Could AI Disinformation Disrupt the Midterms?

Elections in the United States have been a battleground for foreign adversaries dating back at least a decade, with Russia, Iran, and China all using cyberattacks and social media campaigns to disrupt or interfere with the voting process. Add in the advanced cyber capabilities of the latest artificial intelligence models, their ability to generate bespoke information on command, and a highly polarized electorate, and the upcoming U.S. midterm elections begin to look like a potential minefield of misinformation.

But the overarching message from more than half a dozen experts whom Foreign Policy spoke to, including current and former tech executives and government officials, is that it’s not time to press the panic button just yet.

Elections in the United States have been a battleground for foreign adversaries dating back at least a decade, with Russia, Iran, and China all using cyberattacks and social media campaigns to disrupt or interfere with the voting process. Add in the advanced cyber capabilities of the latest artificial intelligence models, their ability to generate bespoke information on command, and a highly polarized electorate, and the upcoming U.S. midterm elections begin to look like a potential minefield of misinformation.

But the overarching message from more than half a dozen experts whom Foreign Policy spoke to, including current and former tech executives and government officials, is that it’s not time to press the panic button just yet.

“There are attempts to use AI to deceive, certainly, but they haven’t been particularly effective,” said Matt Masterson, a senior director for elections and societal resilience at Microsoft. “I haven’t seen anything that’s really taken hold in this election cycle or jumped out at me as really having an impact in that way.”

That’s not to say that U.S. adversaries couldn’t be—or aren’t already—using AI tools to make their disinformation campaigns more effective. Large language models such as OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini (to name only a few) give bad actors the ability to operate at a speed and scale that they previously couldn’t.

“It’s helped them operationalize their work and almost industrialize it—it means that they can spin up content more quickly. They can also do a lot of research more effectively and quickly in terms of their target audience,” said Olga Belogolova, the director of the emerging technologies initiative at Johns Hopkins University’s School of Advanced International Studies.

Investigators have uncovered past Russian and Iranian campaigns because of mistranslations into English or a lack of knowledge of local slang. But AI allows those actors to sound more authentic on social media.

“A lot of what we’re seeing on the back end is changing because it’s helping them draft and polish content, translate, and localize things, which are important changes. I’m not trying to dismiss them,” Belogolova said. “But that doesn’t mean that they’re necessarily going to be any better at persuading and reaching people,” she added. That’s in part because the social and even traditional media channels that they rely on to distribute the content remain the same. “A lot of what you see on the receiving end of information consumption hasn’t changed dramatically.”

Midterm elections in the United States also pose a fundamentally different challenge from an interference standpoint than U.S. presidential elections, such as the one in 2024.

“Midterms are not as easy to target or as popular to target by foreign adversaries. Our federal system, with 400-plus congressional races and 30-something Senate races, makes it hard to target effectively,” said Jacob Breach, a senior advisor at the Foundation for Defense of Democracies think tank who previously spent 15 years in the U.S. intelligence community focused largely on foreign influence operations. “They’re testing grounds for actors’ unique tactics and unique combinations of tactics to attempt to interfere or influence these elections, but it’s not like it would be in a presidential election year.”

Iran and Russia are also engaged in protracted wars that could reduce their wherewithal to target the midterms. AI can help ease that load, but only so much.

“These actors have limited capabilities and bandwidth. When those actors are distracted by other things, like the wars in this case, they have less bandwidth to devote to sustained campaigns,” Breach said.

---

Multiple experts told Foreign Policy that companies, election officials, and voters are simply better prepared this time around to combat misinformation—including by using AI as a solution on the defensive side.

“Back in 2024, it was a rare election official who was really thinking about AI for good—I think most of us were laser focused on deepfakes and the concerns associated with that,” said Bill Gates, who served as a supervisor for Arizona’s Maricopa County during five successive elections until 2024 and is no stranger to election integrity efforts.

Gates (no relation to the billionaire Microsoft founder) now heads Arizona State University’s Mechanics of Democracy Lab, which has launched an initiative sponsored by Microsoft to train election officials in AI applications and harness AI to provide voters with authoritative information.

“The sort of burying-the-head-in-the-sand ostrich approach that maybe a fair number of us were doing in 2024, you’re seeing a lot less of that this time around,” Gates said.

The ASU partnership is one of several that major tech companies have entered into to ensure that their models can serve reliable information. OpenAI has partnered with The Associated Press to supply its vote count data to ChatGPT users who ask about it, while OpenAI, Anthropic, and Microsoft have all partnered with an organization called Democracy Works, which aims to provide tech companies with reliable information about polling place locations and other frequent queries by prospective voters.

OpenAI and Anthropic say they have also tested their models extensively to prevent political bias and clamp down on bad actors’ potential misuse of their platforms. “People should be able to access reliable information, understand the origin of AI-generated media that they might encounter off platform, and … explore political questions without being steered in one particular direction by the model,” Josh Lawson, OpenAI’s head of civic and strategic risk, told reporters in a recent briefing. Combating adversaries requires “adding friction in strategic areas” that make it harder for them to exploit the chatbots. “Bad actors have a lot of options whenever they choose to fly. We don’t want them to fly with us.”

In an encouraging sign, a study published by the Brennan Center for Justice found in August that most of the popular chatbots consistently pushed back against common election conspiracy theories.

The pace at which AI is evolving and the so-called “black box” nature of the technology mean that unforeseen consequences and threats are always a possibility.

“The hardest thing is that we don’t know what we don’t know,” said Katie Harbath, author of the new book Disrupting Politics: A Front Row Seat to the Collision of Technology and Democracy, which is about her decade leading Facebook’s election-focused public policy team until 2021. “You can do as much as possible to try to be ready for any of these potential things, but what if something happens that we weren’t prepared for?” she asked. “The thing that I’m most happy about—I don’t know if that’s the right word—is the fact that there are a lot of people thinking about this.”

That also extends to the platform users and voters themselves, for whom AI-generated content has become a part of life. That has led to a natural skepticism that many experts say is somewhat healthy—provided that it doesn’t tip over into distrusting even authoritative information.

“I do think that the prevalence of AI-manufactured and -created information is causing people to at least look twice at what they’re seeing,” Breach said.

“Societally, it’s not great. I would love to be able to trust what I saw on the internet,” he added. But “if AI is causing Americans to get better at media literacy simply because they trust what they’re seeing less, that’s a good thing.”