How AI deepfakes are upending political advertising
AI-generated political ads flooded TV and social media this midterm cycle. Roughly half of them used likenesses of real candidates. Most didn’t disclose they used AI. What does this mean for the future of truth in elections?
Guests
Erika Franklin Fowler, co-director of the Wesleyan Media Project, which tracks and analyzes political advertising across the U.S. Professor of government at Wesleyan University.
Eric Wilson, executive director of the Center for Campaign Innovation, which describes itself as “the right’s research and development lab.”
Maya Hutchinson, founder and CEO of BattlegroundAI, which provides tools for campaigns to create AI-generated content. Former democratic strategist.
Samuel Woolley, William S. Dietrich II Endowed Chair in Disinformation Studies at the University of Pittsburgh. Associate professor in the Department of Communication.
Also Featured
Hany Farid, Donald Knuth Professor of Computation and Just Communities and professor of Computer Science at Dartmouth University.
____
Transcript of Full Broadcast
The version of our broadcast available at the top of this page and via podcast apps is a condensed version of the full show. You can listen to the full, unedited broadcast here:
Part I
(AI AD OF JOSH SHAPIRO PLAYS)
MEGHNA CHAKRABARTI: That is not, in fact, Pennsylvania Governor Josh Shapiro. Shapiro never said any of that. This is from an attack ad by Shapiro’s opponent, Republican Stacey Garrity, in this year’s midterm elections.
It was generated using AI. As you heard, the AI-generated voice of the AI-generated Shapiro say. By the way, this is what Josh Shapiro actually sounds like.
SHAPIRO: Look, the results are still rolling in, but the picture is starting to come into focus. This is a good night for our commonwealth.
CHAKRABARTI: That’s the governor after he won a resounding re-election just this week.
And then there’s this ad from Michigan. It features an AI likeness of Democrat and now Senator-elect Abdul El-Sayed, dressed as a Greek warrior hiding inside a Trojan horse labeled “Definitely not communism.”
ABDUL EL-SAYED: They’ll never see us coming. They really think we’re just Democrats.
CHAKRABARTI: The ad also includes AI-generated versions of Alexandria Ocasio-Cortez, Bernie Sanders, and Zohran Mamdani. It was created by Republican Mike Rogers’ campaign during his failed run against El-Sayed for Senate. In New York, a deepfake of incumbent Democratic Governor Kathy Hochul appeared in an ad by Hochul’s Republican opponent, Bruce Blakeman.
The ad starts off with the words, “We asked AI what a truthful Kathy Hochul ad would look like.” And then it shows an AI-generated version of Kathy Hochul walking around New York and talking.
(AD PLAYS)
We provided luxury accommodations for illegal migrants in beautiful hotels just like this, and we raised taxes by $8 billion for the programs that matter, like free healthcare for sex workers and lucrative no-bid contracts to the politically connected.
Thanks to our hard work, we’re leading the nation in giving foreign companies energy contracts, which means your energy bills go up.
CHAKRABARTI: And in Texas, an ad by Republican Senate candidate Ken Paxton featured an AI-generated likeness of now Senator-elect James Talarico holding a Bible.
(AD PLAYS)
In the beginning, God was non-binary, and he wasn’t a Christian. And God said, “Let there be six genders,” and there were. Then God so loved the world that he gave us his only son, who was a radical feminist and never asked people to worship him. Look, I always think of myself as a Christian who hates Christianity.
CHAKRABARTI: The attack ad removes context from some things that now Senator-elect Talarico has actually said in speeches, sermons, and podcast appearances over the years, and it splices those together. And again, by the way, here is what Talarico actually sounds like.
TALARICO: Every corrupt defender of this broken political system in this state, in this country, power belongs to the people.
CHAKRABARTI: Okay. Now back to that ad from the Paxton campaign and the things that Talarico actually did say. He did say, “God is non-binary,” during a Texas House floor speech in October of 2021. He has since explained it was meant to be, quote, “deliberately provocative” and that he believes, quote, “You can’t use human categories to define God.”
Talarico has not said that there are six genders, but he did talk about biological sex during a Texas House Public Education Committee hearing in 2021. Talarico said, quote, “Modern science obviously recognizes that there are many more than two biological sexes. In fact, there are six,” which Talarico said surprised him, too.
