Bob interviews Nate Persily, the James B. McClatchy Professor of Law at Stanford Law School and co-director of the Stanford Law AI Initiative, about artificial intelligence and democracy, with particular attention to elections. They discuss what’s driving American anxiety about AI, why Republican and Democratic users alike rate AI models as left of center, and why Persily thinks the greater danger from deepfakes is not voters who are fooled but voters who stop believing what’s true—the “liar’s dividend.” They also take up the risks of relying on chatbots for voting information, state laws requiring disclosure of AI in political ads, and why Persily, though not “an unrepentant doomer,” thinks the time to regulate AI may be running out rapidly.
Mentioned:
Nathaniel Persily, “Misunderstanding AI’s Democracy Problem,” in The Digitalist Papers: Artificial Intelligence and Democracy in America (Stanford Digital Economy Lab, September 2024)
Sean Westwood, Justin Grimmer, and Andrew B. Hall, “Measuring Perceived Slant in Large Language Models Through User Evaluations” (Hoover Institution, May 8, 2025)
Hannah Waight, Eddie Yang, Yin Yuan, Sol Messing, Margaret E. Roberts, Brandon M. Stewart, and Joshua A. Tucker, “State media control influences large language models,” Nature (May 13, 2026)
Andrew B. Hall and Sho Miyazaki, “Voters Increasingly Use AI as Political Advisor. A New Study Shows the Risks.” (Freeman Spogli Institute for International Studies, Stanford University, March 31, 2026)
Stephen Ansolabehere and Shanto Iyengar, “Can the Press Monitor Campaign Advertising? An Experimental Study,” Harvard International Journal of Press/Politics (1996)
This is an edited transcript of an episode of “Executive Functions Chat.” You can listen to the full conversation by following or subscribing to the show on Substack, Apple, Spotify, or wherever you get your podcasts.
Bob Bauer: Welcome to Executive Functions. Today we’re going to be talking about a really complicated but also hugely significant topic: artificial intelligence and democracy, with specific attention to elections. And I am pleased that we have here today Professor Nathaniel Persily, who is a professor of law at Stanford University Law School. Hello, Nate.
Nate Persily: Hello, Bob. Good to be here.
Glad to have you. Let me say briefly something about Nate, who is a national and international expert on this topic and has written and spoken widely on it. As I mentioned, he’s a professor of law at Stanford. He is also co-director of the Stanford Law AI Initiative. He is the co-chair of the American Political Science Association’s presidential task force on artificial intelligence. He’s an expert on both election law and the governance of technology.
Today we’re going to focus on his AI expertise. So let’s begin by talking more broadly, before we zero in on elections, about AI and its impact on the functioning of democracy. We see a lot of anxiety being expressed, particularly in the United States, about the development of AI and its impacts in a wide variety of areas, such as the economy, but also real concerns about what it means for democracy. Can you describe generally (a) what those concerns are and (b) why they seem to be more acute in the United States than elsewhere in the world, relatively speaking?
Thanks again for having me on. I tend to think that we are still living through the techlash that began, say, after the 2016 election. And so the hangover from people’s dissatisfaction with the social media companies is affecting people’s perception of AI, particularly as it relates to AI’s effect on elections and democracy.
If you were concerned about, say, high levels of disinformation on Facebook or other social media, or bots on Twitter, this new technology comes in and you sort of transfer a lot of the anxieties you developed over that period to it. As you suggest, Americans are the most pessimistic in the world when it comes to AI, and they’re also about as anti–social media as any place in the world, despite the fact that we’re the ones who have the companies that are benefiting from it. A lot of this is just that poorer countries see this as a leapfrogging technology. There’s a little bit less anxiety on some of the issues related to job loss and energy usage and other things.
But right now, AI is sort of having a bad political moment, whether you’re talking about the technology in general or its effect on democracy. And I think that most people, when they think about technology and democracy, are still thinking about the problem of disinformation. So when it comes to AI, they’re thinking about synthetic imagery and deepfakes and the specter that we’re not going to be able to tell what’s true and what’s false.
(3:01) So is there any other category—I mean, beyond misinformation, and I’ll clue you into what I’m talking about: bias—that gives people concern about the potential effect of AI on democracy?
I think so. There are concerns about the bias of the model developers—that, as was true with the social media companies, maybe liberals in California are tweaking the models. And many of the studies that look at the political bias of the models do find that they are left-leaning. So I think there’s worry about that. But it’s not just a systematic political bias one way or the other. It’s whether a model reflects the preferences of its company, so that Elon Musk’s model might reflect his preferences, and OpenAI’s or Anthropic’s might reflect theirs. Or, a more complicated problem: whether what’s known as sycophancy—the fact that these models will cater to your expressed political beliefs—might reinforce polarization, so that the models give different answers to the same question depending on who the user is.
Now, you said something that I’m sure would hit conservative ears pretty hard. You said something about the models being generally left-leaning. In what way? Why would that be the case?
So my colleagues Andy Hall, who was here at Stanford Business School and is now at Anthropic, and Justin Grimmer actually had both Republicans and Democrats rate the responses of the models on everything from affirmative action to trans rights to birthright citizenship. Both Republicans and Democrats rated the models as being left of center, some more significantly than others. The Washington Post replicated this over the last six months and still found some of the models to be more left-leaning. Google’s, interestingly enough, was a little more middle of the road.
Now, there’s a hard empirical or methodological question here about how you sense the biases of a model. For example, if you ask the model whether climate change is real, should it give you an “on the one hand, on the other” kind of response? But what was interesting about Andy Hall’s methodology is that they just had Republicans and Democrats rate the models themselves, and found that they tended to lean a little bit to the left.
But what would explain that?
The training data, as well as the guardrails that are put in by the companies. If the modal source of information in the training data is, for example, the New York Times, and the New York Times is somewhat left-leaning, then you would expect that to affect the average answer you get on politics.
And by the way, we see similar kinds of biases in other settings, in different ways. My colleague Josh Tucker at NYU has done a really interesting study, published recently in Nature, about how in authoritarian systems, if you ask questions in the local language, you will get different answers than you will in English. For example, even with the American models, if you ask about Tiananmen Square in English, you’ll get one answer; if you ask in Chinese, you’ll get a much more pro-regime answer. This is not because the makers of ChatGPT want to have pro-regime answers about Tiananmen Square. It’s because so much of the training data that’s in Chinese favors the government’s position.
Similarly, Andy Hall did another paper where he found that in Japan, if you’re a moderate left-leaning voter and you ask the models who you should vote for, the models will recommend that you vote for the Communist Party of Japan. The Communist Party of Japan is not a major political player, so it was a bit of a puzzle. But the answer is that all of the party’s materials are free on the internet, while a lot of the moderate parties and mainstream sources are behind paywalls, so they’re less present in the training data. So it’s a combination of the training data plus different things that the model developers have done to deal with hate speech and other incendiary issues that may have moved the models a little bit to the left.
So how would you reassure people, or address the profound concerns they have about this question of bias? They might say it’s sort of a pick-your-poison situation. Either the companies have control and are manipulating the models for their own purposes, to suit their own political preferences, or, alternatively, the underlying data that the models are training on are tilted—they’re skewed. So don’t we have this problem coming and going? What would be your response to that?
I think that’s right. This is a feature, not a bug, of the nature of large language models, which is that they’re trained on a corpus of information. So it’s not even clear what the right answer is here, even though you see Democrats and Republicans rating them as left-leaning. Large language models are basically a brain that was developed off of the existing material on the internet, plus other sources.
But I want to emphasize the sycophancy point, because what is going to happen as people use these models more and more is that the models are not going to give consistent answers to every person on the same question. That might also be a kind of way out of this mess—that they’re more personalized. But if you care about polarization and people not having a shared information resource, then it cuts against that.
So from the standpoint of the implications for democratic engagement and participation, it sounds like one challenge is for users of AI who become very dependent upon it to recognize that it is not the all-knowing source of information they should rely on. That seems like a very tough hill to climb—or not?
It is. And I think it’s increasingly difficult, because people are turning to these chatbots for information, and so it doesn’t seem unrealistic to think that they will effectively replace Google search. Even in Google itself, of course, right now if you ask a question, you’ll often get the AI Overview first. So the stakes are pretty high as to how these models are going to behave.
There was always a problem, whether you’re talking about social media news feeds or search results: what went into the 10 blue links that Google would return wasn’t necessarily going to be representative on every issue. But at least you as the user could choose among those 10 blue links. Whereas now it gives you an answer. And so the stakes are quite high, not just in politics but for all kinds of information—medical information, consumer information. If they hallucinate, if there are errors, that is a big problem. And it’s a big problem, by the way, from the standpoint of election administration. Or if they’re systematically biased in favor of one party or another, or one type of source or another, that’s another problem.
(9:42) Let’s then turn for a moment specifically to elections. You mentioned deepfakes; you mentioned potentially erroneous election information. At the same time, in your article in that collection of essays, The Digitalist Papers, you in effect frame your piece as one that addresses certain common misunderstandings about the nature and the extent of the problem. Can you talk about that in particular? What are those misunderstandings?
There is no question that when it comes to synthetic media, as well as other forms of disinformation, there are millions or hundreds of millions of examples on the internet at any given time. And YouTube, which I have some conversations with, is overwhelmed with what we call AI slop. There is a lot of AI-generated content.
However, as we think about the political salience and importance of AI for, say, voting behavior—both whether people turn out and who they vote for—AI-generated content is actually a small share of what the average user is going to consume. And so the more that we panic about that 1% of deceptive AI content that might be in people’s feeds—and I’m making up 1%; some small share—the more that people are going to doubt the 99% of their feed that is at least not false, whether it’s true or whether it’s entertainment or something like that. My concern is that the more we freak out about the possibility that people will be duped by AI synthetic content, the more we create another and bigger democracy problem, which we sometimes call the liar’s dividend: people stop believing in true facts.
So from the standpoint of democracy, I think that’s actually the greater challenge. And that is what concerns me, because it also enables politicians to disclaim true things as being false. If you look around the world in the post-ChatGPT era, that is the phenomenon we see more often. Not that people are being duped by fake audio and video, though that does happen, but that politicians are using this as an opportunity and excuse to disclaim true stuff as actually being false.
Let me just be clear on one point, so that everybody knows precisely what we mean by synthetic content or AI-generated material. When we talk about that kind of material, we’re talking about extremely effectively produced, fantastical false depictions of, say, a politician’s behavior, or of something a politician says on mic or off mic or allegedly in close quarters. And I take it that as time has gone by—and this sets up my next question—the true from the false has become, as the technology has evolved, harder and harder to distinguish. Or is that not the case?
No, that is true. I could reproduce this podcast and put different words in our mouths, and no one would really be able to tell the difference. The technology is so good right now that, unless you are trained to look at it, at the cutting edge we can create extremely realistic videos that do almost anything. And we see this, of course, outside the political realm, whether it’s deepfake pornography or commercially available deepfakes—companies running advertisements with totally synthetic human beings selling products. So we’ve already reached the level where you can be deceived.
But what I’ll say is, if you look at the use of synthetic media in politics, and particularly the image-generation tools, most of it is being used not to deceive but to piggyback onto other narratives that are in the political system. So a lot of it is satire. If you look at the 2024 election, I didn’t think deepfake imagery really had much of a role to play in the presidential race, although there were, again, millions of examples of it. But you would see things like Donald Trump hugging dogs and cats, to pick up on the story in Ohio with Haitian immigrants, or pictures of Kamala Harris dressed in a Communist Party uniform. None of these things are actually intended to deceive.
But they are piggybacking onto false narratives to try to push people in a polarizing direction. And so the new tools are being used to achieve the same goals that consultants have always had, and in a polarized political environment, those could be bare-knuckle tactics.
But how would you rate the following concern, just to follow up on this? The technology evolves to the point that it’s very hard to distinguish the true from the false. Something that is not meant to be satirical—it’s meant to be persuasive—is crafted very cleverly against the background of polling data showing that it plays into perceptions of a particular candidate’s weakness. So something wholly fake is created, and it’s micro-targeted in various ways to voting constituencies. It is expected, if it’s done extremely well, to become an issue in the campaign. And news media organizations, by way of reporting on it, begin to duplicate references to it, show it on the air.
Over time, it would seem, in the right race, in the right way, at the right time, on the right issues, that could wind up being a quite significant factor—maybe just in the closing days of the campaign, when the effort to chase after it, to stop it, to rebut it for the public’s benefit is very difficult to achieve.
That is the worst-case scenario. But the role of the media in checking this is quite important. As you were saying, maybe the mainstream media rebroadcasts it, whether to criticize it or to affirm it. That is actually when we see disinformation in general, or deepfakes in particular, gain traction: when they get blown up and picked up by the mainstream media. I don’t want to understate the significance of that event, but the effectiveness of a strategic deepfake in the waning hours of a campaign is going to be conditional on whether the mainstream media checks it or not. It’s quite rare that you’re going to get some signal that’s sent, say, over Facebook or some other social media platform that reaches a certain level of virality and then is not checked by the mass media.
Now, people may not care, and this is part of my point about the liar’s dividend. It may be that just the fact that it’s out there, piggybacking onto other kinds of narratives, matters—that even if the mainstream media covers it and says it’s false, the coverage could repeat the charge, and that could be the effect of it. But while I do think this is a serious concern, I actually think that kind of surreptitious last-minute action is quite rare and not as effective as people think.
Just one last thing on this, because it obviously has captured my attention. Years ago, two political scientists at MIT did a study on the effect of fact-checking of negative advertising, and they concluded that viewers tended to zero in on the content of the advertising and not the content of the fact-checking. Now, wouldn’t that also significantly diminish the value of media fact-checking? They repeat it only to try to undermine it, but the repetition is more important than the undermining.
Exactly. That’s right. We tend to think, as social scientists or as lawyers, that when people approach content online they have a kind of Lincoln-Douglas debate going on in their head, weighing the evidence in favor or against. But the power of deepfakes, and of disinformation in general, is to play on people’s emotions. If the mainstream media repeats the deepfake, even to debunk it, it will have some effect. That’s going to be true of true content as well as false content. This is, I think, a reason to be concerned, but principally because, again, it’s piggybacking onto existing narratives.
The real division, I think, in the American public is not between those who believe something that’s true and those who believe something that’s false, but between those who care whether something is true or false and those who don’t. I think we on the outside tend to overstate how cerebral people are when they’re consuming political content.
(18:06) Let’s turn to another way that AI can be problematic, and that is if people are relying upon it for basic information about their rights to participate—in particular, let’s talk about elections. Where’s my polling place? What time do the polls open? What voter ID requirements apply that I have to satisfy to vote? I take it from your writing that there have been instances—I don’t know how widespread—when AI produces information that is incorrect and voters rely upon it to their detriment. How big a problem is that?
And let me just tack on to that question: Do you imagine that we’re at the point where, in this election, controversies over deepfakes or over that kind of problem—that is to say, communicating misinformation—could be the basis for a claim that an election was fraudulently decided?
Let me address the latter first. It could be part of the claim. I don’t think it’s necessarily legally cognizable. You’re not going to rerun the election because of false imagery or something like that.
I am concerned about this for the same reason I’m concerned about people’s over-reliance on the models for information generally. Now we’re moving from bias to hallucinations and truth—how reliable are these models? Let me give a little short history of what’s happened here. It started out that you would ask the models, like you said, “Where’s my polling place?” and they would sometimes base their answer on old information in the training data, so they could send you to the wrong polling place. In the 2022 elections, and certainly as we approached 2024, the models started refusing to answer those questions and kicking you out to authoritative sources like the National Association of Secretaries of State or Democracy Works or a few others. They didn’t want to have responsibility for the answers.
Now the models are better, though they’re not perfect, and some of the contemporary studies are trying to show that they still are flawed. The models are answering the questions, so the stakes are high. Election officials, I think, are legitimately worried that when people go to the models and rely on them for election administration issues—like you said, whether it’s how to vote or where to vote—they will give them some false responses. Sometimes, depending on the question, the models will have almost a disclaimer at the end that says to make sure to check whether this is right with your local election official, or something like that.
(20:34) Okay, so we’re at a point now, given the problems you’ve identified—and again, I think it’s important to emphasize that we are where we are now at a rapid pace of development, so these problems could be significantly more challenging, more pressing as that development occurs. But let’s talk about the policy challenges here: how to think about the regulation of AI, I suppose particularly in our case, but I think more generally as well, and how that policy challenge could be met to address these particular problems. So can you talk a little bit about that?
This is an interesting area, AI and elections in the United States, because as much as the topic of AI regulation does lead to some controversy in certain circles, and certainly the federal government has not been able to pass a systematic AI bill, most states have now passed some kind of legislation related to AI and political advertising, for example. They alternate between certain kinds of regulations or bans, and most of the states are passing AI disclosure bills, so that if you use AI in the construction of a political ad, you need to disclose it. For the reasons I was saying before, it’s not clear that disclosure does a whole lot in terms of protecting people from false beliefs, and the more that everything is disclosed as AI, the more that people are going to ignore it.
But at the same time, I think it’s a reasonable deterrent to have laws that say, at a minimum, if you’re going to start using synthetic imagery of your opponent in an election ad, you should have to disclose that that’s what you’ve done. Then, when it comes to election information, I think the platforms need to start going back to what they were doing before, to make sure that people are always looking at the authoritative sites of their local and state election officials, because there may be last-minute changes that the models are not picking up on. Take, for example, “Where is my polling place?” in an environment where we have these frequent redistrictings, where you’re going to be changing jurisdictions. There could be lots of polling-place moves, and you want to make sure that you actually know where you need to turn out to vote.
I think it’s important for them not just to try to get the best answer from their training data, but also to have some hard guardrails in there that refer people to these sources. Other countries are taking some more draconian action here. In Brazil, for example, they prohibit the models from recommending that voters vote for particular candidates. That, I think, might be a problem under the First Amendment here, but those are the kinds of things that we’re seeing in other countries.
(23:17) Let me just ask you a closing question, and that is: you talk about various attitudes toward AI and predictions about its future use and its benefits and its costs, and you create a range that goes all the way from boosters to doomers. I think that speaks for itself. As you think about the development of this technology—the success of efforts to regulate it, or the lack of success of efforts to regulate it, its continued extraordinary development—how do you see those trade-offs, of course, between costs and benefits? Are you more on the booster side or the doomer side, or how would you characterize your view of the future here?
As we’ve been talking about the effect of AI on democracy, we’ve been focusing principally on elections, but what we call a democracy problem could include massive unemployment as a result of the new AI economy. It could be AI-inspired cyberattacks on electrical grids on Election Day. All of the other—I’ll say garden-variety, but really serious—AI problems have an incarnation when it comes to democracy.
This is the summer of AI doomerism, even here in Silicon Valley, because of what happened with OpenAI’s agents escaping their sandbox and hacking into Hugging Face, which has now led to revelations that this has happened really thousands of times, sometimes on government websites. I think there is real concern right now that the model developers cannot control the technology that they are developing. Even they are advocating for pauses. OpenAI just recently decided not to release the next Astra model because they were worried about safety.
I am very concerned about the development of this technology, particularly the pace at which it is happening. And I think that it very well may cure cancer. It will have downstream beneficial impacts. Even in the election realm, election officials can use it to communicate with voters. Candidates who don’t have resources can now use AI to develop campaign ads and jingles and all kinds of things they might have had to pay consultants for before. I don’t want to understate the benefits, because I think they are going to be significant.
But this is an incredibly rocky time right now. These next two years are going to be absolutely central to charting our AI future, because the model developers say they are on the brink of what we call recursive self-improvement, where the models basically develop the next generation of models. If we don’t wrap our heads and laws around this process now, it is going to be too late in a short period of time. While I wouldn’t call myself an unrepentant doomer at this point, I would say that I am concerned.
On that sobering note, I think you are going to play a huge role in this debate, and thank you for all you do on it. Thank you very much for joining Executive Functions for this conversation today.
Thanks for having me.
