Quinta Jurassic (4:33)

I'm Quinta Jurassic, a senior editor at Lawfare, and this is the Lawfare Podcast November 29, 2023. Today we're bringing you an episode of Arbiters of Truth, our series on the information ecosystem. Unless you've been living under a rock, you've probably heard a great deal over the last year about generative AI and how it's going to reshape various aspects of our society. That includes elections. With one year until the 2024 US presidential election, we thought it would be a good time to step back and take a look at how generative AI might and might not make a difference when it comes to the political landscape. Luckily, friends of the pod, Matt Peralt and Scott Bobwe Brennan of the UNC center on Technology Policy have a new report out on just that subject. Together with Eugenio Laustri, Laufer's Fellow in Technology Policy and Law, I sat down with Matt and Scott to talk through the potential risks and benefits of generative AI when it comes to political advertising, which concerns are overstated and which are worth closer attention as we move toward 2024. How should policymakers respond to new uses of this technology in the context of elections? It's the Lawfare podcast November 29th. Will generative AI reshape Elections? To start off, can you just set the scene a little bit and tell us about this report that you've put out on generative AI in elections so.

Matt Perault (6:04)

What we were trying to do with this report was explore the craft of translational policy development with respect to generative AI and political ads. And what I mean by that is that we looked at the harms associated with generative AI and political ads. I actually should say the alleged harms. So in the discourse around this new technology, what are people thinking the potential harmful use cases might be? And then we examine those harms in light of the academic literature. So what does the academic literature say about how these harms might manifest themselves and how harmful they might actually be? And then we tried to develop a policy framework rooted not just in the conversation around the harms, but what the academic literature suggests about them. So we looked at four harms. The harms were scale, authenticity, personalization and bias. And then we looked at some of the policy solutions that have been suggested in this area, like watermarks and disclaimers, for instance. And what we found, on balance, is that many of the harms have actually been overstated. The academic literature suggests that the harms in reality will not be as acute as the conversation around these issues suggests they might be. But that doesn't mean there aren't any harms. The harms that we thought the literature suggests we should pay more attention to to not less, are the potential use of generative AI in down ballot in smaller races, as well as the harms of generative AI related to bias. So its ability to exacerbate bias in our society. We found. Now I'm turning to the intervention side. We found that the interventions that have been discussed, watermarks and disclaimers are probably unlikely to be a silver bullet. So the recommendations that we have focus on a different set of potential policy interventions and we group them into two primary categories. The first is that public policy should target electoral harms, not technologies. So we shouldn't just be concerned about the use of generative AI to do problematic things. We should be concerned about problematic things that might happen in our electoral process, regardless of the technology that might be used. And then the second recommendations are focused on learning. So how can we gather more data that will help to fill holes in the current research? Because our view is that even though the research right now points in a certain direction, we think it's incomplete in many areas. We think it's probably inconclusive in lots of areas. And so we think there's more learning to be done to inform governance in the future.

Eugenia Laustri (8:25)

So how speculative do you think all of this is? I was following the election cycle in Argentina and there is an interesting New York Times article that focuses on the use of AI for political advertisement there. And it was clearly, you know, it featured in the election cycle. And I have to say most of the content was clearly identified and identifiable as artificial. And it doesn't really seem in, I have to still see more analysis on this, but it doesn't seem like it had a particular influx on the decision that was made. So do you think that's the type of use cases that we can see for artificial intelligence, or do you see an evolution in how AI can be applied to political advertisement?

Scott Bobbie Brennan (9:17)

So, honestly, I'm not super familiar with the case in Argentina, but what I can say is you're right that much of the discourse around the particular harms of generative AI and political AI is super speculative. And really that's what we were addressing with the report. It was to try and bring some rigor to, to the conversation around what the harms may actually look like. So to kind of, in addition to what Matt has already said about the report, I say we review some of the known examples of generative AI in political ads in the US over the past year. And you're right that they're, I don't say underwhelming, but not likely to have had a huge impact. Things like the Ron DeSantis campaign editing in fighter jets behind him when he's giving a campaign stump speech or, you know, some of the more, you know, you know, potentially kind of interesting ones were again, the, I think it was a super PAC aligned with DeSantis creating an image of Trump and Fauci, Anthony Fauci hugging now, you know, but, but I think you're right to point out that, you know, whether or not, you know, we might be, we might find some of these examples concerning the next question is, well, what sort of impact will these uses likely have on, on the, on an election itself? And that's where, you know, rooted in our analysis of the existing literature on political ads and on misinformation, we concluded that the literature seems to be suggesting that they're unlikely to make a huge difference in the, the candidate that, that voters end up choosing because frankly, misinformation and political ads for, you know, most of the time have limited impact on, on persuasion on, on who we ultimately vote for.