He was talking about scientifically recognized intersex chromosomal variations, such as Klinefelter syndrome, where a male has two X chromosomes and one Y, or Turner syndrome, where a female has only one X chromosome instead of two. Joining us now is Erika Franklin Fowler. She’s co-director of the Wesleyan Media Project, which tracks and analyzes political advertising across the U.S.
She’s also a professor of government at Wesleyan University. Professor Fowler, welcome to On Point.
ERIKA FRANKLIN FOWLER: Thanks for having me.
CHAKRABARTI: I wanted to actually give the context of what Senator-elect Talarico had actually said because that’s 100% what’s devoid from all political ads, but especially these AI ads.
First of all, tell me how much or how many of such ads were out there during this midterm cycle?
FOWLER: Yeah. We’re still counting. So as you heard my co-director, Travis Ridout, say at the top of the hour, in our last count, we released 289 unique ads to the tune of $187 million on AI ads, and that number will grow.
It’s also certainly an undercount.
CHAKRABARTI: Oh, an undercount. Okay. But is it still a relatively small number given the vast overall amount of political ads?
FOWLER: Given the billions of dollars that were spent in this election … you do want to put that in context. 187 million is a lot of dollars, but it’s still a small fraction of the total that was spent.
CHAKRABARTI: Okay. So deepfakes have been around for a while, but I’ve been hearing people say that this midterm cycle is perhaps the first one where there were so many in different markets and primarily on social media, right? Can we say this is like the first election which deepfake ads made inroads into people’s social media feeds for politics?
FOWLER: Yeah, and I think the even more striking thing to us is the inroads on television as well. In some ways you expect to see some activity on social media. But I think the fact that you had so many television airings as well that had deepfakes in them is quite stunning.
CHAKRABARTI: And there are no regulations against that for broadcast television?
FOWLER: There’s no federal regulation of AI in advertising, so what we have instead is a patchwork of state laws that apply to mostly state candidates across the country. And so some of the ways in which we see AI in advertising is because the ad itself will carry a disclaimer that was required by a state.
CHAKRABARTI: Oh, okay. Oh my gosh, I wanna come back to this in a second, Professor Fowler. But regarding the AI ads that Wesleyan, the Wesleyan Media Project has been able to catalog so far, were there different themes or were there patterns of themes?
FOWLER: Yeah, they, I would say that the ads go into roughly three types of buckets.
There’s the kind of AI use that is really subtle and that we would even have a hard time detecting, and the only way we know that it’s been used is because there was a disclaimer on the ad. I’m not sure that type of use is particularly concerning. I think there’s a lot of, there can be benefits to the use of AI for campaigns.
It can make producing ads a lot more economical, and if they’re just using it to change or enhance lighting or visuals or color, things along those lines, I’m not sure that it’s a big deal. But then there are two different types that we’ve identified of deepfakes, and I think it’s the hyper-realistic deepfake that we should be particularly concerned about, because it’s not clear to voters that the information is, that a candidate hasn’t actually said something that they’re depicted as saying or doing something that they’re depicted as doing. So that’s, I think, the most concerning category. And then there’s also this third category of also deepfakes, but the context of the ad almost makes it clear that it’s not real.
In other words, the ad is satirical or sort of parody in format, if that makes sense.
CHAKRABARTI: Oh, it makes total sense because that Talarico ad that we played earlier, or the anti-Talarico ad that we played, the visual is James Talarico’s AI likeness sitting in a bedroom drawing in a Bible with crayons.
So I mean, I think that’s pretty obvious that it is satirical. But you actually made a important point, Professor Fowler, which I want to underscore because AI can be used for a lot. You said it could be used for taking real footage and maybe editing it together more quickly and splicing it together or creating the ad.
But it’s using actual footage or the person’s actual voice. Those are not deepfakes. It’s just AI as a tool to more rapidly create acceptable political ads, correct?
FOWLER: Correct.
CHAKRABARTI: Okay, good.
FOWLER: And in many ways, those ads don’t look at all different to the ones that we’ve seen in the past, if that makes sense.
