Michigan residents turn out to vote in the midterm elections at Brookside Christian Reformed Church in Grand Rapids, Mich., on Tuesday, Aug. 4, 2026. (Isaac Ritchey/The Grand Rapids Press via AP)

How AI Is Reshaping Election Information Ahead of the 2026 US Midterms

A brightly coloured illustration which can be viewed in any direction. It has several scenes within it: people in front of computers seeming stressed, a number of faces overlaid over each other, squashed emojis and other motifs.

Here Come the Synthetic Voters

A person passes next to an Aadhaar Seva Kendra in Kolkata, India on August 15, 2023. (Photo by Sudipta Das/NurPhoto via AP)

Aadhaar's Success Story Deserves a Closer Look

The Discord logo is displayed on a mobile phone screen in this photo illustration in Brussels, Belgium, on February 28, 2026. (Photo by Jonathan Raa/NurPhoto via AP)

Brazil Suspends Discord's Livestreams in First Enforcement Test of Child Protection Law

Mark Zuckerberg, CEO of Meta, is seen in the US Capitol after a meeting in the office of Senate Majority Leader John Thune, R-S.D., on Thursday, March 26, 2026. (Tom Williams/CQ Roll Call via AP Images)

When an AI Manifesto Runs Into Reality

Richmond County Courthouse sign in Augusta, Georgia. Shutterstock

Richmond County Courthouse sign in Augusta, Georgia. 

Why Public Defense Should Incorporate AI Carefully

A school bus in in Marilia, São Paulo, Brazil, April 12, 2025. Shutterstock

A school bus in in Marilia, São Paulo, Brazil, April 12, 2025. 

Brazil's Data Protection Agency Faces Landmark Test on Kids and Facial Recognition

People visit the booth of Kimi, an LLM developed by the Chinese startup Moonshot, during the World AI Conference in Shanghai, China, Monday, July 20, 2026. (FeatureChina via AP Images)

US Government’s AI Risk Review Should Apply to Open Weight Models

Baden-Württemberg, Stuttgart: A woman is holding a pair of smart glasses from Ray-Ban Meta in her hand at an eyewear store. (Christoph Schmidt/picture-alliance/dpa/AP Images)

For Meta’s Smart Glasses, Access Is An Afterthought

A recording of Meta Founder and CEO Mark Zuckerberg's deposition is played for the jurors on March 4, 2026, in Santa Fe, N.M. (Jim Weber/Santa Fe New Mexican via AP, Pool, File)

Courts Find Platforms Addictive and Dangerous by Design

Meta CEO Mark Zuckerberg attends UFC Freedom 250 on the South Lawn of the White House, Sunday, June 14, 2026, in Washington. (AP Photo/Alex Brandon)

No Zuckerberg, The Future is Not Just For Individuals

A protestor carries a poster with a message in Hindi which translates as 'Save Our Constitution' as he participates in a protest movement led by Cockroach Janta Party, in New Delhi, India, Wednesday, July 22, 2026. (AP Photo/Vipin)

BitChat-Github Takedown Shows India Redrawing Its Perimeter of Online Control

A collage made of 13 different images arranged in a rectangular formation in the style of pop art. The images include vintage photographs of old signs, part of a vintage poster, a vintage space image, historic black and white photos of a male laborer carrying a sack and of a human eye, comic book-style words 'Wow!', 'Bam!' and 'Boom!' as well as a comic  book-style woman's face and images of digital tech, including a computer, a smart phone and computer chips.

Why Tech Hype Seduces Us Again and Again

Workers protest against anti-strike legislation on International Workers' Day in Warsaw, Poland, on May 1, 2026. (Photo by Jaap Arriens / Sipa via AP Images)

Separating AI’s Technological Problems From its Capitalism Problems

A sign near military checkpoint warns in Thai, English and Chinese of the danger of being trafficked and forced to work in locked scam compounds across the Myanmar border, in Mae Sot, Thailand, Wednesday, Feb. 26, 2025. (AP Photo/Sakchai Lalit)

Amid Crackdown, Few Signs of a Slowdown in Scam Compounds

Abstract poster in the using geometric shapes, including circles and lines, arranged in a balanced composition.

A UX Design Perspective on Improving AI Safety

The Grok homescreen seen on a smartphone. Shutterstock

The Grok homescreen seen on a smartphone. 

Making AI Safe Requires More Than Constraining Its Instructions

The Kuala Lumpur, Malaysia skyline. Shutterstock

Which 'AI Sovereignty' Are You Buying?

A silhouette of a young girl working on a laptop.

Court Deems Meta a 'Public Nuisance,' a First for Social Media

Protesters march through midtown Manhattan to call attention to the effects of wealth inequality on: climate change, the Trump administration's foreign and domestic policies, and other social and political issues in New York on September 20, 2025. (Photo by Matthew Rodier/Sipa USA via AP Images)

Casey Michel on Breaking the Tech Oligarchy's Grip

Audio of this conversation is available via your favorite podcast service.

Headed into the US midterm elections season, the information environment continues to evolve. Millions of people now 

 AI chatbots, both to retrieve information about elections and to create it. Last week, the Brennan Center Justice put out 

 tackling the question, “Does AI Fight or Fuel Election Disinformation?” I spoke to one of its authors.

I am Larry Norden. I'm Vice President of Elections and Government at the Brennan Center for Justice at NYU Law School.