Matt Perault (11:28)

Eugenia, I think your question is in some ways that is the question that animated our desire to do the report. So where generative AI is today in terms of its use in the electoral process is not where it's going to be in tomorrow. And so we wanted to look at how do things look in an off year election in 2023 with the idea that we should be better prepared going into 2024. And I think, as we've been discussing, it's really hard, I think, to come out with what we would conceive of as smart, data driven, informed public policy going into 2024. Given how much we don't know right now, there's an enormous amount that we don't know. But the idea is to, if policymakers were to implement some of the recommendations in the report, I do think that we would be better prepared going into 2024. That's primarily the set of recommendations that are focused on addressing electoral harms. And then, and then, as Scott and I have both been describing, we also would be well, better positioned, I think, to learn about the potential harms and about the value of various different potential interventions.

Quinta Jurassic (12:39)

I want to go back to the discussion of how much of a difference any of this makes to begin with because Scott, you mentioned a really interesting part of your report which has to do with the question of, you know, not only just does generative AI make a difference, but do political advertisements make a difference at all? And I think that it's very striking how you have this sort of literature review that essentially concludes nobody really knows if political ads make a difference or not, or to the extent that there is an answer, it seems like they might not make any difference. So could you talk about that a little more and sort of sketch out how that explains how we should be thinking about generative AI? I mean, does that mean that the freakout is overstated?

Scott Bobbie Brennan (13:25)

Sure. You know, I was surprised to see at the end of our review of literature that first, you know, given, you know, understanding that there are, you know, there's always sort of need for additional research that the kind of current sort of thinking on the impact that political ads have on, on persuasion is, is essentially zero. There was a, there's a great review piece that came out recently that looked across, I think was 49 different field studies and concluded that, that, you know, I think they said something like, you know, our best estimate of the effects of campaign contact and advertising on Americans candidate Americans candidate choices in general elections is zero. You know, that being said, right. I don't want to just kind of suggest that political ads have no value or any sort of impact at all. You know, there seems to be consensus that they are more effective at influencing behavior. So turnout for elections, donations, signing up for an email. Right. So like giving your personal data to a campaign. All of these are things that political ads can have some sort of impact on. But, but that's a, that's a big deal, right? To, to recognize that political ads are very unlikely to change who we vote for. And, and given that, I think, you know, to, to go back to your final question. Yeah, I mean, I think some of the concerns around generative AI and political ads might be overstated if we recognize this.

Matt Perault (15:06)

I think the data points in the direction that, as Scott said, some concerns are overstated, but others are understated. We didn't do like a media analysis to see how press is covering misinformation in the election. But my guess is that the overwhelming volume of reporting is focused on the idea that the content moderation practices on X are going to have a dramatic effect on whether people prefer Joe Biden or Donald Trump. And I think the literature suggests very, very strongly that, that that is a dramatically overstated concern, that if you're trying to persuade someone to move from voting for Biden for president to Trump for president, that's not likely to happen via political advertising. But as we said at the outset, that might be less true in smaller down ballot races. So the kinds of races that Scott and I were voting in just a few weeks ago in the 2023 election are more likely to be impacted probably than the presidential election. So when we were voting for mayor of Durham and Durham City Council, those are the kinds of races where there's going to be much, much less focus, much less media attention on the race, which means probably much less ability to detect misinformation campaigns. There's going to be a smaller volume of content. So the use of something like generative AI to increase the volume of problematic content might be more likely to have an effect. And so those are the kinds of things that we think where there should be more focus, more media attention and public policy should be directed at trying to address those kinds of harms.

Quinta Jurassic (16:41)

Yeah, I think that point about local races being different is really important. And it also makes me think about, in the same way that you see a lot of coverage, for example, of generative AI enabling pornographic and sort of violating deepfakes of people in high school who are being bullied by their peers, that there's, there's maybe, obviously these are very different situations, but I do wonder whether there's kind of a similarity there insofar as the technology is more powerful when it's used against someone who has less sort of stature, less ability to push back. You know, if there is a deep Fake of Biden saying something egregious or doing something egregious. He has a huge platform to get out there and say this didn't happen. Whereas if it's, you know, a kid in a high school or a local candidate for school board or something like that, people are less interested in finding out the truth. And there's less of a avenue for that person who was targeted to kind of put the truth out there. And so in that way, it strikes me that the impact can be so much greater in these sort of less prominent spaces.