CHAKRABARTI: Yeah, they wouldn’t, right? Because it’s just, it’s editing but faster with AI. Exactly. So your point is really well taken on that. But the one that seems like you’re concerned about are these deepfake ads. But I do wonder, the aesthetic of them, at least the ones that I’ve been able to see, they do have that sort of cheesy, hyper-realistic, like this was definitely AI look to them.
Are there ones that you have seen that aren’t that obviously, as the kids say, just slop?
FOWLER: I do think there are some. And I do also think that there’s a generational difference in who can identify them. My undergrads here at Wesleyan that are coding our ads are often the best detectors, because they can point to exactly the point that that they think the AI is. And so I think it really part of the concern, I think, is on the part of the viewers. If you are a user of AI or if you’ve seen a lot of AI, you’re much more readily able to detect it and to notice it.
But I’m not sure that’s true for the entire electorate.
CHAKRABARTI: Okay. A couple other quick questions before I have to let you go, professor. … Did one party make more of these AI ads than another, or was it equally distributed?
FOWLER: Yes. Of the kind that and particularly of the deep fakes that are, that sort of populate our database, Republicans are definitely the ones who created and are distributing the ads much more so than Democrats.
CHAKRABARTI: Okay. And do we have a way to measure the reach or the impressions of how many people saw these ads?
FOWLER: Yeah, that part is still coming, and one of the challenges there is because we’re tracking across platforms, how to measure reach is actually quite difficult. So we’re- we’ve been tracking the spending, and we’ll do comparisons across platforms in spending, but the impression part will be harder.
CHAKRABARTI: Okay. We have about 30 seconds left with you, professor. You are a professor of both of government and in this Wesleyan Media Project. Is this a new era of political advertising we’re in? Is this kind of one of those moments where we’ll look back and say, “Wow, that was the year that everything changed”?
FOWLER: I think a lot depends on what happens in the federal regulation space, and also potentially among the states. I think if our legislators decide that they want to make changes, we might cede to something different in future cycles, but I do think that the use of AI broadly and the use of AI in video is likely here to stay.
Part II
CHAKRABARTI: I’d like to bring in two folks who are in this world from opposite sides of the aisle. Maya Hutchinson is a former Democratic strategist. She’s founder and CEO of Battleground AI, which provides tools for campaigns to create AI-generated content, and she joins us from Denver, Colorado. Maya Hutchinson, welcome to On Point.
MAYA HUTCHINSON: Hi, Meghna. Thanks for having me.
CHAKRABARTI: It’s great to have you. Eric Wilson worked on several Republican political campaigns, including Marco Rubio’s 2016 run for president. He’s now executive director of the Center for Campaign Innovation, which describes itself as, quote, “The right’s research and development lab,” and he’s joining us from Washington, D.C.
Eric, welcome to you.
ERIC WILSON: Thank you, Meghna. Glad to be here.
CHAKRABARTI: It’s great to have you. So let me just ask both of you, and Eric, I’ll kick off with you. Just give me your broad sense as to the import of AI in political advertising and how you think or how you assess how Republican candidates use them this time around in the midterms.
WILSON: I’m really excited to see more campaigns using AI because it has the opportunity to do several things, including level the playing field, bring down the costs of producing ads so we can create more ads, maybe make those ads more creative, more pleasant to watch, for those of you who might be in battleground states and enjoying your TV time now.
But the AI is another tool that allows us to be more effective in our campaigning and really help get us past, who has the most funds and who has the best ideas. I
CHAKRABARTI: s that what you meant by leveling the playing field, the funding discrepancy?
HUTCHINSON: Absolutely. You’re allowed … you’re now able to do things with your messaging that previously would’ve been restricted to wealthy candidates or was otherwise cost-prohibitive or time-prohibitive for campaigns to do. And I’m jealous of folks who get to work on campaigns right now with all of the possibilities that happen with AI.
CHAKRABARTI: To be fair, Republican candidates who didn’t cross the finish line in these midterms this week, many of them were really strongly funded, more so than their Democratic opponents. So funds in and of themselves aren’t gonna really make a difference in terms of who wins and who doesn’t.
