- Artificial Intelligence
- Congress
- Public Opinion
- Revitalizing American Institutions
Visiting Fellow Sean Westwood explores electoral reform, AI’s impact on American politics, and issues such as partisanship and polarization. He spoke to the Hoover Institution’s Chris Herhalt about his recent work exploring the impact and media coverage of politically motivated violence, how AI is turning polling on its head, and the rise of what he and his colleagues call “conflict entrepreneurs” in Congress.
Chris Herhalt:
Sean, thanks so much for taking some time to speak with me. I really enjoy your work both in your role at Dartmouth and for the Center for Revitalizing American Institutions. And I want to start out with the discussion you had with the folks from The New York Times. Regarding political violence, you said that the rate of politically motivated violence today is not really at the same levels that it had been earlier in American history, but that this activity still presents harms to the political climate. What do you mean?
Sean Westwood:
Any political violence is troubling. We would obviously want to live in a society where our politicians can do their job without risking their lives, their safety or their family's safety. But it's also important to contextualize just how much of a threat political violence is to the United States relative to what we've seen in the past. There are several researchers who've constructed incredibly alarming polls suggesting 20, 30, 40, 50 percent of the American public support political violence. And that's just not where we are. Not only is it vanishingly rare, but public support is, when competently measured, close to the margin of error in most polls.
In 1856, a congressman from South Carolina walked onto the Senate floor and beat a senator from Massachusetts nearly to death with a gold-headed gutta-percha cane. Admirers across the South sent replacement canes, some inscribed “Hit him again.” In 1898, a white supremacist mob in Wilmington, North Carolina, overthrew a duly elected multiracial government, murdered dozens of Black residents, and installed its own officials by force--the only successful coup d'etat in American history. In the eighteen months from January 1971 through mid-1972, there were over 2,500 domestic bombings in the United States, an average of nearly five per day. None of these events produced a second civil war. None ended the republic.
The point is not that these things were acceptable. The point is that the claim saturating contemporary commentary--that we are on the cusp of a civil war--is a claim that can only be made by people who have a cursory understanding of history.
This is largely undergirded by polling that often asks the wrong question. We don’t ask would you, yourself, commit political violence? Instead, we ask would you support someone else who committed violence—often with no context on the crime or the motivation---to help your party?
And that's a nebulous question that people answer very expressively. If you're mad at the Democrats or Republicans and someone asks, "Would you support violence against the other party?" It's not clear that your answer to that question is actually a genuine, sincere expression of support for violence and not just a way of saying, "Yeah, I'm really upset with the other party, and I want to signal my displeasure by answering a survey question in a way that maximally conveys my anger."
This then has the potential to become a self-fulfilling prophecy. When the media report that 40 percent of the public support political violence, that has a real risk of generating a false sense of support, or the false sense of a mandate in the mind of mentally disturbed individual. It could be enough to nudge someone on the fringes of society to commit violence.
We need to remember that American history is not a story of civic harmony occasionally disrupted by conflict. It is a story of conflict remembered with retrospective impressions of harmony.
Chris Herhalt:
Thanks, that’s really interesting. I want to move also to the paper you did with Mark Jacob and coauthors where [speaking of members of Congress,] you guys coined the term conflict entrepreneurs. I found this fascinating. Basically, you call them legislators who resort to personal attacks and insults in order to acquire additional media coverage beyond what they’d get as an ordinary run-of-the-mill member of Congress. Is that accurate?
Sean Westwood:
Every member of Congress wants recognition, power, and influence. There are basically two ways to do this.
One is to put in the time: spend years in Congress building a reputation by introducing legislation, co-sponsoring bills, sitting on committees, and dedicating time to the boring, unglamorous work behind the scenes. It's slow, and it pays off slowly.
But not everyone has the motivation or skillset to do the work of legislating. Some representatives don't have the policy background or education. Some don't have the staff. And some just don't have the patience to spend six or eight years building seniority. Instead, these folks take a shortcut. They start baiting the media by bullying others and making outlandish statements. They know that sponsoring a bill that will in all likelihood die before leaving committee does not reap anywhere near the attention of a well-timed X post. These representatives—what I call conflict entrepreneurs—really do get more media coverage than colleagues who keep their heads down working on policy.