Matt Perault (17:52)

Yeah, that seems exactly right to me. Yeah.

Scott Bobbie Brennan (17:55)

I honestly don't have anything to add to that.

Matt Perault (17:57)

You summed up really well.

Scott Bobbie Brennan (17:59)

Yeah, no, I mean, we're. I think we're done here.

DeleteMe Advertiser / Sponsor Voice (18:01)

I don't.

Quinta Jurassic (18:01)

I think there's nothing more to say, Quinto.

Matt Perault (18:05)

I mean, I think. I think what you're part of what you're getting at. And apologies if I'm going a different direction rather than just restating, but I think with any new technology, there are so many nuances that we don't know. And those nuances are so important, I think, in figuring out what the right regulatory framework might be. And that's why we put so much emphasis on research in the report, with a real hope that that is not just a throwaway line, that that's not just perceived as a call to do nothing, but is really oriented around understanding how the technology creates benefits in the world and how it creates harm in the world. One example, I think, where we can see the train leaving the station, and I would say based on the past work that Scott and I have done, I wouldn't say I'm very optimistic about our ability to. About the likelihood that we will study. It is the idea of watermarks and disclaimers, which now two significant platforms have rolled out as their own policies. So Meta and Google both will have policies in the 2024 election that require some level of disclosures on digitally altered political ads. And the. The thing that Scott and I would hope for is that we use the 2024 election to understand that intervention better, that we use it as an experiment to. To understand what is the impact that that has on what people think. Does it result in a decrease in volume of misinformation or a decrease in how people perceive potentially false information that they see? Does it create competitive effects? Like, is that. Is that an intervention that's easier for large companies to implement than small companies? And our fear, I think, is that we will go through this election cycle not learning that so not unlocking and exploring some of these nuances that we hope will inform smarter public policy in the future.

Eugenia Laustri (19:58)

So I want to jump in and take us back a little bit to something that you were mentioning at the beginning. You set out four different categories of potential harms that could derive from the use of generative AI and political advertisements. They were scale, authenticity, personalization, and bias. And I think it's useful if you could walk us a little bit through the literature review that you did. How do you understand these harms and what we know and what we don't know yet about each specific one?

Scott Bobbie Brennan (20:34)

So the four harms that we identify so, you know, really come from, you know, a broad reading of the. Really the kind of public, almost popular kind of discourse around, around generative AI and political ads. You know, I'll say right off the bat, we don't have much of a scholarly, you know, academic literature, you know, offering great empirical insight into this question just because it's so. It's so new. But looking kind of broadly across.

Eugenia Laustri (21:09)

The.

Scott Bobbie Brennan (21:10)

Conversation, we identified these four harms. And I guess I can start with scale. So the idea that generative AI will make it will lower the cost and the effort that go into producing deceptive content in political ads. Again, while we don't have a great sort of sense of how much more will generative AI increase the, the quantity, the supply of false and deceptive content, we do have some literature first on what sort of impact do, you know, false ads or do ads or do false, false content have on voter choice? We talked about that a few minutes ago. But also, you know, what does the literature say about. About repetition of, about, you know, that we see in things like political ads, you know, you know, as I sort of pointed to earlier, again, while the literature is sort of inconclusive here, and recognizing that, you know, repetition is all like, has been a very sort of key strategy for both political advertisers and disinformation producers. That again, the literature kind of suggests that the harm here might be somewhat overstated, sort of similar kind of conclusions around things like, around authenticity, where the concern was will generative AI allow for, you know, bad actors to create more sort of convincing deceptive content. And here, you know, I'll just point out that the sort of existing literature on visual mis and disinformation really points out that, you know, despite deepfakes being around now for a few years, despite there being, you know, pretty great kind of methods of, of. Of convincingly editing, you know, images and video for some years, Most of the recent visual disinformation has used really sort of simple techniques, you know, cheap fakes rather than deep fakes. So for, you know, the best example of this I think is the like the Nancy Pelosi video that was released some years ago where I think the video was just sort of slowed down to make it look like she was, she was slurring her speech. So given that right, we have to sort of separate out the questions about the, you know, how, how, how photorealistic is a, is a piece of deceptive content from the sort of impact that it will have. And then, you know, the other two harms personalization. So we talk through the, we look at the literature on what know about the effectiveness of ads as they're increasingly targeted to smaller groups and then finally on bias. And you know, I'll just say I think this was one of the areas that I think, as Matt has already said, really suggests that the concern here has been maybe understated rather than overstated. It's that we have a lot of good indication that a lot of the large language models have sort of built in biases. And there's a real concern that when they're used to produce political ads, especially if it's the sort of like boilerplate political ad content, they very possibly could be replicating or amplifying those biases.

