a really one medium. And a lot of times it circulates on platforms like WhatsApp, not publicly

[00:23:08] [SPEAKER_01]: on social media, like ex or Facebook. It's circulating in these close communities on encrypted platforms.

[00:23:14] [SPEAKER_01]: It's incredibly difficult to detect and harder to debunk. India and Pakistan, they both have

[00:23:19] [SPEAKER_01]: a lot of really fabulous tech talent. And we see a lot of companies coming up to service that market.

[00:23:26] [SPEAKER_01]: But in general, most of the AI that tools that we're seeing and the tools to detect AI

[00:23:34] [SPEAKER_01]: generated content are trained on and built for data from the Western market. When we're looking

[00:23:39] [SPEAKER_01]: at markets in the global south, people are recording stuff on phones that are not as advanced

[00:23:46] [SPEAKER_01]: as like an iPhone, meaning that the baseline quality of the content might be lower. And that makes

[00:23:52] [SPEAKER_01]: so much harder for these detection tools to flag if something's fake. False positives and false

[00:23:58] [SPEAKER_01]: negatives are so much more common when stuff is coming off of these lower quality phones.

[00:24:04] [SPEAKER_01]: Secondly, when you're working with markets that speak English in a non-Western way. So like

[00:24:10] [SPEAKER_01]: pigeon forms of English, accent it English. There's more likely to have content that's flagged

[00:24:16] [SPEAKER_01]: deep fake. Like AI notoriously bad when it comes to non-white people. And so I think when we're talking

[00:24:24] [SPEAKER_01]: about places I've pocketed on in the global south, there is a lot of interest in these technologies

[00:24:29] [SPEAKER_01]: there. But also the ways in which they are used and the ways in which they can or cannot be

[00:24:35] [SPEAKER_01]: detected is totally different. And these are all massive problems where like this kind of

[00:24:46] [SPEAKER_01]: ones that frankly, these companies, they're barely prioritizing working on issues here. You know?

[00:24:52] [SPEAKER_04]: Totally. I'm curious with all of these instances, is it actually changing how people ended up

[00:24:59] [SPEAKER_01]: voting or like change voter turnout? So I think that really depends on where you look at. A really

[00:25:06] [SPEAKER_01]: great example came from actually a wired story that I didn't write. It was written by friend of

[00:25:10] [SPEAKER_01]: Nilesh Christopher and his reporting partner Varsha Bansal out of the Indian elections. They're

[00:25:17] [SPEAKER_01]: at the time of AI companies that are coming up in India now servicing the Indian market. And obviously

[00:25:22] [SPEAKER_01]: India is a massively diverse place in terms of language, religion, culture, etc. And they found

[00:25:27] [SPEAKER_01]: that in the lead-up to the Indian elections, a lot of local politicians were employing these AI

[00:25:34] [SPEAKER_01]: companies to create deepfakes of themselves similar to the Joe Biden Robo call, but they were authorizing

[00:25:39] [SPEAKER_01]: it and they were calling people's numbers, the numbers of their constituents. And you know in the

[00:25:43] [SPEAKER_01]: US we might consider that kind of thing really annoying and invasive. But what Nilesh found was

[00:25:48] [SPEAKER_01]: that actually people felt that sort of personalized outreach, even though they knew it was

[00:25:53] [SPEAKER_01]: an AI tool, even though they knew it was automated. They felt very seen by that and they felt very

[00:25:57] [SPEAKER_01]: considered by that. And it actually did make them want to vote for a particular candidate to feel

[00:26:02] [SPEAKER_01]: that they were getting personalized attention through the use of these AI tools. So in that way,

[00:26:09] [SPEAKER_01]: you know, it can be extraordinarily powerful. In Indonesia, for instance, during the campaign,

[00:26:15] [SPEAKER_01]: probably Osobiyantho who ended up winning the election, the former general under the Suharto

[00:26:21] [SPEAKER_01]: dictatorship. They used a majority to create a avatar of him looking very cute, friendly,

[00:26:27] [SPEAKER_01]: grandpa sort of vibes and they shared those all over TikTok and it got 19 billion views and it

[00:26:33] [SPEAKER_01]: definitely helped make him popular with young people who don't remember the Suharto dictatorship.

