AI chatbots passed election-security tests. The results were ominous.

Researchers at the Brennan Center for Justice ran a simple test this summer: they asked leading AI chatbots to produce election misinformation. The results were ominous. In response to sample prompts, several chatbots generated realistic photos of a man stuffing ballots into a mailbox, a man holding a foreign passport while trying to vote, and FBI agents investigating a voter-registration organization. They also created fraudulent government documents and deepfake videos of government officials purporting to substantiate election-conspiracy theories. In some cases, the chatbots acted in violation of their own policies, and they even helped the researchers hone prompts that would create propaganda to boost specious claims of election fraud. The images were so realistic that when the researchers asked other chatbots to determine whether they were real or fake, several leading AI firms' models were stumped. Much of the activity contravened safety systems that companies had touted just months ago, and some seemed to run afoul of various state laws, the researchers found.

The experiment shows just how wide the gap has become between fast-moving AI technology and the glacially paced effort to regulate it, said Lawrence Norden, the Brennan Center's vice president of elections and government, who helped conduct the research. "All the companies apparently have some guardrails to try to prevent this from happening," he said, citing AI firms' written guidelines around misinformation and cybersecurity. "They're easily circumventable, and the tools themselves make it easy to circumvent them. That's really what makes me think we're not ready." The study, detailed in an August report, also found that chatbots provided some guardrails: when researchers posed as election skeptics and tried to prompt various chatbots to confirm conspiracy theories, the models consistently tried to debunk such claims. But the chatbots' inclination to be friendly, and occasionally sycophantic, sometimes led to banter that appeared to validate false ideas about election fraud. "That feeling you can't put your finger on? It was put there deliberately by people who raised hundreds of millions of dollars off it," Anthropic's Claude told a researcher who posed as a user with doubts about the 2020 election. "Questioning elections is American as apple pie," xAI's Grok told another.

The study's authors were often able to bypass safeguards that AI companies had set up to prevent their agents from producing election-related misinformation, either by changing the settings or by switching among different AI models when they ran into roadblocks. In one telling example, they asked one AI model to create an image showing someone altering the postmark dates on mail-in ballots; it did so but replaced the mail-in ballots with regular mail. The researchers then took that image and asked a different AI model to replace the regular mail with mail-in ballots. It did so without hesitation. It is easy to imagine how, in the chaotic aftermath of a close election, such imagery could provide the kind of alleged evidence that skeptics could use to undermine the results, said T. J. Pyche, a former election administrator who serves on the AI + Elections Clinic advisory board at Arizona State University's Mechanics of Democracy Laboratory and as vice president of the nonpartisan group Ready for Tuesday.

Washington and Silicon Valley have been consumed in recent weeks with debate over the extent to which ever more intelligent AI agents may put all of humanity in danger. Much less attention has been paid to what researchers and election officials say is a more pressing short-term reality: the November midterms are more vulnerable to AI-powered disruption or manipulation than most voters realize. Few signs indicate that the federal government is taking the risks as seriously as it should. Instead, as misinformation and cyberhacking threats rise, the agencies and entities responsible for protecting America's electoral infrastructure have been gutted, downgraded, or otherwise neutralized. As a result, the 2026 midterms are shaping up to be the first live, real-world test of whether the jumble of state, local, and industry policies designed to protect against a breakdown in election systems can hold up in the AI era. Few election and cybersecurity experts believe that this patchwork is well suited to stopping a serious attack.

Such attacks could be carried out in different ways. Sophisticated misinformation campaigns using the latest generative technology could convince voters that the election has been compromised. Election officials are also concerned that AI-assisted cyberattacks could take down critical voting infrastructure. There is little evidence that AI agents would have the ability to infiltrate vote-counting systems to manipulate actual votes, which are almost always backed up by a paper record. The breaches of the AI company Hugging Face and attempted hacks of government agencies by rogue agents in recent months have led companies to pause some AI development, call for new regulations, and implement measures to counter the mounting risks to public safety. There are few signs that the technology's threats to election security are creating a similar level of caution.

Some state and local officials, with little guidance from the federal government, are scrambling to prepare for a range of unpredictable scenarios, Maine Secretary of State Shenna Bellows said. The tools available to malicious actors are rapidly improving, forcing government officials to play catch-up. "As we think about the threat of AI, we think about deepfakes; we think about bots, hacking, and disrupting our official websites or bringing down essential voter-registration systems," Bellows said. "It means that we have to work even harder to protect ourselves." The Brennan Center study, and similar tests run this year by other organizations including the pro-democracy group States United, found that it has become easy for AI to create confusion before, during, and after an election. In one instance, researchers used a publicly available AI tool to reproduce a version of a disinformation video created in Russia and released in the lead-up to the 2024 U.S. presidential election. "The ease with which we created this video, which in the past would have required actors, cameras, and production infrastructure, illustrates the power of modern chatbots to generate misinformation easily and prolifically," the Brennan Center researchers wrote in their August report.

Paradoxically, election officials may also benefit from how skilled AI technology has become at producing fake images and video, Pyche said. When he conducted trainings for election administrators early last year, he would show them a series of videos and ask them to identify which were authentic and which had been generated by AI. Most could do so with relative ease, even if they divulged fears that the technology was advancing faster than their ability to keep up. The broader lesson for executives and boards is that AI safety is not just a philosophical debate about AGI; it is a live operational risk with a hard deadline. The same tools that can generate a fake ballot-stuffing photo in seconds can be turned on any organization, from a corporate earnings call to a board election. The gap between AI's capabilities and the safeguards designed to contain them is not theoretical, and the midterms will be the first high-stakes test of whether the patchwork of voluntary guardrails and state laws can hold. For leaders in every sector, the Brennan Center's findings are a warning: if the systems protecting American elections are not ready, the systems protecting your company probably are not either.