Matt Perault (24:47)

Why choose a sleep number Smart bed?

Eugenia Laustri (24:48)

Can I make my site softer?

Matt Perault (24:50)

Can I make my site firmer? Can we sleep cooler? Sleep number does that cools up to eight times faster and lets you choose your ideal comfort on either side your sleep number setting. Enjoy personalized comfort for better sleep night after night. It's our Black Friday sale.

Scott Bobbie Brennan (25:05)

Recharge this season with a bundle of cozy, soothing comfort.

Matt Perault (25:08)

Now only $17.99 for our C2 mattress and base plus free premium delivery. Price is higher in Alaska and Hawaii. Check it out at a sleep number store or sleepnumber.com today.

Quinta Jurassic (25:17)

This holiday, give the gift that says let's cancel plans and just lounge. Meundies has dropped their new holiday collection and it's made from maximum cozy. We're soft as snow, ultra modal fabric, festive prints and loungewear so comfy your couch might get jealous. Onesies, hoodies, joggers, even delightfully quirky holiday designs. You're welcome. Knock out all your holiday gifting needs with deals up to 60% off at Meundies.com acast enter promo code acast that's Meundies.com acast code acast Amazon Pizza Hut.

Scott Bobbie Brennan (25:56)

Audible How'd they get so big without soul destroying complexity? On Founders Mentality, the CEO Sessions, we're.

Quinta Jurassic (26:03)

Gonna find out who's number one is the customer. Whose Walmart is it? My Walmart.

Matt Perault (26:07)

If you looked at Audible, it was kind of like growth, growth and then growth.

Quinta Jurassic (26:11)

It separates Amazon and AWS from anyone else.

Scott Bobbie Brennan (26:15)

Join me, Jimmy Allen, partner of Bain and Company to hear surprising stories from.

DeleteMe Advertiser / Sponsor Voice (26:20)

The world's greatest leaders.

Scott Bobbie Brennan (26:21)

Subscribe to Founders Mentality to CEO Sessions now.

Quinta Jurassic (26:27)

So I wanted to ask about the scale question because one of the arguments that I've seen about why the ability of generative AI to allow campaigns to generate just more potentially false information, sort of a pushback against that, is that the bottleneck for this kind of thing is at the level of distributing this material rather than generating it. And I've seen that argument made in context of thinking about state run influence operations or disinformation campaigns that it's not that hard to come up with a lot of false material. The problem is building those networks to make sure that people actually see it rather than just kind of sending it out into the void. I'm curious what you think of that argument in connection with the concerns about scale and whether it holds for political advertising conducted by a professional campaign as opposed to say, sort of COVID social media influence effort.

Scott Bobbie Brennan (27:33)

I mean, I think that's a great question and a really good point. The first thing that came to mind for me was a lot of the early Twitter studies on misinformation from, you know, right around 2016, 2017, you know, we're much more interested at sort of counting the, the amount of misinformation and disinformation on the platform rather than looking really closely at how it was distributed at things like viewership. I think it's a great point that, that you know, the thing that we really care about is what sort of impact, you know, do falsehoods have on voters. And, and there is not much evidence that there is such a simple relationship of if you just produce more bad content that it will, people will be persuaded by it. Absolutely. You know, I think you're right to point out that like there may be something slightly different between, you know, a foreign influence operation and a political campaign which does have a sort of more robust distribution network. But you know, it's, I think it goes to, you know, you know, are just sort of like lack of understanding of what really a, you know, a massive campaign of deceptive content run by a major political campaign might actually look like. We just, we just don't really know. So, yeah, much, much to sort of keep an eye on for the next election.

Matt Perault (29:07)

But Quinta, I think the one key thing about your question is that it's sort of premised on the idea that someone seeing misinformation is going to result in a change in that person's viewpoint. And so, as Scott said, the literature just suggests that at a minimum, that's much more complicated than people think, and even potentially that it really has very limited impact, at least in sort of isolated incidents. And so if that's the case, then then the difference between creation and distribution matters dramatically less.