[00:26:40] [SPEAKER_01]: They didn't live through it and so they were susceptible to this sort of softer image of him

[00:26:46] [SPEAKER_01]: and that was totally authorized by the campaign. You know, there are ways in which this technology

[00:26:51] [SPEAKER_01]: can be used for campaigning and image management and all this stuff and that can really

[00:26:56] [SPEAKER_01]: maybe affect how people perceive a particular candidate. Absolutely. I mean, it's interesting

[00:27:01] [SPEAKER_04]: the indie example that you also brought up earlier. There was a company called Refraise AI

[00:27:07] [SPEAKER_04]: that would characon one of the biggest Hollywood actors. He did like a Cadbury chocolate advertising

[00:27:13] [SPEAKER_04]: campaign where the call to action like purchase the chocolate was like the local confectionary store

[00:27:19] [SPEAKER_04]: across all these regions. It's like, hey, if you're going to go buy this stuff, go buy it from here.

[00:27:24] [SPEAKER_04]: And now it's interesting to see this being applied in a political context. I think I could make

[00:27:28] [SPEAKER_04]: huge difference. It feels like personalized outreach and then as far as distribution goes,

[00:27:34] [SPEAKER_04]: there's the one to one stuff that we're talking about. There's the TikTok example maybe on

[00:27:38] [SPEAKER_04]: the other end that you mentioned where it's like, yeah, if these videos go viral, of course,

[00:27:43] [SPEAKER_04]: that's going to make a dent at least like put you top of mind for a lot of presumably,

[00:27:47] [SPEAKER_04]: you know, in this case, young voters, the stuff in the middle going back to WhatsApp seems more

[00:27:53] [SPEAKER_04]: concerning because it feels like harder to moderate and you know, stop the spread. Like I

[00:27:58] [SPEAKER_04]: recall a few years ago in India around Kashmir. There was a little bit of instability and sort

[00:28:03] [SPEAKER_04]: of like the internet got shut off, but then also WhatsApp basically around that time,

[00:28:10] [SPEAKER_04]: instituted, hey, there's only a limited number of people we're going to let you forward a message to.

[00:28:15] [SPEAKER_04]: And that the idea is like, that's one way to sort of cap the spread because, you know,

[00:28:22] [SPEAKER_04]: post a, you know, stop the spread of misinformation and by the time the analysts go and send

[00:28:28] [SPEAKER_04]: it to like, I don't know, experts in Europe to go, you know, assess the defect. Like the damage has

[00:28:33] [SPEAKER_01]: been done, right? Yeah. Well, I think, you know, again, if everyone was sort of agreeing like,

[00:28:38] [SPEAKER_01]: hey, we're going to watermark or we're going to use a hashing tool for it. So like that's how they

[00:28:44] [SPEAKER_01]: deal with encrypted messages. That's how they deal with like child sexual abuse materials hashing.

[00:28:50] [SPEAKER_01]: So a image will be hashed. It'll be entered into a database. It'll have sort of the specific

[00:28:55] [SPEAKER_01]: signature. And if you try and send it before it even leaves your phone to sort of go into that

[00:29:02] [SPEAKER_01]: space of being encrypted, it'll be sort of like caught by the platform and you will be unable to

[00:29:08] [SPEAKER_01]: forward it because you're looking it up against this data base exactly because it'll it'll be hashed

[00:29:12] [SPEAKER_01]: again, that sort of digital signature that hash will sort of ensure that. So maybe there would

[00:29:17] [SPEAKER_01]: be a way to say we're going to hash everything created by AI anything that, you know, gets forwarded

[00:29:24] [SPEAKER_01]: that that can be sort of checked against this hash database auto label as AI. There might be a