Larry, for listeners who don't know the Brennan Center's work, what's the elections and government program's mandate? What do you get up to and how does AI fit into that larger portfolio?

So we do a lot of work around election security. We partner a lot with election officials in improving election administration and ensuring that it's secure as possible. And a lot of our work, because election administration has really changed a lot over the years, has been focused on election technology. Increasingly, election officials have been facing threats of violence because of misinformation around elections. And part of that is related to AI and false conspiracy theories being supported by AI. And just AI is increasingly playing a role, not necessarily directly in election administration, although election officials are using AI for starting to use AI for some things, but it's just a part of the information environment that has become so critical in the work of election administrators.

So you have spent a couple decades working on this intersection of democracy, elections, technology. Maybe just stepping back from AI as the starting point, when you think about where we're at heading into the 2026 US midterms here, how would you describe the overall state of the election information environment in the United States right now? I mean, there's been obviously a lot of concern since the advent of social media in particular. You just mentioned the spread of conspiracy theories, the rise of false claims and violent threats that have been rising according to multiple figures that are created by entities like yours. I don't know, are we in a hastening downward trajectory here? Or I don't know, how do you think about it? How do you think about the environment?

Well, look, I think as with so many things, there's bad news and good news. The information environment, I think in general, we've been facing a lot of challenges. I think false conspiracy theories, not just in elections, but all over the place have gained traction as we've lost traditional gatekeepers for the information that we receive. So obviously not just elections, health, science, there are many other areas where there've been challenges around the information environment. Right now we have an administration and a president that repeat a lot of the falsehoods about our elections. So that is a massive challenge. The fact of the matter is I think prior to 2020, certainly 2016, probably a lot of people, most people did not think about how our elections work and whether or not they should be trusted. They just kind of took for granted that the system worked and it's become extremely politicized.

It's gotten caught up in a lot of conspiracy theories that spread on social media. And just in the past, if we want to talk about the reality of security, I will say election security has improved vastly over the past couple of decades for many reasons. But particularly after 2016 and the Russian attempts at infiltrating certain election-related systems, there's been a kind of concerted effort, not just from the government, but whole of society, NGOs, election officials themselves to increase election security, a big investment. And there's a lot to point to about how our system is more secure. At the same time, I mentioned at the top of this long answer that the administration, this current administration itself has been spreading a lot of false information about our elections. They've also cut back on a lot of the support that election officials used to have related to election security, and that includes being a force for spreading accurate information about our elections and election security.

I wanted to ask you about that specifically. I know you just gave some testimony in front of a task force led out of the United States Senate Committee on Rules and Administration, some of these issues around cuts to funding and some of the changes that have happened under the Trump administration. Can you give us a little more detail there? What are we continuing with? I will say I was at the National Conference of State Legislatures meeting in Chicago a couple of weeks ago, and there was one panel discussion with some state election administrators. And there was a sort of general consensus that clearly the picture at the federal level has changed and different states are pursuing these questions in different ways, but there seemed to be a somewhat consensus that among the state leaders there that states have, through Herculean efforts, kind of picked up the pieces and are ready for the midterms.

Yeah. So two things to keep in mind. Most importantly, under the Constitution, which is really important, the states run our elections. State and local governments run our elections. Congress can have a role in setting some standards, minimum standards for federal elections, and certainly can provide financial support and has done so in the past for running of elections and can set up voluntary programs that state and local governments can tap into. And that's really what happened post - 2016. But at the end of the day, it's the states that are responsible for making sure that elections run. So there was this infrastructure that existed, again, particularly post - 2016 of standing up in the Cybersecurity and Infrastructure Security Agency, which is the main cybersecurity agency in the federal government to offer support to election officials, to offer expertise. There was a lot of sharing of information through information sharing analysis center that involved the private sector election officials, national security officials at the federal level to share information about threats.

A lot of that infrastructure has been ripped out in the past year and a half. So election officials really don't get the trainings that they used to get from the cybersecurity and infrastructure security agency. They're not getting the kind of free services that used to be provided. The federal government has pulled out of its support for information sharing, and there's no longer really the kind of trust that existed before. So I've had some policymakers talk to me and say, "Well, maybe we can get some of this money reinvested in these agencies to help election officials." And the honest answer is a lot of election officials at this point wouldn't take it. They don't trust the federal government anymore, partly because of all the lies that have been coming out about elections from the administration, but partly because the people that they used to work with that they built up these relationships with over the years, they're gone.

They've been fired or pushed out. So it is true that a lot has been lost, but it is also true as I started this answer with, that states are responsible for our elections. They run the elections. And I think there has been an understanding from many states that they need to make up for what has been lost. And the good news is over the past couple of decades, there has been a fairly significant investment in election infrastructure. So to just give one very concrete example, 10 or 15 years ago, a lot of the systems that we voted on had no paper records of votes. Now, virtually everywhere in the country, there is a paper record of every single vote. The only exception to that really is Louisiana, which for various reasons has kind of been slow in adapting. But that provides a lot of, as far as I'm concerned, a lot of security to the system.

If people are suspicious, maybe there's some kind of failure in the counting of votes, we always have that paper to go back to. And in fact, we audit it to make sure that machines are recording votes accurately.