Scott Bobbie Brennan (29:39)

You know, I've been a little, a little worried that someone reading our report might believe that we are saying that misinformation never has impact on the world and misinformation can never do any harm. That's, that's, that's false, right? Like, we have examples of significant harm happening as a result of, of, of large disinformation or misinformation campaign. So I think that the key is to try to understand, well, what, you know, what's different about those few examples that we have. Climate disinformation, you know, anti vax movement, the election denialism, you know, these are honestly sort of the outliers. These are the extreme cases. And to me, the lesson there is, you know, while the literature is pretty clear that a single piece of misinformation spread online probably has minimal impact on what someone thinks or what someone believes, a targeted multimodal across many different media campaign that it is deeply rooted in political ideology and identity can have a pretty significant impact. But given that, what difference is it going to make for a campaign within a campaign like that, within an effort like that, create one, you know, a couple extra pieces of, of somewhat more realistic falsehoods? I'm not, I'm not sure what difference that will make.

Matt Perault (31:10)

I think part of what Scott is getting at here is that if you look at the overwhelming majority of the nature of the dialogue around this issue, you, you're seeing things like what is Facebook going to do about COVID misinformation content? Or what is X's position going to be on a particular kind of information that some people, or maybe a lot of people think is like, is really deeply problematic. And I think the implication of what Scott's saying is those questions just have much less impact than the policy debate suggests that they do. Because in order for there to be impact, there has to be a lot of connection to other things throughout society. The presence of the information on other media or its connection to identity politics or what's happening in terms of offline conversation. And I don't think that means we should bury our heads in the sand and not care about what Meta's content moderation practices are or X's content moderation practices are, but in isolation shifts to those practices are, I think the literature suggests, very unlikely to have a significant impact. That's where an enormous percentage of the policy conversation on misinformation is. And what Scott and I try to do in the report, when we get into the recommendations, is to sort of pivot the focus to things that I think have gotten much less attention, but that we see as really impactful. Like it continues to be the case. And we are like, continue to be shocked at this, but it continues to be the case that there is no federal law on voter suppression. It is not a violation of federal law to use deceptive practices in voting with the intent of suppressing the vote, which is astonishing and I think is the kind of thing that is technology agnostic, can target a problem that a lot of people think is a problem and open up areas for law enforcement to actually take steps to address a real harm.

Scott Bobbie Brennan (33:01)

Yeah, I guess the other thing I'll say is that I think there is a discussion to be had about the value of partial solutions, whether that's watermarking or X's moderation or policies that. Well, it's true that many, maybe, you know, what X decides to do, as far as, you know, what content to keep up or take down is not going to solve the global problem of mis and disinformation. It can make a sort of small difference, meaningful difference.

Matt Perault (33:42)

And.

Scott Bobbie Brennan (33:42)

Well, of course there's the concern, right, that like, too much attention being given to, you know, in these, in these kind of, you know, these single kind of questions about platform policies is drawing attention away from the sort of, like, broader conversation that we could be having. You know, I do think that there might be some value. You know, I don't want to say that there's no value in those solutions at all. And here I'm. My brain is telling me to stop talking, but I really want to talk about this Bertolt Brecht play about communist rice planters, but actually, I will stop.

Matt Perault (34:18)

So.

Quinta Jurassic (34:19)

Well, now I'm curious.

Scott Bobbie Brennan (34:21)

Well, yeah, it's. There's this play by Bertolt Brecht where these two communists spreading the gospel of communism go into the rice fields, I believe, in China, and one of them, they see the, the pain and toil of these rice workers and one of them basically because the rice workers are like spending their days like knee deep in water and one of them invents these platform shoes basically. Or I think it just has has the suggest that the workers can stand on these blocks so that they don't have to wait in the water all day and their communist masters end up, I think, punishing this the sky because by lessening the, the burden, the suffering of the workers is going to delay their sort of awakening into, you know, class consciousness and will ultimately delay the the revolution.

Quinta Jurassic (35:21)

Let me ask about something that may be the equivalent of the platform shoes, which is that. So without focusing too much on platforms, I do think it's worth talking a little bit about the role of the platform, since they are one of the big stakeholders here and about how they envision their responsibilities, particularly given that we've seen recent change in Meta's policies around political ads in terms of allowing content about the supposed theft of the 2020 election. A lot of different platforms are sort of figuring out where they stand on generative AI. There's also this additional complication where a lot of platforms are kind of rolling back the resources that they're putting toward countering election misinformation. And maybe this is a bit of a stretch, but I would be remiss in not mentioning the ongoing chaos at the leadership in OpenAI, which seems to be in turmoil when it comes to thinking about how the company should prioritize profits over safety concerns. So I'm curious how you see the role of companies in this landscape and how you understand these companies as kind of weighing the priorities and where that fits into mitigating potential harm from generative AI in elections.