[00:29:29] [SPEAKER_01]: way to do that if everybody agreed and invested in that. But that's an immense amount of time

[00:29:34] [SPEAKER_01]: technology. Like, and that's the thing that I think people don't really think about when we're talking

[00:29:37] [SPEAKER_01]: about all this creation of AI generated content is now you're requiring the other side of it, which is

[00:29:42] [SPEAKER_01]: detection. You know, now you've created a whole other industry that's also about detection and that

[00:29:48] [SPEAKER_01]: also requires an immense amount of technology and investment and time and money to scale up

[00:29:55] [SPEAKER_04]: to respond to this problem. Totally. I really want to talk to you about AI Steve, you know,

[00:30:00] [SPEAKER_04]: my mind was really blown and I think it's super, it's a super interesting story. So why don't

[00:30:05] [SPEAKER_01]: you start by introducing AI Steve? Yeah, so he was a British political candidate who stood for

[00:30:11] [SPEAKER_01]: Parliament from the town of Brighton. And AI Steve was literally the candidate. He's the digital

[00:30:19] [SPEAKER_01]: avatar of actual Steve, a real man named Steve who was running for office and sort of the way

[00:30:25] [SPEAKER_01]: the campaign described it to me was actual Steve would be the physical embodiment of AI Steve.

[00:30:31] [SPEAKER_01]: He would be in Parliament doing the negotiations talking to people but all of his decisions would

[00:30:37] [SPEAKER_01]: be informed by AI Steve. And AI Steve was a sort of in that similar way, an AI avatar who could

[00:30:43] [SPEAKER_01]: respond to constituents and collect their their questions and the point of having this AI model was

[00:30:49] [SPEAKER_01]: that constituents could say here's what we care about, here's how we want you to vote on things,

[00:30:54] [SPEAKER_01]: here are the issues that matter to us and then real Steve's job was to go to Parliament

[00:30:57] [SPEAKER_01]: and do the actions dictated by AI Steve and AI Steve did not win. But in principle, I think that's

[00:31:06] [SPEAKER_01]: a really interesting idea and actually the campaign told me that the two things that people are

[00:31:10] [SPEAKER_01]: most interested in when they first launched it were the conflict in Gaza and trash collection.

[00:31:16] [SPEAKER_01]: But you know, I think even then that has sort of its own limitations right because members of Parliament

[00:31:21] [SPEAKER_01]: members of Congress get classified briefings all the time and are making decisions based on information

[00:31:27] [SPEAKER_01]: that the general public may not have and so there would still be a negotiation you need to have

[00:31:32] [SPEAKER_01]: with that sort of campaign commitment to make all decisions in line with what the AI has been

[00:31:38] [SPEAKER_01]: able to collect from constituents because the reality is the AI is not going to the classified

[00:31:44] [SPEAKER_01]: national security briefing. The AI is not being given special documentation and numbers

[00:31:51] [SPEAKER_01]: in the way that actual members of government are. And so there probably is a way for AI technology

[00:31:58] [SPEAKER_01]: to be incorporated into a sense of good governance but I think that's not what we're seeing

[00:32:04] [SPEAKER_01]: prioritized in what's being built and that's certainly not necessarily I think possible when

[00:32:10] [SPEAKER_04]: the AI itself is the candidate. All right, well you certainly answered my question of like how

[00:32:15] [SPEAKER_04]: would this even work? And so it's almost like this is like a digital twin or proxy in the public domain

[00:32:21] [SPEAKER_04]: for a politician and really it's a way of sort of getting a pulse, you know, over your constituents

[00:32:27] [SPEAKER_04]: like a pulse of the nation if you will but this isn't the only example right there been other

[00:32:32] [SPEAKER_04]: examples of AI candidates around the world. Can you tell me the story about the Belarus elections

[00:32:37] [SPEAKER_01]: earlier this year? Belarus is a dictatorship and one might argue a sort of proxy country of Russia.