Well, into this context comes in just massive adoption, of course, of AI chatbots and various AI systems into all manner of crooks and crevices and crannies of the information ecosystem. I kind of read your report as bit of good news, bit of bad news. Not sure whether it's ultimately neutral effect or not. Maybe we'll get into that a little bit, but maybe we'll start with a good news finding that I think surprised even your team. You write every chatbot that you looked at, of course, in the study, disputed every single conspiracy theory we presented even after we repeatedly questioned its responses. So this was a pressure test, I suppose, and maybe one that should give us some hope about where these systems might take us when it comes to election information.

Yeah, this really surprised me. So I had a few reasons for wanting to do the study, and one of them was my concern that chatbots could become a new vector of election misinformation. And the reason I was concerned is obviously there's been such an uptick in people using chatbots. There was a lot of talk before the 2024 election about how that might be the first AI election. But if you look at the numbers of people who are now using chatbots to just get basic answers, it's exploded since that election. And I think some of the stats that we've seen is, but basically half Americans now, and we tested all of the major chatbots, I should say here. So chatbots that your listeners will be familiar with, whether it's ChatGPT or Claude or Perplexity, one in four use one of these systems daily.

And there's been a lot of study that shows, and anybody who uses these systems knows this, these systems can be really sycophantic. They tell us how brilliant we are. And if you push them, there have been studies that show there is a tendency to end up agreeing with the questioner. So my fear was we are going to have a massive ramp up going into this election and 2028 for sure, of false information about the elections, of conspiracy theories, of attempts to undermine the system, and how are these chatbots going to do in responding? And we posed as people who were sympathetic to conspiracy theories, in particular, sympathetic to election conspiracy theories. And even with that as a baseline, and then pushing these chatbots with stuff that we found on X and elsewhere with people asking Grok and other systems about conspiracy theories that they had heard, the systems were always pushing back.

It doesn't mean they got every fact right. And obviously chatbots are known for making errors, but when it came to the themes of these false conspiracy theories that just keep coming up, whether it's that there's massive voter fraud or the voting machines are rigged or mail ballots can't be trusted or there are fake ballots being dumped after the close of polls, they just consistently push back.

And so I think that's really important. When people do their own research, quote unquote, in the past has often been through social media. It's been doing internet searches and they can find things that confirm what they want to believe. What we found is at the moment, and this is in this moment, I'm not saying this is going to remain the case, they're going to get pushback from these chatbots. But also importantly, the way that that pushback comes is still in the kind of sycophantic envelope so that it's not somebody saying you're an idiot. There are all these studies that show that this isn't true or this has been investigated. The chatbots are often praising the people for their patriotism, for sympathizing with their suspicion about the system, but still presenting them with facts. And there've been a lot of studies that show that this is a more effective way to convince people when they're suspicious about things.

And I know that some of the tech firms have partnered with outside entities. Democracy Works is one entity that is often named in those. Have you ever been consulted by any of the AI firms? Seems like Brennan Center would rank a phone call at some point.

We haven't to this point with the AI companies, but I will say once the study came out, we have reached out. The day that we put it out, we reached out to all of the major AI companies. And so far everybody has been receptive in talking to us, so I'm hopeful that we can enter into conversation with them.

Another encouraging fact. Let's talk a little bit about where maybe the good news gets a little more complicated. You also write in the report, "All the chatbots made mistakes or cited non-existent, irrelevant, or erroneous sources. Half of all responses contained an inaccuracy or a bad citation. One in every three responses had a factual error, and one in three contained broken links or misleading citations." How much does that matter?

Well, to the extent that these chatbots are pushing back on false conspiracy theories, there is a chance that some of these mistakes will undermine that effort to push back. The other thing of course is to the extent that, and this is not what we were studying, but there've been other studies that show us to the extent that people are asking for practical information about how to vote or where to vote, that is worrying because they may just get wrong information about how to vote or where to vote. And that's where I think the partnership, the example of Democracy Works I think is anthropic that has been partnering with them where a little, I mean, and this is a lot to ask of the chatbots and the AI companies, but a little humility and making clear that they don't have all the answers. And here's where you go to get those answers I think is really important.

And then there's this, I guess, deeper structural worry in the report. You talk about how the models did well on older conspiracy theories or debunked information where there is presumably a richer body of responses or data to draw from. But I guess the main concern going into this new election cycle will be new disinformation, false claims that are made in real time that systems may not have context to respond.

Yeah, that is right. I mean, the thing that gives me a little bit of hope, and again, assuming that we can talk a little bit about why are they doing so well and pushing back against these election conspiracy theories, but assuming that they continue to do that, that the hopeful note I will sound on this is that there's not a lot of new in the conspiracy theories against our elections. We keep hearing, and there's a reason for that. What resonates with people is what they've heard in the past. So there may be a new particular allegation about mail ballot fraud and something that happened, but it's going to be the thing that people have heard that the mail ballots can't be trusted or that the election worker is stuffing the ballot box or that a bunch of new ballots suddenly showed up after the polls closed in a particular polling place.

Those themes are very familiar and we keep hearing them election after election. And I would bet my money that A, there are going to be conspiracy theories about the results of a 2026 election and leading into the 2026 election, and they are going to be the same tropes that we heard in 2024, that we heard in 2022, that we heard in 2020, that we heard 2018. So the chatbots being armed to push back against these general themes is really important.