But Maya, I want to ask you the same question that I asked Eric in terms of, how do you see Democrats having used AI ads and what do you think the advantages are?
HUTCHINSON: Yeah, absolutely. Democrats have also been adopting and implementing the technology. I started Battleground AI as a compliance-focused advertising technology so that we can really harness the power of this tool and use it to be more effective and to deliver messages and content in a more efficient and effective way, but also meet those regulations and requirements both from platforms, from state requirements and the regulations that already existed in this space and being able to do both of those things, and we can see that play out across the board.
And it’s not just one party versus the other, but we both need to understand how to effectively and safely use this technology.
CHAKRABARTI: Okay, effectively and safely, we’ll come back to that. But are there any particular ads that you saw, I don’t know if you worked on any during this campaign that you thought were particularly effective, Maya?
And then Eric, I’ll ask the same question of you, but go ahead, Maya.
HUTCHINSON: So there’s lots of effective use of this technology. I know you played a lot of examples where there were deepfakes involved or fake scenarios or things that were not truthful that were said. But the reality is that most of the time this technology is built into almost every video editing platform at this point.
We choose to build the guardrails so that you cannot create fake images of people using our technology. We chose that, and that’s a trade-off because folks can do that across a lot of other technologies now. But for the most part, it’s really in speeding up efficiency, clipping making content more rapidly available across multiple platforms to reach viewers and audiences no matter where they are, and campaigns need to do that quickly, they need to do that effectively, and AI is allowing them to do that.
CHAKRABARTI: Yeah. Okay, so Eric, let me ask you, are there any specific ads you can think of that you thought, “This was really good. It was effective, and we probably couldn’t have done it in the pre-AI era”?
WILSON: Yeah, the one that I keep going back to is an ad for the South Carolina governor’s race earlier this year where they recreated sort of a Family Feud-style program where they were educating voters about one of the candidates’ records, and it had a full studio audience, and contestants, and a host.
That just would not have been something that people would have been able to do without AI. And I think that’s the kind of creative, interesting storytelling that people would really be excited for in political advertising.
CHAKRABARTI: And in that ad, it didn’t feature deepfakes of any of the candidates.
It was just like artificial Family Feud participants?
WILSON: That’s correct, yep.
CHAKRABARTI: Okay, great. For both of you, I actually want to play some ads from your opposite party here. So let’s start off with a political attack ad that was launched against Georgia Democratic Senator Jon Ossoff, and it included AI deep fakes of Ossoff sitting in a beauty salon having his hair, nails, and makeup done as a female narrator criticizes him.
Here’s part of that ad, which was released by the Republican-aligned Senate Leadership Fund PAC.
Turns out that boy spends more on hair and makeup than any D.C. politician, including women. I know what I’m trying to cover up. Ossoff, he’s hiding something else, like voting to raise taxes on retirees making under $75,000.
Jon Ossoff talks pretty. His actions, ugly.
CHAKRABARTI: And again, the visual is of a deepfake of Jon Ossoff. So Maya Hutchinson, how would you assess that ad?
HUTCHINSON: So I think that’s fabrication. That’s replicating a candidate. We have building out regulations around this use case, and there’s disclosure requirements, and they’re not federal but many states have passed deepfake bans or disclosure of the use of AI.
While that is not the whole answer here, we need to make sure that we have trust and understanding in the content that voters are seeing, and that they can trust that the information is accurate. And that’s really what’s most important. Again, this is viewed by and approved by an organization.
That should be held accountable if they’re not providing accurate and truthful information, regardless of the technology. So we really need to focus on separating out the content and context of the ad from the technology that’s being used.
CHAKRABARTI: Got it. Okay. So Eric, your turn here.
WILSON: Yeah, no, I thought that was a great ad, right?
Because it took something that was otherwise very boring and dry and turned it into a story, and I think voters are smart enough to tell a cartoon from reality.
CHAKRABARTI: Maybe now, but the problem I think is that AI is just getting better and better, and I can easily imagine that in the next election cycle in two years, what if we see more ads where obviously the cartoonish aspect of the production is simply not there, and it’s an ad that shows a candidate saying and doing things that he or she literally never did, and it’s not satirical?
Do you, would you have a problem with that?