Matt Perault (36:39)

So I don't know if it's a mistake to keep going with the Brett analogy, but I think it's actually sort of interesting in that it seems to show to me the importance of cost benefit analysis in evaluating various different solutions. And that, I think is an important and helpful framing for evaluating lots of different policy solutions in lots of different contexts, including not just government efforts to address policy issues, but also company efforts. So I think what we're seeing here is in the absence of federal law, and largely in the absence of state law as well, platforms are taking steps to try to address some of the alleged harms. I think the question is, do those interventions have benefits that outweigh costs? And that's something that we really don't know. I think it's very interesting that Google and Meta both have approaches here that are focused on, that create the impression that watermarks and disclaimers have value and have positive value. I don't know that that's definitively the case. And as I was saying earlier in this conversation, my hope is that we use the 2024 election to learn about those things. So we learn about what are the benefits of disclaimers, what are their costs. I hope that we hear from a range of stakeholders about that and not just from Google and Meta, and that we understand more about the impact that, that they, that those kinds of interventions might have on smaller companies and, and, and that platforms share enough information or make enough information available so that credible third party researchers can evaluate this and that we use the 2024 election as a test of some of this and, and that we're able to learn enough from the 2024 election that when we go into the next one in 2025 and 2026 and 2028, that we will actually be able to make more informed decisions about watermarks and disclaimers.

Eugenia Laustri (38:32)

Well, since you brought up watermarks and disclaimers, let's turn it to the executive order on AI that came up from, what was it, two weeks ago? So the executive order requires the Department of Commerce develop guidance for content authentication and watermarking to label AI generated content. How effective do you see these interventions considering what you just said, and if we're not sure about them, what else could the executive branch do at this point?

Scott Bobbie Brennan (39:09)

Yeah, well, I think honestly we don't know as a starting point. I think having that the executive order directs there to be guidance more thought behind how we might actually go about trying to identify generated content is a really good thing. But the question of what effect labels on generated content might have on something like, you know, voters understanding that an ad is false or is a piece of deceptive content. Yeah, we just don't really know. I think it's still too early to have good sort of empirical data about, about the effectiveness of watermarks. We do though have some sort of empirical studies about, about the effectiveness of political disclaimers around, around like payment. So when an ad says, you know, this ad was paid for by, you know, this campaign or this political action committee and you know, well, that literature is, seems to suggest, you know, some studies suggest that, you know, they have some impact. Other studies suggest that people don't really pay attention to them and they have no sort of real meaningful impact. I think that the takeaway there is just they're not going to be sort of silver bullet, you know, solutions that are going to, to solve all of the problems about generated content.

Matt Perault (40:41)

I really like this provision of the Executive Order. And I actually think it's similar to lots of provisions in the Executive Order. And I like the spirit of it, which is focused on trying to develop standards and best practices that would not be mandatory. There's no liability or penalty for companies that fail to implement the best practices. There's just an emphasis on trying to work hard to develop them. And I think that's exactly right. I think there are sort of two things that I think are missing. So the first is kind of, as we're saying, we're putting the emphasis on, well, do watermarks have value? And so if all you're doing in the Executive Order is focusing on standardization, what should watermarking look like if it is implemented? You're kind of missing the key point, which is, does watermarking have value? And so I would have liked to see something in the Executive Order or in future government action that would do more formal research on that, that would fund third party researchers to evaluate those questions. So again, we could learn more going into the next election. The other thing, and it's maybe, maybe this is like kind of an attenuated concern. Quinta, I think it's something that you've thought and written about in the past, but I do have a little bit of a concern of moving in a homogenous direction with approaches to remedies here. And so I think this has kind of been true of the last 5, 6, 7, 8 years of content moderation practices generally, which is that there has been less diversity of approaches, which gives consumers fewer options in terms of being able to pick platforms based on having really divergent content moderation practices. So when I worked on the public policy team at Meta, and when I started there, Facebook had a real name policy and Twitter did not. And I thought that diversity was really important. People who really wanted to be able to have a voice, not in their real name, had a choice in the market about, about the platform that they could use if they had that preference. And when I see a lot of emphasis on standardization and best practices, again, I generally think that's a really good move. But I do have a little bit of concern that it cuts out some innovation at the edges that platforms might experiment with and instead they just race to be in the pack in.