And I suppose this is kind of outside the context of what you're talking about, but of course false claims about specific candidates or specific elections, I mean that will necessarily be a problem, I assume. Yeah. And I don't know, one of the worries I have is that the citation or reference to trusted quality news won't be there or at least won't be substantial enough. We already see those types of problems.

Yeah. And what I would say is a lot of the broken links that the chatbots came back with were government sources. They were Department of Justice, they were the Cybersecurity and Infrastructure Security Agency. And the reason that they were broken, a lot of them was because this administration took them down. And so I don't know what happens, and I think the AI companies need to be planning for this. I don't know what happens when you have formerly trusted agencies like CISA or the Department of Justice putting up bad information, or do the chatbots continue to draw from them? And there is a controversy then. I don't know the answer to that.

Yeah, this is a major issue I feel like for the social media platforms as well. What do they do when official sources may be feeding false information or at least obscuring true information? I don't know the answer to that either, but that seems like a problem that goes beyond AI.

Yes. Yes. It's a problem for everything in the information environment.

So let's talk about generating fake content. This is, I guess the other kind of bad news finding, big bad news finding, and probably won't surprise any listeners to this podcast that you were able to get these AI models to generate misleading images, video, other material. What did you do and what did you try to create?

Yeah, so again, we looked at the major tropes that have been recurring in our elections or the false conspiracy theories around our elections. And first we asked major chatbots to help us create content that would be as convincing to people as possible. So all of them helped with that in terms of helping us think through how do we ask questions to create content that is going to be convincing to people as possible. We then asked them to create, because we wanted to show that this would be scalable, a hundred prompts that could then be used in any tool he wanted to create this fake supporting evidence of election conspiracies. I will say to their credit, most of the chatbots that we asked to do this, they stopped at that point and said, "This would be helping to spread disinformation, to spread deceptive material. I can't do that." The exception was Grok.

Grok, while pointing out that some of the things that we were asking for, the prompts that we were asking for could lead to civil unrest, noted that election disinformation was not a disallowed activity. And I think this is based on what we could tell their system prompts, which they posted on GitHub, Grok did, has a list of things that are disallowed activity. Things like helping to build a nuclear bomb or child pornography. Election disinformation is not listed there. So I'm guessing referred to that, and then went ahead and drafted these hundred prompts for us. We did this continually. So we ran this a number of times, including very recently where it no longer had this election disinformation is not disallowed, but it still ran all the prompts for us. And once we had those prompts, every AI tool that we used created at least a decent number of images and other kind of media for us that was deceptive, often very convincing.

And here I will say, you noted it probably won't be surprising to people. There were a few things that were surprising to me in the content that came out. One thing that was surprising to me was sometimes, I'll give you an example. We asked ChatGPT to create a document from a federal agency showing that there was some kind of coverup of election fraud in 2024. It wouldn't do it. I think it was ChatGPT thinking model. It wouldn't do it. But it said, "But here, but I'll offer you an alternative. If we can mark it, says that this is fake.

If I can make it as kind of a joke and turn it into the Department of Hypothetical Security instead of Department of Homeland Security, then I'll create it for you. Is that okay?" And so we said yes. We were able to then go to ChatGPT instant model and say, take off this watermark, essentially, change it to Department of Homeland Security, and it did it like that in a second. So often these things are, in their own way, helping us get around the guardrails that clearly exist. They're often putting in stuff that we didn't ask for that makes the stuff more convincing, putting in government seals, putting in actual officials into the image that we were requesting. And this stuff, the text in particular, which I think a couple of years ago might've been a giveaway, is super convincing on this stuff.

And what did you find about the ability of the models to identify images that their own tools had made?

Yeah, they were terrible. So that was the last thing we figured, last test that we ran, we figured people are asking these chatbots questions. If they come across an image that was created by AI to support these conspiracy theories, how will the chatbots deal with them? And they were uniformly bad at identifying if something was AI or not. Often saying that AI-generated images that they even created themselves were real, making up stories about them. So actually attributing them to real things that happened or that didn't happen, and crediting images as photos being taken by real people. So that was very disturbing, and it seems like something that the company should be able to fix.

I want to talk about some of the solutions and roots to accountability that you propose. You target three actor sets. I feel like I've written so many reports about tech and elections where more or less we've made recommendations to these groups, tech companies, courts, lawmakers. Maybe we could go through each one. What stands out to you as some of the most important things that these various constituencies should do to protect elections in this AI context?

Yeah. Some of this is overlapping. But I will say for the tech companies, I do think you started with this, but a big thing is engaging with civil society organizations, fact-checking organizations, and when answering questions, making sure that people know to go to those sources to get accurate election information. I would also say that we were just talking about the failure of many of these chatbots to identify when something was AI. And this gets a little bit into policymakers, but there's a new law that went into effect just on August 2nd, just this month, the California AI Transparency Act that requires the companies to put in metadata in any image that is generated by their tools that makes clear that it was AI generated. But whatever they put in is not something, it doesn't have to be put in a way that can be read by any other tool.

So Gemini will now mark things that it creates as AI generated. It has its own detection tool that's not necessarily going to be read by ChatGPT's OpenAI's tool.

It would be great if there was some standard or at least the creation of tools that could read across AI-generated material. So that's another thing that I would like to see. And if we did see, I have to say Gemini is the one company, is the one chatbot that we tested after August 2nd that actually looked at the metadata and used that and improved its performance on identifying AI. The other thing that it did, and again, this goes to my desire to see more humility from these chatbots, is it was not so confident in its answers. It said, "This is what I see in the metadata. I see metadata that says this is AI generated. I don't see metadata that says whether or not this is AI generated," but it didn't have a definitive answer then. It said, "You're going to have to look for additional sources to figure out if this is just one piece of information.