WILSON: The things that you’re worried about are already illegal. It has nothing to do with AI. You can’t slander candidates. You can’t deceive voters. You can’t suppress the vote. We don’t need additional laws and regulations just because AI came onto the scene.
CHAKRABARTI: Maya, what do you think about that?
HUTCHINSON: I think states are already passing regulation. We see almost 30 states or more are having some sort of AI regulation already passed. So states are leading on this. They are creating regulation. We also do need to set a federal regulatory framework and standard at least as a floor here.
I think that’s a critical part of trusting in new technology. This is ultimately a trust question, and do voters trust the content that they see? And that is in part from regulation. I think as an organization work with the American Association of Political Consultants, we put together the meaningful disclosure framework for the industry to evaluate content that they’re creating and making sure that that is accurate, that they are telling the right information.
You can create attack ads. Those are not new to this industry. That has been around for a long time. The information and the tool that we use to do that just needs to be separated from that.
CHAKRABARTI: Okay, let me just ask you a follow-up because Eric said that, in my future scenario, maybe even present scenario, where you have video that is purely AI-generated of a candidate, and the candidate never said or did things in the video.
Eric says that using that is already illegal, do you understand the law in the same way, Maya?
HUTCHINSON: Yeah, so I think there you have to have truth in advertising. You have to be truthful in the content that you deliver and advertising is held to a higher standard, right? That type of content is often, could be posted organically.
The rules are not the same when it comes to advertising. There are higher standards around that and platform requirements as well as state laws.
CHAKRABARTI: Okay, organically you mean like user-generated content?
HUTCHINSON: Correct, yeah.
CHAKRABARIT: Okay. So not pol- political campaigns or PACs. Okay, we’re gonna come back to this in a second, but Eric, I did wanna actually get your take on a Democratic ad that uses AI.
So this is from, it’s not from this past election cycle. It’s from Josh Gottheimer’s primary campaign for governor of New Jersey last year. And again, this is from, this is a Democratic ad, and the ad shows an AI-generated Gottheimer in the boxing ring fighting President Donald Trump.
(AD PLAYS)
CHAKRABARTI: And again, the visual there is of an AI-generated Gottheimer in a boxing ring fighting an AI-generated President Trump.
Your assessment of that ad, Eric?
WILSON: Yeah, good ad. I’m not in the habit of complimenting Democrats, but I think voters are smart enough to realize that he didn’t actually do a boxing match with President Trump. I’m having a hard time understanding the concern here.
CHAKRABARTI: Okay. Great. This may be one of those few AI conversations that I’ve had on the show that perhaps we’re just making a tempest in a teapot. But I do have this question that I wanted to ask you about what Maya said regarding skirting the line between having things be just obvious but then daring to maybe mess with the truth.
Are you still, Eric satisfied with the idea that it’s so obvious that when people cross that line using AI that it’s already illegal?
WILSON: Look, campaigns are job interviews with voters. And you don’t get hired to do a job when you lie to the people who are hiring. And so there is already a clear guardrail of voters don’t want to be deceived.
They want accurate information, and campaigns that mislead voters or deceive them will pay the price. We already have that in place. We don’t need more regulation. We don’t need more laws. We don’t need to protect voters from themselves because I think they’re smart enough to understand what’s true and what’s not.
CHAKRABARTI: , point well taken. And in fact, in some of the ads that we played at the top for example, there was the deep fake of Pennsylvania Governor Josh Shapiro. Actually, in the middle of that ad, whoever made it, they said, the deepfake Josh Shapiro actually said, “I’m an AI.”
Would you actually be okay, though, if would you support regulation that would say in some way, shape, or form it, an ad has to identify when AI was used in it, even if it’s just a little watermark at the bottom of the screen?
WILSON: Yeah. I don’t support turning political ads into pharmaceutical advertisements that are just loaded up with something. We already have to put too much in there as it is, in my opinion. Some of the most over-regulated advertising in the country. But we’ve also seen and Maya pointed this out, that it’s really hard to make an ad today without using any sort of AI, and the way some of these regulations have been written is you basically would get an AI warning on every single ad that came out, and that doesn’t help voters understand, is this true or not?