[00:32:48] [SPEAKER_01]: The Russian military currently is staged out of Belarus for its word Ukraine and it's dictatorship,

[00:32:54] [SPEAKER_01]: the Lukashenko dictatorship is one of the last remaining ones in Europe and there has been

[00:33:01] [SPEAKER_01]: incredible crackdowns on dissidents. One of the dissidents who is an exile created an AI candidate

[00:33:07] [SPEAKER_01]: called Gas Gas Bidar to run in the country's parliamentary elections and obviously this candidate

[00:33:13] [SPEAKER_01]: did not win the elections in Belarus earlier this year were widely regarded to be unfree

[00:33:18] [SPEAKER_01]: but because dissidents is so criminalized in Belarus and because many of the people who might have

[00:33:25] [SPEAKER_01]: actually stood for elections are in exile this was sort of a way of using AI to represent those people,

[00:33:34] [SPEAKER_01]: you know he can't be arrested he cannot be subject to the ways in which the government has used

[00:33:40] [SPEAKER_01]: force to crack down on dissidents in the past. So I think this was actually a very clever use of AI

[00:33:46] [SPEAKER_01]: just like you know we saw people in the Middle East use social media to really power the

[00:33:52] [SPEAKER_01]: Arab Spring. Like there are good generative democratic creative uses of these technologies and people

[00:34:01] [SPEAKER_01]: will find those but that doesn't mean that's always like what they're geared towards or the most

[00:34:05] [SPEAKER_04]: common use of them. Totally right I mean this is a perfect example of AI being effective if you

[00:34:10] [SPEAKER_04]: are a political dissident right like I was blown away by that one quote in that article which is

[00:34:15] [SPEAKER_04]: basically like yeah this person can't be imprisoned because he's just code and I was like wow

[00:34:21] [SPEAKER_04]: that's one way to give a voice to the voiceless. So we've gone through a lot of aspects of the current

[00:34:27] [SPEAKER_04]: state of you know let's just call it the state of AI elections in the US and some lessons

[00:34:33] [SPEAKER_04]: learned from other countries. I think at this point I just have to take a step back and ask like

[00:34:39] [SPEAKER_04]: how concerned should we really be about the ways that AI could affect the 2024 election?

[00:34:45] [SPEAKER_01]: In the US I don't know yet. I think there are a lot of things at play and I think one of the

[00:34:52] [SPEAKER_01]: things we didn't touch on which is important to note is that social platforms have rolled back

[00:34:57] [SPEAKER_01]: their investment in trust and safety and trust and safety are the people and teams who make

[00:35:02] [SPEAKER_01]: sure that hate speech, disinformation all that stuff stays off the platform and then on top of that

[00:35:07] [SPEAKER_01]: we're sort of adding this extra layer. So I think we are certainly going to see a ton of AI

[00:35:15] [SPEAKER_01]: bullshit on all the platforms. I think we're going to see more of it but I think the real thing is

[00:35:22] [SPEAKER_01]: personally I am less concerned about the elections themselves and more about what happens after

[00:35:31] [SPEAKER_01]: we are in a moment where there is less and less and less trust than there's ever been. We have

[00:35:38] [SPEAKER_01]: now Maskuhon's ex he has already started seeding narratives around you know illegal immigrants

[00:35:47] [SPEAKER_01]: voting things that could very easily sort of form the intellectual foundation to question a

[00:35:53] [SPEAKER_01]: democratifictory. Those are old problems. The platforms still haven't solved and I think we'll see

[00:35:59] [SPEAKER_01]: AI play a role in those whether that's through disinformation campaigns, whether that's through

[00:36:02] [SPEAKER_01]: AI generated media, whether that's through for instance, Grock AI recently returned answers saying

[00:36:09] [SPEAKER_01]: that Kamala Harris had missed the deadline to register to be on the ballot in nine states and the

[00:36:15] [SPEAKER_01]: secretaries general of five of those states had to write to ex and say your AI chatbot is spitting out

[00:36:21] [SPEAKER_01]: bad information. I think we'll see things like that but I think the core issues underlying this which is

[00:36:27] [SPEAKER_01]: lack of investment interest and safety, a lack of investment in thinking about the implications

[00:36:33] [SPEAKER_01]: of these technologies and the safeguards necessary and real lack of trust and sense of reality

[00:36:40] [SPEAKER_01]: in terms of the shared world we're living in and trust in institutions and systems. Those

[00:36:45] [SPEAKER_01]: underpin this question more deeply than any sort of variation in the technology could.

[00:36:51] [SPEAKER_04]: That is a very profound and nuanced answer to end on. Vittoria, thanks for being on the show.

[00:36:57] [SPEAKER_04]: Thank you. After talking with Vittoria, I had a couple of ideas about how we can prepare for the next

[00:37:06] [SPEAKER_04]: few months of the election season. The first thing I'm thinking about is actually something we

[00:37:10] [SPEAKER_04]: covered in the very first episode of this podcast. When I spoke with Sam Gregory from Witness,

[00:37:16] [SPEAKER_04]: he recommended a good exercise in media literacy that I think still applies beautifully.

[00:37:21] [SPEAKER_04]: Sift. As a refresher that stands for stop, investigate the source, find alternative coverage and

[00:37:27] [SPEAKER_04]: trace down the original content that you think might not be legit.

[00:37:31] [SPEAKER_04]: Especially since the political landscape is changing at this unprecedented pace,

[00:37:36] [SPEAKER_04]: it's very important to always hit those four steps and have a wide range of trusted sources

[00:37:41] [SPEAKER_04]: at your disposal. But the second thing is a little harder to break down into specific steps

[00:37:47] [SPEAKER_04]: because it has to do with this concept that Vittoria brought up. The liar is dividend.

[00:37:52] [SPEAKER_04]: That's the idea that in a world where anything can be falsified, nothing is real

[00:37:57] [SPEAKER_04]: and so it kind of doesn't matter if you can identify disinformation out in the wild.

[00:38:01] [SPEAKER_04]: Its message can still very much influence people, even when they know it's not true.

[00:38:07] [SPEAKER_04]: It's why memes can be such powerful weapons for political messaging. They can bypass

[00:38:11] [SPEAKER_04]: that part of our brains that question if something is real and go straight to our emotions.

[00:38:18] [SPEAKER_04]: When I saw those AI-generated memes that I mentioned at the top of the episode,

[00:38:21] [SPEAKER_04]: I realized that introducing this tool months before election day was opening the door

[00:38:26] [SPEAKER_04]: for full-blown, mimetic warfare, which means we're all going to have to come to terms with

[00:38:31] [SPEAKER_04]: the fact that none of us are exempt from being influenced by content. AI-generated or not.

[00:38:37] [SPEAKER_04]: It's going to be hard but we've got to modify our information diets.

[00:38:42] [SPEAKER_04]: That means adjusting how we let content impact our perceptions of reality.

[00:38:47] [SPEAKER_04]: There are some really exciting possibilities for AI technology to make politics more accessible,

[00:38:52] [SPEAKER_04]: more representative and maybe even more revolutionary. But ahead of our first AI election,

[00:38:58] [SPEAKER_04]: it'll be up to us to learn from the rest of the world and charter own path towards the future.

[00:39:07] [SPEAKER_04]: The Teddy I show is a part of the TED audio collective and is produced by TED with Cosmic Standard.

[00:39:13] [SPEAKER_04]: Our producers are Ben Montoya and Alex Higgins. Our editors are Ben Bencheng and

[00:39:18] [SPEAKER_04]: Alejandra Salazar. Our showrunner is Ivana Tucker and our engineer is Asia Polar Simpson.

[00:39:24] [SPEAKER_04]: Our technical director is Jacob Winick and our executive producer is Eliza Smith.

[00:39:30] [SPEAKER_04]: Our researcher and fact checker is Christian Aparta and I'm your host, Belabel Sadoo. See y'all in the next one.