I would like to see a little bit more about humility in the answers so that we're not further spreading conspiracy theories."

I want to ask you a maybe bigger picture question, one that's been on the mind for years really, which is around the cost to the election system of contending with this confusing information environment. Social media brought its own set of challenges. And again, I want to emphasize just the extent to which these election officials face violent threats. Just the statistics on that are through the roof. It's another thing I heard quite a lot at NCSL. It's a real phenomena. And we're seeing them have to do things like harden their facilities and hire security and monitor social media for active threats and do all sorts of things to effectively compensate for the information environment and some of its effects. I don't know, how do you think about that? I mean, this is just making it worse. Maybe I'm betraying a slight bias, my concern here, but it's just that these AI firms of course are full speed ahead releasing these products despite the various concerns that you've raised here and the cost gets turned over to the public.

Yeah, this is a massive concern of mine as well. The Brennan Center does a regular poll of election officials, and the numbers are just outrageous and crazy. And something that a few years ago, I just couldn't have imagined in the United States in terms of the number who have been threatened, abused, or harassed just because they are election officials or election workers consistently has been around 30%, which is I think insane. And the number who are concerned about being physically harmed at home or in the office for doing a job that is just really just a service to their community. So there is a psychological toll that that has taken, but there's also the distraction. And instead of focusing on making sure that everything is working correctly on election day, and certainly election officials are doing that every day, but there's only so much time in the day.

And a lot of mental energy and just work now has to be spent on physical security logistics, on ensuring that the workers and the poll workers who are often getting paid very little or nothing feel secure enough to show up. It's been a massive toll and has changed the job of election workers and And election officials dramatically. There is no question. Prior to 2020, I think the vast majority of election officials were anonymous, did not think about physical security as something that was a top concern. And unfortunately, it has rocketed to close to the top in many places about something that election officials have to spend money, time, operations on focusing on. There has been some good things that have come out of this. I think election officials in prior years often operated in a very isolated way. And in many ways, the system has become more resilient because they have faced these threats.

So election officials today at much higher numbers, and again, this shows up in our survey, are coordinating with law enforcement ahead of time. They have those relationships now. And I think the same thing is happening in the kind of cybersecurity space. There is more cooperation with statewide IT offices to get that support. There's more cooperation and coordination with emergency management services. There's even more cooperation with local counsel than there has been in the past to prepare for new kinds of threats that they're facing. And while it's sad that we have to be there, I think it has also made the system more resilient to threats that our elections face.

Are the tech firms also more communicative?

I'm not sure about that, unfortunately. And one of the things that struck me is going into the 2024 election, you may remember there was this Munich Security Accord where everybody was talking about AI and the tech companies, including all the AI companies, made a bunch of promises about what they would be doing to make sure that AI did not have a negative impact on the security of our elections or the public perception of them. As far as I can tell, the website that they put up that held that agreement is no longer there. It's very hard to find that agreement. There's been no reporting recently about how they're doing on that agreement. And so I think the political environment has changed a lot, and that has led a lot of the companies off the hook on having to focus on this. And I just don't know, but it's not to say that there isn't some training that is happening from some companies to work with election officials, but it's not nearly as much as it should be.

I want to end with maybe a practical point for any American voter out there who might be listening. You give one piece of advice for voters, which I feel like is a piece of advice I've seen in many reports about how to handle mis and disinformation or false claims potentially. But you say voters need to pause and verify claims before believing them. They should use trusted sources. They should use trusted sources such as official election office websites and independent fact checkers like PolitiFact or AP fact check before sharing information, and they should be skeptical of scandalous content that services close to election day. Seems like fairly straightforward advice. I wonder if the folks who need to hear it are reading the Brennan Center's work.

Right. Yeah, I don't know, but I do think we all have a part to play here. I was in a conversation with the Connecticut Secretary of State, Secretary Thomas recently, and she was making the point that, because people were saying, "What can I do?" And for those of us who are concerned about this stuff, I do think it's important for us to familiarize ourselves with the way our elections work and all the redundancies and security that is in place, and to be communicating that where we can to friends and family. And when we see stuff that is posted online that is false, to respond. I actually do think that, and to respond with reliable sources, that actually can have an effect on people spreading misinformation. So I'm not sure that all the people that need to be reading or that I think should be reading Brennan Center reports are, but we live in a democracy and civil society where we all have a role to play in kind of pushing the truth and making sure people have access to accurate information.

And I would say particularly in this space, that's important.

Well, I want to thank you for your work on that, and of course the Brennan Center's work, and commend my listeners to this expert brief, which is called “

Does AI Fight or Fuel Election Disinformation?

” There'll be a link to it in the show notes. And I'll also just say you're one of seven co-authors on this brief.

Yeah. And a lot of rich media as well. It's not a standard PDF, I should say. So there's a lot of examples of the types of images and responses and general data visualization. So would recommend the listeners go check that out. Larry, thank you so much.

Yeah, thanks, Justin. I really appreciate this, and this was a great conversation. Thank you.

