Imagine this: you’re trying to do your civic duty, but the polling place hours are confusing and the candidate’s record is fuzzy. So you turn to an AI chatbot for help—maybe it’s embedded in your search engine, maybe it’s a standalone assistant. It sounds convenient, even responsible. But what if the answer is wrong? That’s the troubling reality as the United States heads into the November 3 midterm elections. A new study from Forum AI, a company that evaluates AI systems, found that nearly 40 percent of chatbot responses to election-related questions contained at least one factual mistake. The researchers tested eight popular AI models and discovered that the errors weren’t random. The bots were significantly more likely to stumble on basic voting procedures—things like deadlines, registration rules, and mail-in ballot requests—than on questions about candidates themselves. That’s particularly worrying because procedural mistakes can have real consequences. If a chatbot tells you the wrong date to request an absentee ballot, you might miss your chance to vote. If it gives you the incorrect address for a polling station, you might show up somewhere that doesn’t exist. The study suggests that these tools, which millions of Americans are increasingly using as a first stop for civic information, are far from reliable. And while the occasional wrong answer might seem like a harmless glitch, in an election context it can undermine a person’s ability to participate in democracy. The researchers aren’t claiming that chatbots are deliberately misleading people. Rather, they say, the systems are still too immature to handle the complexity and precision required for election guidance. And that creates a dangerous gap between the public’s trust in these tools and their actual performance.
The Forum AI researchers shared some specific examples that show how easily chatbots can slip up. When asked about Don Davis, a Democratic representative from North Carolina, Anthropic’s Claude described him as a former school superintendent. That’s incorrect—the model apparently confused Davis with someone else entirely. It’s a small but telling error. In another case, Google’s Gemini 3.6 Flash was asked how to vote in Pennsylvania. It told the user that they could request a mail-in ballot until November 3. In reality, the deadline was a week earlier. A voter who followed that advice would have asked for a ballot too late to use it. That kind of mistake is not merely an inconvenience—it’s the difference between having a voice and being disenfranchised. The study also found that chatbots repeatedly misrepresented statements made by candidates, public officials, and observers. Sometimes they paraphrased inaccurately; other times they simply fabricated quotes. Worse, the bots often relied on low-quality sources, including Reddit discussions, to shape their answers. That means they weren’t just making occasional arithmetic errors. They were absorbing the chaotic, unverified noise of the internet and serving it back as confident facts. For a person who doesn’t have time to cross-check every answer, the risk of being misled is high. And because chatbots often deliver their answers with authoritative language, it can be hard to know when to trust them. The underlying issue is that language models are designed to generate plausible-sounding text, not to verify the truth. They prioritize fluency over accuracy. So when a user asks a straightforward question about an election deadline, the model will happily produce a neat, confident answer—even if that answer is completely wrong. This isn’t a problem that can be solved with a simple fix. It’s a structural limitation of the technology, and it’s especially dangerous in a high-stakes context like voting.
This all comes at a time when AI tools have become deeply embedded in the everyday platforms Americans use. The November 3 vote is the first US federal election in which major technology companies have integrated AI assistants across their services—search engines, social media, productivity apps, and more. That integration means a huge number of voters might encounter AI-generated election information without even realizing it. They might ask a question about a ballot initiative, get an answer that sounds authoritative, and never think to double-check it. A survey by the nonpartisan National Conference on Citizenship found that two in five Americans had used AI tools to search for election information, including details about candidates and ballot measures. But the same survey reveals a profound lack of trust. Only 40 percent of respondents said they trusted these tools to provide practical voting information, such as when, where, and how to vote. Six in 10 said they had little or no confidence in the tools’ ability to give unbiased information about candidates. That split is telling. People are curious enough to try AI, but they’re not sure they should believe it. The gap between usage and trust suggests that many users are engaging with these tools cautiously, perhaps as a starting point rather than a final authority. But even cautious users can be led astray if they don’t know which answers to double-check. The survey also highlights a broader anxiety: not only are chatbots making factual errors, but they may also be presenting biased or slanted perspectives on the candidates. In a deeply polarized country, the last thing voters need is a technology that reinforces their existing biases or invents new ones. The platforms themselves say they are working to improve. Google, for example, has said it has strict quality and policy protections for election-related responses and is helping voters find reliable information across its services. But the independent research suggests that those protections are not always working. With millions of voters turning to AI, even a small failure rate can affect a very large number of people.