But my work shows that it doesn't lead anywhere. It doesn't buy them more influence over policy—they can't turn the attention into actual legislation, even on the issues they say they care about. It doesn't give them an edge in elections. And it doesn't bring in the money you'd need to run for higher office.
This strategy leads to headlines and cable news segments, but not much more.
Chris Herhalt:
So that was my natural next inclination of things. It gets them the lead 30 seconds on pick your big broadcast network. But if it doesn't result in additional fundraising, if it doesn't result in any legislative success, is that worth it? What is the pathway? What is the motivation there then?
Sean Westwood:
It could be vanity. It could be that individuals just want to be important and to be part of the national conversation. Maybe they want to go to bed at night knowing that what they said on the floor of the House was on every cable news outlet and the pundits discussed it at length. But it could also be that this is more strategic. They may recognize that they don't have a future in Congress and are trying to build a kind of infrastructure to transition to becoming a pundit or hosting a podcast. It could be just a strategy of building a set of not constituents, but a set of followers.
Chris Herhalt:
There are personal examples of this that come to mind, but we don't need to go there.
Sean Westwood:
The most important thing to understand is that this happens in both parties, and that it is, despite what one might intuit from watching the news, genuinely rare. Ask an ordinary American what they think of Congress and you'll hear something bleak. People are deeply disappointed in the institution, and part of the reason is a mistaken belief: that Congress is hopelessly combative.
But the truth is that the vast majority of members never engage in the taunting and personal attacks that the average citizen has come to assume are routine. They go to Washington, they do the job, they meet with their constituents, they sit through committee hearings, and they return to their districts. But that is just not newsworthy.
What the public sees instead is a vision of Congress built on the behavior of a small subset of irresponsible members who simply do not take the responsibility of the office as seriously as we have every right to expect. This matters because it makes the actual work of governing harder to do, makes Congress harder to trust, and most damaging of all, it erodes Americans' confidence in democracy itself.
Chris Herhalt:
Let’s move on to the other aim of your recent work, which is examining AI’s impact on politics. I read about how last year, you built an AI agent that could fool online political polls into believing it was a real human voter 99.8 percent of the time. Given the pace at which AI has advanced since then, is it safe to assume that now there’s an AI agent that would be even more capable of deception and be even easier to build than it was when you embarked on it? And what does that mean?
Sean Westwood:
American intelligence agencies have long tried to influence elections overseas, and part of that effort has involved manipulating public opinion polling. It's reasonable to assume that if we are doing it, other intelligence agencies are doing it too.
The key point is that the tool I built doesn't require a PhD to produce. I'm a political scientist, and I was able to do it. I have to assume that trained coders in the Russian FSB or in Chinese intelligence can not only do what I've done but do it far better and at scale. That means, in effect, that almost anyone can interfere in almost any election—regardless of language or context—if they have enough awareness, motivation, and resources.
The ultimate question is whether this is actually happening. It's a frustrating question to ask, because these tools are designed to be undetectable. If we can't detect them, that could mean they don't exist—or it could mean they're simply good enough to evade the traps we've built. So, I can't say with confidence that it's happening.
What I can say is this: American elections are decided by very thin margins, and districts have been gerrymandered so heavily, that even a small number of synthetic poll respondents could matter—in presidential and House elections alike. Are all surveys dominated by these tools? No, absolutely not. Could a poor-quality survey contain a small number of these agents? Yes. And could that be enough to flip the top-line result? Yes.
Chris Herhalt:
Where did you get the idea to try and do this?
Sean Westwood:
It was actually in a conversation with one of my co-authors. We were talking about the effects of AI on public opinion. I realized that I was pretty sure I could build a tool that would use AI to answer surveys. It was a personal challenge: could I do it? I didn't expect it to work as well as it did. But once I realized that, I just put everything else down and wrote the paper.
Chris Herhalt:
So given all that, should we do something about it? Is there something the United States could do to prevent the use of agents domestically?