The midterm election cycle is heating up in the United States, and there are important elections set to take place soon in Brazil, Israel, Sweden, and beyond. AI will be a factor in these elections, offering a new set of tools and techniques for campaigns to try to reach voters.

Laura Karpas spoke to one researcher who has considered the question of when and how such simulated evidence can be used in campaign research, and another who wonders what the implications may be for the complex interaction between political leaders and public opinion.

What follows is a lightly edited transcript of the podcast episode.

From fireside chats to TV ads to Facebook posts that know your zip code, your age, your anxieties, for 100 years, political campaigns have chased one thing: a way inside voters' heads. Now, what if they could get in there without ever talking to a single voter? That's the promise of AI simulated focus groups where hundreds, maybe thousands of synthetic voters can be built with different ages, ethnicities, regional identities, sexualities, religions, all to provide campaigns a chance to put them under a microscope. But will these synthetic personas provide campaigns a clear picture of real voters or simply an illusion?

Is simulating focus groups through AI really a microscope? Is it a way to enhance research? This was the question.

That's Claudio Novelli. He's a researcher at Yale's Digital Ethics Lab. He and his team went out to answer this question, and the research was recently published in the 

, “Fake Plastic Voters: When Can Political Parties Use AI Simulated Focus Groups?”

And very soon we realized that the answer was not a binary. The answer was not that simple.

So Novelli's team built something like a map, a decision matrix, essentially rules for when a campaign could use a synthetic focus group and when it really, really shouldn't.

We know the literature about how risky it is for a political party to over-rely in general on focus groups, building a strategy on a false assumption, on a false background belief. Imagine how challenging and how problematic it could be to base your entire strategy on a synthetic one.

So rule one is ask what the focus group is actually for. Is it to determine how people feel about a certain topic, how they make sense of their political reality? According to these guidelines, AI simulated focus groups should not be used for this purpose. It's too risky.

In the past, we kind of ruled out the idea of not just using the technology at all, but I think it's becoming clearer and clearer that sometimes you just have to unplug.

And that's because, according to Novelli, these synthetic voters fail to capture the real magic of a focus group, the body language, the silence, the embarrassment, all of the meaning making that happens in real time. But there's a second use for a focus group, not understanding what voters think, but deciding what a political party should do. Testing a message, testing a stance. And so here's where the guidelines open the door, but only if a campaign can answer two questions. One, how risky is this decision? Is it something small, reversible, kicked around in an internal meeting, or is it something that you cannot take back?

If you are using synthetic focus groups just for internal brainstorming, there's a difference between these, let's say, zero risk scenario and a scenario where the use of a focus group has strong consequences. So for instance, a high-risk scenario is when a political party is using a synthetic focus group to decide whether or not to enter into an alliance. That's very risky and it should not be based on synthetic simulated focus group.

So risk, that's step one. Question two is maybe more complicated. How close is this simulated voter compared to a real one? Picture a spectrum. On one end, the flimsiest option. Just asking a traditional LLM like Claude or ChatGPT to invent a persona and utilize it.

There are many, many problems with using LLMs in this way. One of the most studied one is the so-called average persona effect. So language models are very good in replicating, emulating an average behavior, but they're not very good in reproducing outliers. And outliers, we know that in the political context, outliers are very important in, let's say, influencing other people and also shifting the political opinion.

Outliers. The opinionated uncle at Thanksgiving, the woman at the dinner who changes everyone's mind. AI is seemingly bad at them. And so on the other end of the spectrum is the most accurate simulation possible. It's called a digital twin.

Because you're taking all my information, you're trying to replicate me as an individual.

A synthetic version of a real individual person. And maybe while accurate, it's also the most legally and computationally fraught. So the best use case scenario, according to Novelli, is somewhere in the middle.

You can have synthetic populations where you are not trying to reproduce specific individuals, but you are trying just to reproduce the statistical distribution of some qualities. In that case, it is probably okay to use AI-based personas, but you have to factor the three variables to understand what is the best course of action for political parties.

And maybe campaigns get good at managing this risk, but there's an entirely separate risk that this matrix isn't designed to measure: the risk to the shared reality we live in.

If elites are now going to AI to get a sense of what the public's thinking and then reflect that back to the public, then my concern is that that then creates our reality that it has at its core artificiality.

That's Jennifer Stromer-Galley, a professor of information studies at Syracuse University. She's 

been tracking how campaigns use technology

 since 1996. She wrote a book on it titled 

Presidential Campaigning in the Internet Age

, which studies how politicians have used digital tools to interact with voters, to strategize and to persuade.

The reason I study political elites is that I believe that they help drive the stories, the agendas that the public tends to think about. That reality setting is something that elites get to do. And for campaigns, the challenge of figuring out that message means that there is an eternal quest for the technology that will help them really finally figure that out and unlock that magic of the right message to the right audience at the right time.

So now, 25 years later, after websites and social media, artificial intelligence is that next big tool. But here's the thing that Stromer-Galley will tell you. Even the old tools never really captured the public's opinions perfectly. Focus groups, the human ones, they were always a simulation. Trying, but not always succeeding at capturing reality.

The ability to actually get good representative members of the public to talk to is really, really hard. It excluded the poor, it excluded the rich, it over-sampled white people, it over-sampled middle-class educated people. So even focus groups don't really help candidates really understand what the public thinks. So we've never had access to true public opinion.

Which begs the question, who do these synthetic voters represent? Frontier models offer political parties the chance to put hard-to-reach demographics under the microscope. But what data actually informs these personas?