All it says is that AI was involved in some level of production of this ad. It doesn’t inform them.
CHAKRABARTI: Yeah, so that’s a good point. And Maya, I’d love to hear you your thoughts on this because AI is so deeply embedded already, as you said, in the technologies, in media technologies that if you were to call for more regulation, what would it look like that would stop short of having every single ad people ever see have that AI disclaimer on it?
HUTCHINSON: Yeah, absolutely. I definitely agree with Eric. I think labeling everything as AI then doesn’t tell you anything about the content or the context of the advertising itself. That’s why we’re really, I’m really focused on meaningful disclosure. This is, again, separating that technology that folks can use which is really powerful from the content and the message that voters and viewers are seeing and hearing and telling them when, to your point, when something is satire or parody or a reenactment.
That happens in many cases. So if a creator is wanting to do that, they should disclose the meaningful context behind the information, not the technology that was used to create it. Because at this point and especially since we started doing this in 2024 election cycle, just the use of AI and the power of it has increased dramatically, and we need to be able to account for that.
CHAKRABARTI: Sorry to interrupt you, Maya, but can you give me a specific concrete example of what you mean by meaningful disclosure?
HUTCHINSON: Yeah. So from the meaningful disclosure framework again, that we worked on and Eric was a part of, too from the association, American Association of Political Consultants, is a good example of that.
When content in the ad display, walk through whether that is truthful information, whether you have evidence for the truth that you’re going to, that you’re going to put in that content of that advertisement. You highlighted a lot of places. If the truth in that ad from the Talarico attack ad if that content was actually truthful or not you should make that decision about how you’re going to make that ad.
If it’s not true and you’re fabricating that should not be an ad that’s run … from industry standard.
CHAKRABARTI: Sorry, we only have both of you for a couple more seconds, so forgive me, but Maya I’m a little confused as to how you add, like, all that content to a 30-second spot.
HUTCHINSON: Oh, I think from the get-go you decide if you’re having a truth in your advertisement and whether it should run or not.
I think it’s actually a lot easier than I’m explaining. If it’s not truthful, you shouldn’t run the ad.
Part III
CHAKRABARTI: We asked deepfake expert Hany Farid, are people really falling for these AI-generated campaign ads?
HANY FARID: Oh, 100%. Look, why do we still get emails about Nigerian princes and their multi-million-dollar inheritance?
They work. You don’t have to change the minds and the hearts and the thinking of millions of people. Think about the last national elections. What, a couple hundred thousand votes in a handful of states can move an entire election. You don’t have to move that much.
CHAKRABARTI: Farid is a professor of computer science at Dartmouth University and co-founder of Get Real Security, a digital forensics company.
FARID: I don’t think we are going to see the equivalent of a nuclear bomb of a deepfake. I don’t think we’re gonna see days, hours before an election, this catastrophic deepfake that just changes an election. It’s a caricature of the reality. The reality is this will be death by a thousand cuts, that what will happen is the slow, steady chipping away of trust.
CHAKRABARTI: Farid says social media companies could do a better job of keeping deepfake ads from spreading online. Congress could pass laws that regulate deepfake political ads. But he also thinks it’s very important that campaigns and others disclose when they have used AI.
FARID: I think of it like the labels in the grocery store on food.
You want to eat barbecue potato chips, it’s fine, but this is what you’re going to be eating. Informing consumers, particularly around political ads, but I would say more broadly. But look, Congress can just say, “No AI in political ads.” Maybe that’s too broad, but maybe that’s okay right now while we’re trying to figure out what this all is, because it is clear that you can do some pretty damaging things in a way that you couldn’t do a few years ago.
CHAKRABARTI: Farid also says that with or without regulation, whether we like it or not, this is just our new reality.
FARID: We as consumers are going to have to live with the reality that everything we are seeing is some combination of sloppy AI, convincing AI, and then real content. And then the answer there is stop getting your news and information from social media.
If you want to be lied to and you want to be deceived and defrauded, sure, by all means, spend more time on X and Facebook and Instagram and YouTube and TikTok. But if you want to be an informed citizen, get off of social media.
CHAKRABARTI: Hany Farid is professor of computer science at Dartmouth University and co-founder of GetReal Security.