Another layer of concern involves foreign disinformation. Researchers have warned that AI chatbots can amplify influence operations by repeating false claims that are already circulating online. In late October, AFP fact-checkers debunked a likely AI-altered video that appeared to show a former student of Texas Senate candidate James Talarico accusing him of sexual misconduct. The baseless allegation bore the marks of a Russian disinformation campaign. But when a disinformation watchdog called NewsGuard tested the major AI tools, Google’s AI Overview repeated the allegation as if it were true. Two other tools, Google’s Gemini and Microsoft’s Copilot, also reproduced the false claim. In a separate study of ten AI tools, NewsGuard warned that users were at risk of being misled by foreign influence operations and other malicious actors. This isn’t just about bad code or sloppy data. It’s about foreign governments using AI systems as unwitting amplifiers for their attacks on American democracy. The good news, if there is any, is that the situation can change quickly. In a later AFP test, after the allegation had been widely debunked, the same tools cited professional fact-checkers and no longer repeated the claim. That shows that chatbot responses are not fixed—they evolve as new information becomes available. It also means that the window of vulnerability is often short, but in an election cycle, even a few days of false information can cause damage. Google said the issue had been addressed as its systems were updated to reflect the latest reporting. Microsoft said the claim emerged during a rapidly developing situation when reliable information was limited, and added that Copilot later identified it as fabricated. Anthropic did not respond to a request for comment. The incident illustrates a fundamental challenge: AI chatbots are caught between being fast and being accurate. When a false story starts spreading, they may pick it up before fact-checkers have a chance to catch it. And by the time the correction comes, the damage may already be done—people may have seen the claim, shared it, and changed their minds about a candidate.
But the researchers also found that AI tools aren’t all bad news. In fact, they can sometimes serve as a bulwark against election misinformation. The Brennan Center for Justice, a nonpartisan policy institute, conducted its own study and found that leading chatbots rejected common election conspiracy theories—including claims that voting machines had been rigged—even when users repeatedly challenged them. The chatbots often made mistakes in other areas, but they did not endorse the conspiracy theories. Instead, they stood their ground, explaining that there was no evidence to support such claims. The center suggested that these tools could play a constructive role in fighting election misinformation. That might sound surprising, especially after all the examples of errors. But it makes sense when you think about how the models are trained. They’re designed to avoid repeating harmful claims, and they’ve been reinforced with guardrails around election content. So while they might stumble on a deadline or mix up a biographical detail, they often know enough to avoid amplifying wild accusations about voting machines or stolen elections. That doesn’t mean they’re reliable sources, but it does mean they aren’t wholesale engines of conspiracy. The London-based Institute for Strategic Dialogue has described the November 3 vote as the first US election in which AI tools are built into every major technology platform. That is a critical moment. It means the potential for harm is widespread, but also the potential for good is perhaps unprecedented. A chatbot that can push back on baseless conspiracy theories could be a valuable ally in a country where false claims about election fraud have led to real-world consequences. The key is figuring out how to encourage that positive behavior while minimizing the errors. Some researchers suggest that platforms need to be more transparent about AI’s limitations, perhaps by adding caveats to responses or directing users to official sources. Others say that AI should be trained specifically on verified election data, rather than scraping the entire internet. Whatever the solution, the Brennan Center’s findings offer a small dose of optimism in an otherwise concerning picture.
As the election draws closer, the stakes are rising. Voters are increasingly using chatbots to access political information, whether they’re looking for a candidate’s stance on healthcare or trying to figure out whether they need to bring a photo ID to the polls. And while AI can be a helpful starting point, the evidence is clear: it is not yet ready to be the sole source of truth for civic decisions. The combination of a high error rate on procedural questions, the tendency to rely on low-quality sources, and the vulnerability to foreign disinformation means that anyone who uses a chatbot for election information should do so with a heavy dose of skepticism. That doesn’t mean abandoning the technology altogether. It means treating it as a complement—not a replacement—for official sources. The easiest way to verify an answer is to check your state or local election office’s website, or to use a trusted nonpartisan voter guide. It also means being aware that chatbots can be manipulated, both by bad actors and by their own limitations. The platforms have a responsibility to strengthen their safeguards, and they say they are working on it. But in the meantime, voters need to be their own fact-checkers. That’s a lot to ask of people who are already busy and overwhelmed by the flood of information during an election cycle. Yet the alternative is worse: a democracy in which a significant number of voters are acting on false information, not because they chose to, but because a machine made a confident error. The November 3 midterms will be the first test of how well Americans navigate this new landscape. It won’t be perfect. But by understanding both the promises and the pitfalls of AI, voters can make more informed decisions—not just about candidates, but about the tools they use to learn about them. After all, democracy depends on an informed electorate. And being informed means knowing when the technology in your pocket is lying to you.