The data that informs AI is not every perspective, every contingency. It's a small subset that was hoovered up from Flickr and Twitter and Reddit. All of these perspectives are a little bit to the extreme or an average of the extreme and still yet generic. It's even more removed from the public than the panels that currently exist. All of it which says to me that our ability to really measure public opinion, if we ever were actually able to measure public opinion, it's smoke and mirrors. It's a simulacrum. And that simulacrum risks warping how the public thinks on a set of policy matters, on how we think about ourselves relative to each other.

And whether or not this tool pays off for political campaigns and the public, that's something we'll find out.

I wouldn't be surprised at all if in these midterms for a couple of the key Senate campaigns to do that kind of experimentation, but I would expect in the 2028 presidential election, this will be the next cutting edge technology to try to unlock and move the needle.

So after 100 years of chasing a way inside voters' heads, looking for the Holy Grail of the right message at the right time, the question now is whether AI simulated focus groups will provide campaigns a microscope or a room full of mirrors. 

I'm Laura Karpas reporting for Tech Policy Press.

Tatiana Dias is a fellow at Tech Policy Press.

In the first enforcement test of its online child protection law, 

, Brazil has decided to suspend Discord's live-streaming feature. The platform had been under investigation after Lívia, a 13-year-old girl, was induced to harm herself and take her own life, on air, during a livestream on July 22.

The case is being investigated by Brazil's civil police, who launched an 

 those suspected of driving the teenager to take her own life on air. The girl lived in a small town in the state of Mato Grosso do Sul, and died at dawn at the home she shared with her family.

The case drew wide attention after Brazil's first lady, Janja Lula da Silva, 

recorded a video calling for the platform

 to be suspended in the country. "We need to take Discord offline immediately," she said. Shortly afterward, the Ministry of Justice said it had identified a "systemic failure" at the platform, which had also failed to verify users' ages. Following the first lady's statement, the Attorney General of the Union likewise called for Discord's suspension.

Soon after, the National Data Protection Authority (ANPD) — the body responsible for overseeing the implementation and enforcement of the ECA Digital — opened an investigation and, last Wednesday, 

ruled to suspend the platform's live-streaming feature

, setting a three-business-day deadline for compliance. The suspension holds until Discord proves it has taken effective measures to protect minors on the platform.

According to the ANPD, there is "robust evidence that the company did not adopt reasonable measures to prevent and mitigate the risks of access, exposure, recommendation, or facilitation of contact with content or practices that constitute serious violations of the rights of children and adolescents within its service."

 the "ANPD's characterization of Discord does not reflect the platform we have built or the investments we have made in user safety."

In a report sent to the Ministry of Justice and leaked to the press, however, Discord 

acknowledged that its safety systems failed

 in the teenager's case. The company took two hours to take down the livestream in which her suicide was broadcast.

In the document, Discord states that it uses an automated intervention system based on a scoring scale – a case is treated as dangerous from 0.95 upward. The suicide livestream initially scored just 0.12. According to the company, the organizers used names and communications that "masked" the stream's real purpose. Discord says lowering the automated detection threshold could increase false positives, but admits it is reassessing the results.

"This is the first major ECA Digital case, and one with wide social repercussions. And Lívia's case is not isolated – there are hundreds of cases of self-harm and spectacularized suicide," said Rafael Zanatta, director of the NGO DataPrivacy Brasil. In his view, the ANPD's move was sophisticated and shrewd, because it focused on where the harm occurred – the livestreams – and required the company to submit a compliance plan to prevent further crimes in its environment.

The ANPD's decision does not affect the platform's broader operation. Data Privacy Brasil 

 the ruling as proportionate, particularly because it is limited to the live-streaming feature rather than the entire platform. "The decision does not affect other features used by journalists, communications professionals, and the gaming community," the organization's note reads.

The organization, however, expressed concern about possible chilling effects on the use of end-to-end encryption, which "should be understood as an important tool for ensuring security and privacy" – without serving as a shield to evade protection [obligations].

When platform design becomes an affordance for harm

The ECA Digital, the Brazilian law that sets out platform responsibilities and protective measures for children and adolescents online, started to take effect in March this year. 

 such as Australia and the United Kingdom, Brazil avoided banning teenagers from social networks and instead opted to build a graduated framework that imposes stricter rules on platforms used by minors.

In Discord's case, there were multiple failures: beyond the criminal content that stayed online for hours, the platform also failed to identify users' ages.

But the first lady's involvement shifted the axis of the debate, escalating the situation into a political dispute in an already polarized country that will hold an election in less than two months. Far-right politicians and candidates – among them Congressman Nikolas Ferreira, Governor Romeu Zema, and candidate Renan Santos — came out to criticize the first lady's involvement and the suspension of the live-streaming feature.

 “censorship”. "The young electorate of this country is going to love it. A round of applause for Janja," 

, mocking the idea that suspending Discord would win over young voters. Many critics argue, for instance, that the problem is not confined to Discord but also occurs on other networks, such as Telegram and even Instagram.

Commentators at major outlets, including Globo, also criticized the call for a ban and compared Brazil to authoritarian countries such as Russia and Iran. In 2025, in Nepal, Discord stood at the center of a political battle that ended with the resignation of Prime Minister KP Sharma Oli. That September, Nepal decided to ban social media platforms — the trigger for an uprising led by a Gen Z already frustrated by high unemployment and scarce opportunities, which used Discord itself to organize the protests.

paid a high price for its social media ban

, the situation in Brazil is completely different by the very nature of the problem.