Eugenia Laustri (42:54)

In some of your previous research, you've studied how state governments stepped into the void that was left by the federal government when it failed to move forward on tech policy. So when it comes to generative AI, are states Doing something different. Are there particular concerns that they're trying to address or is it kind of the same discussion that we're seeing at the federal level?

Scott Bobbie Brennan (43:18)

Yeah, that's a great question. A little of both. States are certainly not waiting for the federal government here. One estimate I saw was that I think 125 bills were introduced this past year at the state level, not just introduced. And I think that was, I cast a very wide net on what counts. And those proposals range pretty, pretty widely from a couple of states at least, like Massachusetts and California introduced kind of comprehensive regulation, didn't pass. Right. But introduced this, you know, that would cover sort of a wide range of different kind of aspects of AI regulation. And other states, you know, took a far more sort of targeted or sector specific approach. And as far as, you know, looking at the. I think it's probably, I think it's somewhere in the, like, you know, under, under two dozen bills that were actually passed. Most of them are that sort of very narrow, sector specific regulation. There's a law in Georgia that passed that prohibits the use of AI for, for eye exams or there's a law in West Virginia that requires the use of AI for state road inspections. So we see that. We also see a lot of efforts by states to just build up their capacity to study AI to make better regulations. Commissions, working groups, state sort of audits. So that's what we're seeing so far.

Quinta Jurassic (44:51)

So I want to make sure we give you an opportunity to talk more about some of your policy proposals for how we should handle this. And you've touched on a bunch of them over the course of this conversation. But I want to give you space to turn back to those. If you wanted to discuss in more detail or if there's anything we haven't touched on that you want to make sure we discuss.

Scott Bobbie Brennan (45:12)

Sure, I can name one. Going back to the concern around local and state elections and how generative AI and political ads might be more effective at the state and local level. We picked up this idea from Steve Bannon, right. Who once described his, his approach as flooding the zone with shit. Right? Flooding the zone with, with problematic, confusing content. And we realized that the real concern at the local and state level is, you know, there just is less sort of discussion. There's less, there are fewer ads, there's less content of any sort, which is why a single piece of deceptive content might be more potentially effective. So the recommendation is, you know, state, local governments should flood the zone, right, with good factual content to help kind of potentially Drown out or dilute, you know, efforts by bad actors to introduce deceptive content.

Matt Perault (46:14)

And I can add one as well. So in the targeting electoral harms section, we have one recommendation that focuses on allocating additional funding for law enforcement to enforce existing civil rights law that protects the electoral process. So this was one of the things that. That we developed before seeing the White House's executive order. And the executive order doesn't exactly do this. I don't think it allocates additional funding, but it does call for coordination led by the Civil Rights Division of the Justice Department. I actually worked in the criminal section of the Civil Rights Division one summer in law school. And the criminal section is filled with devoted attorneys who wake up every day trying to figure out how to build cases to prosecute the country's civil rights law and ensure that it is enforced and hold perpetrators, violators of civil rights law to account. And that's an example, I think, of where existing. Existing law, you don't need new law here, but where existing law can really be a powerful tool to address what, you know, Scott and I have said repeatedly, really is, we think, a concern associated with generative AI that is supported in the literature. And that's the issue of bias. And so I think it's important to remember that attorneys don't just get handed cases, like, out of thin air and all, and just take those cases that they're handed to the courthouse to prosecute them. They need to build cases. They need to identify facts that are a violation of law. They need to make that connection. What is the factual record that supports a violation of civil rights law? And that takes skill and expertise. And it's not just skill and expertise around the civil rights law, which obviously current attorneys clearly are skilled in that area. But it will also require deep skills and understanding about the technology, and so allocating additional funding that can build up that expertise and really provide the technical support to people to build cases to enforce existing civil rights law, we think is really critical.

Quinta Jurassic (48:19)

Yeah, this is an issue that we've talked about separately, but I do think it's crucial. There's some indication that the Justice Department may be beginning to take this kind of question of social media and voter access more seriously. There's recent prosecution and conviction of a pro Trump Twitter troll who directed voters to text their vote in, so essentially encouraging them not to vote in a way that actually counted in 2016 under a criminal statute that dates back to reconstruction. And so I do wonder if that may turn out to be a model for what the Justice Department could do if it has the funding and if that conviction holds up on appeal, which is an open question. But to close out, I want to kind of look forward and ask you both what your predictions are. It's the million dollar question for whether generative AI is going to create chaos for the US in 2024 and for other elections around the world that are coming up.

Matt Perault (49:22)

So maybe I can give the optimistic view and Scott can jump in either with optimism or pessimism, but my sense is he may have a different perspective here. So I think generative AI is going to be a powerful tool for advancing human rights in and that on balance, it will advance human rights more than it contracts them. So one thing I think we actually didn't touch on, that we, that we do include in the report is a discussion of, of generative AIs benefits and how it can actually reduce barriers to entry and barriers to speech for people who might historically have not been able to speak in elections. So I think going forward, there certainly will be chaos. There certainly will be examples that we point to of the use of these tools in nefarious and problematic ways. But often in the discussion of them, we sort of focus on numerators rather than the relationship between numerators and denominators. So we, we identify harmful use cases, but we don't really put them in context. And my guess is that, that these tools will be, will be used powerfully by a range of different types of people to explain themselves and to express their viewpoints and that on balance, it will be positive. I probably, on balance, I guess, am somewhat more pessimistic about the direction of travel. On the policy side, I do think the executive order is a really positive step forward. I think it calls for the kinds of deep thinking and learning that I think we really would hope to see and that we express optimism about in this report. But my fear is that we will see examples of how this technology is used in the 2024 election and we will see examples of either from the government side or from the platform side of attempts to implement interventions to constrain those harms. And yet we will be here if we do this again, Quinta and Eugenia, in a year. And you say, well, what did we learn in this election? What do we know now that we didn't know a year ago? My fear is that we won't have a lot we'll be able to say.

Scott Bobbie Brennan (51:22)

In response to that. I guess I'm both optimistic and a bit pessimistic.

Matt Perault (51:28)

Both.

Scott Bobbie Brennan (51:28)

So, you know, I echo, you know, I agree with Matt on a lot of what he said. I am quite optimistic that just with the amount of attention and interest and thought that's going into, to these problems, to AI regulation more generally, I think it's a really good thing. I like that there are, you know, unending, you know, Senate hearings and about, about AI and just every day new new research. I think I am a little less optimistic, you know, than Matt about the benefits of generative AI in elections. Truth is, I am concerned. I'm not concerned because of, I believe that we're going to see a whole lot of false, deceptive content that's going to make people vote for candidates that they wouldn't otherwise vote for. I'm worried that it will just exacerbate a, an existing trend, a radical decline in trust that whether or not generative AI or falsehoods or political ads have any sort of real impact on what we believe, believe. People believe that it does and people believe that it is a big problem. And that is concerning to me. We know that trust across, in, across institutions is falling and continues to fall every year. I don't see how this cannot worsen the problem. There's the concept of the liar's dividend. You know, the way that bad actors can use things, you know, can use AI or generative AI to, to claim that real, you know, evidence of their wrongdoings is fake. Right. And when there's a sort of low trust environment, I think it's, it's, it's, it's easier for that, that sort of thing to happen. So. Yeah, so, so I guess, I guess a bit both. Right. A bit optimistic about at least that these problems are getting a lot of attention. A bit pessimistic though about what the future will bring.

Quinta Jurassic (53:44)

Well, let's end there on that mixed note of optimism and pessimism. Matt Scott, thank you so much for joining us.

Matt Perault (53:51)

Thanks so much.

Scott Bobbie Brennan (53:52)

Yeah, thank you.

Quinta Jurassic (53:56)

You've been listening to Arbiters of Truth, a Lawfare podcast series on the information ecosystem. The Lawfare Podcast is produced in cooperation with the Brookings Institute. You can get ad free versions of this and other Lawfare podcasts by becoming a Lawfare materials supporter through our website, lawfairmedia.org support. You'll also get access to special events and other content available only to our supporters. The podcast is edited by Jen Patya Howell and your audio engineer. This episode was Noam Osband of Goat Rodeo. Our music is performed by Sophia Yan. As always, thanks for listening. Listening.

Matt Perault (54:45)

Group Health Insurance can challenge company budgets with large and unpredictable rate increases. Now, a new form of employer coverage called an ichra or ichra, can help ichras make health plan expenses more predictable and affordable by tapping into a much larger risk pool. That means businesses know what to expect and what to budget. If you're tired of expensive surprises, plan on something predictable. Plan on an icra. Learn more at ambetterhealth. Com ICRA.