I’d like to bring Samuel Woolley into the conversation now. He’s the William S. Dietrich II Endowed Chair in Disinformation Studies at the University of Pittsburgh, and he joins us from On Point station WHYY in Philadelphia. Professor Woolley, welcome to On Point.
SAMUEL WOOLLEY: Thanks so much for having me.
CHAKRABARTI: I actually think this is the 10 most important minutes of today’s conversation because I am quite obsessed with this question of how people can discern fiction from truth these days.
So where do you fall on the spectrum in terms of concerns regarding AI use in political advertising?
WOOLLEY: I’m deeply concerned. I think that AI interrupts the political discussion. It interrupts political communication. It’s intended oftentimes when it’s used in many of the examples that, that have been given on the show today to confuse voters.
It’s not intended to clarify the system oftentimes. It’s intended to make it unclear what’s actually happening, and that’s a problem for the democratic process.
CHAKRABARTI: How is that any different though than different attack ad strategies we’ve had for a generation or two?
WOOLLEY: That’s a really good question, and we’ve had attack ads.
We’re gonna continue to have attack ads, and there is, there are some similarities between attack ads and AI attack ads, but never before have you been able to actually put words in a candidate’s mouth and make it look like they’re actually saying and doing something. And beyond that, something that also bears saying is that the context matters.
So for instance, with the Massie ad that you just talked about, it shows him walking into a hotel with two other Democratic women. Context matters, verbal cues matter, non-verbal cues matter. It’s not just about what is actually said. It’s about what people are seeing, and it’s about introducing an idea to people that they’ve never thought of before.
CHAKRABARTI: Okay, so tell me more about that. Help dissect that ad more. So the idea there that is being planted is that Massie associates with these Democratic congresswomen, whether or not that’s true?
WOOLLEY: And it’s suggesting a sexual relationship between Massie and the Democratic congresswoman.
That’s what the subtext is. It doesn’t say it out loud but it’s using AI to show him with those women going into a hotel. And planting the seed in propaganda is the biggest thing that you can do. Planting the seed and then continuing to fertilize it. As Hany said earlier the idea is it’s death by a thousand cuts.
And so you plant the seed and then you keep on talking about it. It’s not about some monumental catastrophic deepfake that changes everything. It’s about changing reality and breaking trust.
CHAKRABARTI: Okay, so what you’re saying here, and hope this isn’t too crude, but it’s not that a voter would see this ad and be like, “Oh man, Thomas Massie is sleeping with AOC.”
It’s just planting this idea that maybe even if we took the sexual element out of it that, hey, maybe he … Or that’s just weird. I think is that enough actually in politics these days for people to be like, “Wow, that’s weird. That Massie guy, he might be weird”?
WOOLLEY: I think so, and if you study legacy propaganda like Soviet propaganda and then, more recent stuff with what my colleagues and I call computational propaganda, the idea isn’t often or always or even necessarily to change people’s mind about a candidate or to get them to vote a different way.
It’s to create some kind of confusion, some kind of, as you say, weirdness, some kind of apathy, and to oftentimes to get people to turn away from voting or to vote a different way just because they have a bad taste in their mouth.
CHAKRABARTI: I see. So a kind of an association that may not have been there before.
That is really interesting because I think one of the most effective, to use that term, ads from the last presidential election was the Trump campaign ad about Vice President Kamala Harris that ended with the line of, and this wasn’t even AI. This was just a narrator who said, Kamala Harris is for they, them.
Trump is for you. But that, would that be an example?
WOOLLEY: Yeah exactly. It’s just introducing some idea that is rooted in the culture wars at the end of the ad that may or may not be true, but, it’s getting in punches. And AI allows you to do that with increased amplification, increased sophistication.
Yeah, there’s some ads out there that are generated by AI that are very obvious, as one of the previous guests said, but there’s also plenty of ads out there that are not obvious at all and do not have disclosure.
CHAKRABARTI: Okay, so that to me seems to be one of the biggest differences, that we can take the old tools of propaganda, but we can use those tools in orders of magnitude of more intensity, right?
In terms of the speed at which these ads can come out, the velocity at which with they can, course across social media or even television, the number of times people can encounter these ads. How does the sheer volumetric aspect of AI play into your concerns?