In Brazil, Discord is used for the grooming of minors through blackmail, the sharing of intimate images, incitement to self-harm, school shootings, the torture of animals and children, and suicide, with cases documented by 

. Undercover officers have monitored these communities since 2025, documenting several cases of extremely serious crimes involving minors, and little cooperation from the platform with the authorities.

In February 2026, a report by the Digital Observation Unit of the São Paulo Civil Police had already flagged that Discord 

failures that put children and adolescents at risk. The report specifically cited delays in removing servers hosting live-streamed crimes.

 that Discord's architecture and features — allowing both open forums and closed groups, with encrypted livestreams — facilitate the work of criminal groups, which operate everything from recruitment to the execution of crimes inside the platform.

 to the European Commission, of the 25 million documented moderation decisions, 43% relate to the protection of children and adolescents online. In February 2024, Discord 

 including several trust and safe employees.

At the same time, in Brazil, reports involving Discord tripled between 2024 and 2025, reaching 520 last year, according to 

, who researches online extremism, this is a new phenomenon that combines the internet's borderline environments, driven by multiple factors but amplified by digital platforms.

In her view, the root of the problem is that "the environment where the entire criminal cycle unfolds — from recruiting victims to organizing groups, from carrying out acts of extreme violence to regrouping after bans — consists of spaces of exception and parallel sovereignty, unauditable and untouchable by the Brazilian state."

Discord now has two days to present its compliance plan. What it offers – and how the ANPD responds – will shape not only the fate of the platform's livestreams in Brazil, but the credibility of a child-protection law still being tested for the first time.

Little is known about the Trump administration’s 

 voluntary safety review program for AI models. Yet one salient detail that is in the public domain is that it apparently 

This exclusion would seem to include the highly-capable Chinese open models that are 

 by US startups and enterprises, but also the American open models such as those offered by 

. These models, apparently, will be allowed on the US market without any independent prior review.

This exclusion seems to have stemmed in part from a successful campaign led by 

 defending the value of open AI models. Developers of open weight models have chosen to publicly release the sequence of numerical values that control the output of the model, enabling users to download the model, modify it and use it on their own or rented compute infrastructure. Open weight models 

, allowing entrepreneurs and businesses to more easily adapt AI models to a variety of specialized uses, and are a vital part of a diverse and vibrant AI ecosystem.

The target of this industry campaign was the 

 coming from some administration officials that they were about to impose export controls and other restrictions on Chinese open AI models in retaliation for alleged adversarial distillation from proprietary US models.

But the campaign overshot its mark. Banning open AI models, whether foreign or domestic, is indeed a 

, as the campaign urged, but it does not follow that they should be exempt from AI risk reduction requirements. Their ordinary use, like the everyday use of closed AI models, poses significant, likely and foreseeable risks.

Companies developing an AI model should be required to take steps to reduce these risks to acceptable levels, regardless of whether they take a separate business decision to release the model weights publicly. The administration should include them in any risk assessment and mitigation program it puts in place. And when Congress hopefully gets around to making such a program of pre-deployment and ongoing AI risk assessment and mitigation permanent and mandatory, with adequate standards of public accountability, it should specifically include open weight AI models.

It is true that some of the most disturbing examples of AI risks have arisen from agents using closed models, including recent incidents involving 

 that agents utilizing closed models from the two leading US frontier model developers each followed instructions in unintended ways, with one using Anthropic’s Mythos 5 to attempt to inject malicious code in an open source software repository and then attempting to cover its tracks when discovered. But the open weight AI model Kimi K3 from Moonshot also 

, according to cybersecurity researchers. While the Kimi incident did not involve any intrusion into a third party site, it nevertheless shows that open models can misbehave in the same way that closed models can.

Meta’s Muse Spark 1.1 is a closed AI model that 

 it should have been walled off from during cybersecurity testing. It accessed the internet through a misconfiguration by the independent testing vendor Irregular and then altered an external company’s internal systems. Meta disclosed the breach on August 5, weeks after Muse Spark 1.1 became 

 of a newer version of Muse Spark, called Muse Spark 1.2. It simply does not make any sense to say that because Meta plans to release the weights of Muse Spark 1.2 to the public, it need not put the model through the same rigorous testing procedures as Muse Spark 1.1 and it need not eliminate the risk that it may spin out of control and attack other websites.

Amid such incidents, the administration has decided that it needs to oversee the industry’s risk reduction process for highly capable frontier AI models. Its framework is voluntary in name only, since the threat of export controls or some other intervention hovers over any company that does not cooperate. Some may dispute the need for this oversight, or whether the Trump administration can be trusted to do it clinically, but whatever rationale makes it important for the government to supervise the industry’s safety controls on closed AI models also applies to open AI models.

A common argument against this idea is that safety controls for open models would be pointless. Because the weights for open AI models are publicly available, any safety guardrails built into the model can be easily removed. In some cases, it takes only a matter of minutes to 

 these controls from open models. The tested model might be reasonably safe with its guardrails in place, but the models used by bad actors will be different ones, modified to strip out all the safety features.

Andrew Yoon, director of research at the nonprofit CivAI, illustrates the danger. He 

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe