U.S. Midterms: AI Chatbots Could Lend Influence Campaigns New Credibility

Jessica Brandt is a leading expert on the national security implications of artificial intelligence (AI), foreign malign influence, and the governance of emerging technologies. She previously served as director of the Foreign Malign Influence Center at the Office of the Director of National Intelligence.

On October 20, the day after early voting began in Texas, a likely AI-generated video appeared online, targeting Democratic Senate nominee James Talarico. The video falsely accused Talarico of misconduct just weeks before ballots are counted in this closely watched race that could determine control of the Senate. More than 850,000 people viewed the video in its first day, until X, formerly Twitter, suspended the account that posted it. 

Social network analysis company Graphika and other independent researchers connected the video to Storm-1516, a Russia-linked group that has a long history of posting fake whistleblower videos—including deepfakes—to advance the Kremlin’s interests.  

But the morning after the video was released, widely used artificial intelligence (AI) chatbots—including Google’s AI Overview and Gemini, and Microsoft’s Copilot—repeated the false claim, according to NewsGuard, a company that monitors online falsehoods. The incident illustrates that AI isn’t just a tool for generating influence content—it’s a distribution method. 

Foreign influence operations tend to achieve the most reach when they land their content in established, trusted channels. For years, Moscow has worked to disguise its campaigns as the product of authentic domestic voices. In 2016, it used a false persona to place articles in several real, independent, alternative news sites. In 2020, it recruited genuine, unwitting freelance journalists—including some Americans—to publish articles targeting left-leaning audiences. In 2024, it funneled payments to prominent right-leaning American influencers with large followings. In each case, and in many others, Russia’s goal was to leverage the audiences these outlets and individuals have built, along with the credibility they carry. Its goals haven’t changed. 

In this latest episode, we see AI chatbots functioning as the newest channel for Russia to launder its influence content. The chatbots gave Russia’s content the authority of an answer rather than a post, helping to more effectively reach a much broader audience than any meme posted on social media ever could. The mechanism was simple. Russian influence actors seeded the content online, multiple large language models (LLMs) absorbed it, and some of them repeated its false claim to users, at least in the hours after the video surfaced. That is especially likely to happen at the outset, when there is no authoritative rebuttal. In this case, almost twenty-four hours in, when prompted on whether Talarico had been accused of sexual abuse, Google’s Gemini and Microsoft’s Copilot repeated the false claim. So did Google’s AI Overview, using citations that made no mention of the matter. 

Google and Microsoft did not respond to NewsGuard’s requests for comment. 

It’s not clear that Storm-1516 was deliberately targeting LLMs as much as benefiting from them incidentally, but it’s not the first time Russian influence content has benefited from tools built to surface information. In the previous decade, search engines returned Kremlin propaganda on topics for which there was little authoritative content.

This is a genuine challenge for AI companies. When a new claim emerges, authoritative information about it may not yet exist. That gap is known as a data void—or an information gap that bad actors can easily fill. Fast-moving, breaking news events are particularly prone to them because media organizations haven’t yet been able to publish relevant reporting. Search engines have struggled with data voids for years, and cuts to fact-checking and newsroom resources could exacerbate the problem. That puts the onus on AI companies to ensure their chatbots flag when there is limited or thinly sourced information, as many search engines do. This is especially true close to high-consequence events such as elections. 

Both Google and Microsoft had announced midterm election safeguards in September, but Microsoft also acknowledged that election information can pose challenges for AI because it “changes quickly and can be incomplete, contested, or stripped of essential context.”  

The Talarico case marked the reappearance of an operation with experience targeting U.S. elections: Storm-1516. This Russian group tried almost exactly the same play last cycle, peddling a similar false accusation against Minnesota Governor Tim Walz. Storm-1516 had been remarkably quiet this cycle, until the video targeting Talarico dropped. Since then, we’ve seen two additional examples of similar content targeting Democratic Senate candidates running in tough races: Sherrod Brown in Ohio and Troy Jackson in Maine. This flurry of activity suggests that the group’s efforts are picking up as Election Day approaches. 

AI chatbots appeared to respond better to these most recent examples. That may be because the relevant companies tightened their practices in the wake of the Talarico episode. Near-real-time debunks from fact-checkers also appear to have helped. 

That could be needed in the aftermath of the midterms. Storm-1516 released a series of videos close to Election Day in 2024, as it sought to undermine confidence in the vote itself. I’ll be watching for similar activity in the days ahead. Last cycle, the U.S. government worked to rebut the videos quickly and publicly. This cycle, it will be especially important that journalists and fact-checkers write debunks with LLMs in mind.  

This moment does not call for panic, however. NewsGuard’s results were limited, and they only reflect a particular moment in time. But they do highlight an important new vector for laundering foreign malign influence operations, one that AI companies, journalists, and fact-checkers should be alert to now and in the future.