WOOLLEY: It matters a heck of a lot.
It’s a huge part of the concern. Every time I give a talk on computational propaganda or the ways in which AI and automation make propaganda more sophisticated, someone says to me, “Propaganda’s been around forever.” “What about old attack ads?” And I say, “Listen, the speed and scale at which these things spread, the amount of times people come across them matters a lot.”
And then also, the thing, another thing that matters, which Hany kind of alluded to, is that you can demographically target specific groups of people. If elections are won by a couple hundred thousand seats, what happens if you’re able to target people in swing districts in Ohio with highly specific AI-generated ads, or in swing districts in South Florida for that matter, that are very much speaking to them?
And you can beta test 100 different messages to see which one hits the hardest, and then go with that, and then show it to everyone that you can through social media’s own ads dashboards.
CHAKRABARTI: Yeah. So exposure really matters here. I completely get that. But I also continuously come back to what Eric Wilson said.
He was the Republican political ads guy we talked to earlier, who said, “Look, if stuff is illegal, it’s already illegal,” right? That you can’t put words into a candidate’s mouth that person never actually uttered and claim that it is true. So the same applies for AI. We already have regulation in place, so what that we can just do it more?
Your thoughts.
WOOLLEY: I just, I don’t buy that. I think that’s not, that’s misrepresenting the situation. AI change substantively changes the ways in which we can attack people, substantively changes the ways and the rate and speed at which we can attack people, and oftentimes it’s happening without disclosure because of a lack of laws that exist in the space.
States have a pastiche of laws on the books, but we have no federal law about this. And I think that the point is well, we’re already attacking people. AI, we’re not saying things they didn’t say, but the fact of the matter is of those videos that the Western media project tracked, that have been talked about, those 200 some videos, 80% or more of which were spread by Republicans, many of them did include completely factually inaccurate content.
And so perhaps, yeah, we should use the laws that we have on the books regarding libel and slander, but also there needs to be some new laws to track this new technology.
CHAKRABARTI: Wouldn’t these laws then be applicable to all political ads? Because I can think of a lot of attack ads that were also just factually untrue, but they made it to air or people’s feeds anyway.
WOOLLEY: Look, I think that we get into really dicey territory and that’s why the previous guest talked about this with laws related to the First Amendment and free speech. The question is, is the AI substantively changing the context? Is it changing the ways in which people are saying something?
Is it misrepresenting something someone never said or did and making it look like they actually said or did it? Or is it one of those ominous attack ads of old where they say, “James Talarico, xxx,” there’s a substantive difference between Talarico himself saying it out of his own mouth and some ominous voice saying it.
I think that we have to think through the differences there, and it’s not an easy process, and I do not admire lawmakers who have to figure out how to make this policy.
CHAKRABARTI: Okay, so where would you start, though? Because I’m always still looking for solutions here, and your concern seems both warranted and quite profound.
And as you said, we’re getting down to technology and speech, which is such a huge issue. So we’ve got about a minute and a half left, professor. Where would you start in terms of advising, it sounds like you’re calling for some kind of federal regulation or at least framework.
WOOLLEY: I think that Dr. Farid, made a really good comment when you played his tape. What he said is, “Look, at the very least, we need disclosure,” and I absolutely agree with that. We need to have labels on these videos, and many of the videos that we’ve talked about the video here in my state, in Pennsylvania- one of the ones of Shapiro, the disclosure didn’t come until 45 seconds into the advertisement. We’ve gotta have disclosure about the fact that AI plays in these ads. Anyone that says differently is, I think, misleading us about what we deserve to know in this democratic process. And we need to have robust discussion at a federal level about what laws look like in this space.
And yes, I do think we need regulation. I think that we will probably see, the Democratic Congress begin talking about this because this has been such a huge issue.
CHAKRABARTI: Do you ever wonder what the Soviets would’ve done if they had AI?
WOOLLEY: Oof, they would’ve had a field day. And you know what?
Russia has picked up the post-Soviet toolkit, and so has everyone else. We’re all using that toolkit now. That’s what we’re seeing.
This article was originally published on WBUR.org.