Sean Westwood:
I think that it's important to know that there's a gradient of survey quality. So, there are providers of survey data that are just overrun by foreigners, by fraudsters, by computer bot labs. We know who those providers are and we should be very skeptical of the data they produce. But there are also providers that have really tried to robustly approach the problem. YouGov is a good example of that. They're aware of the problem. They've built mechanisms to try and identify when it's occurring and because of their infrastructure, I think they're probably the only panel provider that can detect this in real time.
I don't think it's that we should abandon public opinion polling, and I don’t think we need regulation. Both of those things would be destructive. I think what we need to do is penalize companies that are doing a very poor job of auditing the data that they collect and reward those companies that are doing a good job. The open market and competition are the solution.
Chris Herhalt:
Could you tell us a little bit about the book you’re working on right now?
Sean Westwood:
An enormous amount of money has been spent in this country advocating for electoral reform. In 2024 alone, several states put ranked-choice voting on the ballot. But the case for ranked-choice voting runs into a couple of fundamental problems.
The first is that the evidence for its effectiveness is deeply mixed. In a meta-analysis I conducted—the work that led to this book—the results are, on average, negative or null. In other words, ranked-choice voting has the potential to make things worse, not better. For example, a lot of Americans already distrust elections; the recent mayoral race in Los Angeles is evidence enough of that. Ranked-choice voting makes a particular kind of outcome possible: a candidate can win the most votes on election night, only to lose the seat two weeks later, once the rounds of counting play out. That is exactly the kind of result that erodes trust further.
The second problem is that we don't even agree on what reform is supposed to improve. Most prior research has focused on things like representation, confidence, or satisfaction with democracy, while ignoring the negative side effects of these reforms.
This book is meant as a neutral assessment "Is voting reform really the path to a healthier democracy?" I think the answer is a resounding no.
Fixing American democracy is not simple. There is no magical fix, no single lever to pull. It's the slow, difficult work of nominating and supporting better candidates. It's reforming the rules of Congress. It's pushing back against gerrymandering, and against the rules that have allowed extremists to control the legislative agenda. All of this is hard. All of it requires action from elites, and all of it unfolds over very long timescales—but that is where we actually have purchase, where there are levers we can genuinely pull.
It's a mistake to think groups can step in with hundreds of millions of dollars and fix the system through a couple of referenda. The best evidence for that is what happened in the states where these measures were on the ballot, where the spending gap was enormous. In Oregon, supporters of ranked-choice voting outspent their opponents by nearly 700 to 1—the opposition spent less than $20,000—and the measure was still decimated at the polls.
There simply does not appear to be public demand for this, and to me that signals something important: voters already understand that there is no easy fix. They know democracy has its problems, but they also know there is no silver bullet to be deployed.
Chris Herhalt:
Right. Now I want to move on to my final question for you. When you talk about acts of violence, there's been two instances to date, but one I'll focus on one. There was a mass shooter who repeatedly consulted a large language model in the days, weeks, or months leading up to an attack that resulted in multiple fatalities. And in the one instance I'm talking about, the operator of that large language model saw enough of this person's behavior to become concerned, to hold meetings and to actively discuss whether to alert the authorities to these activities, but ultimately chose not to do so. That's not the last time we're going to see that. So, I'm wondering what you think?
Do you think large language models have a responsibility in that instance, or do you think that requiring them or suggesting to them to do so is a violation of some kind?
Sean Westwood:
It's already the norm for social media companies to alert the police or the government when they observe communications on their platforms that plan or contemplate violence. I don't see why AI should be treated any differently. If someone is using an AI tool to plan a crime or map out illegal activity, I see no reason they shouldn't expect a commercial vendor to monitor that and report it to the authorities.
In fact, I'd argue that this is exactly what consumers expect. People don't want these tools generating child pornography. They don't want them explaining how to build biological weapons. They don't want them helping someone strategize how to get away with murder. If companies fail to prevent this, they risk losing the public's trust—and once that trust is gone, it becomes far harder to do business or to recover.
There's a further risk, too: neglecting these responsibilities could invite overly aggressive government regulation. What we should be aiming for is to persuade these companies to act responsibly on their own and to be good social actors.
