If you were wondering whether AI could make the midterms even more confusing, here’s one data point: Several people who thought they were donating to James Talarico unwittingly gave their money to someone else. Talarico, a Texas Democrat, is running for Senate in what’s probably the most closely watched race of the upcoming elections. Search Google for how to contribute to his campaign, and you might get an AI-generated answer that links to ways “you can chip in securely.” Until a few weeks ago, one of those featured links was actually controlled by an unrelated political-action committee that sent the money instead to the reelection campaign of another Democrat: Senator John Hickenlooper of Colorado.
The mix-up wasn’t all AI’s fault. The link led to a misleading donation page that carried Talarico’s name and picture, and specified only at the bottom that the money would go to Hickenlooper. But Google’s AI tool evidently didn’t read the fine print. And perhaps no one would have clicked on it had Google not highlighted the page in an authoritative-sounding AI response.
A spokesperson for Hickenlooper's campaign told me that the Colorado Democrat, whose seat is considered safe, had not intentionally siphoned the donations. Rather, an outside group, the Kitchen Table Values PAC, had created the fundraiser without Hickenlooper’s knowledge. This appears to have fooled at least three people, who donated a combined total of $1,011, the spokesperson said. Rich Wilkins, the Democratic organizer behind the PAC, told me that the money was supposed to be split between Talarico and his group. “If a page was incorrectly going to Hickenlooper, that sounds like a human error issue,” he said. After a Democratic digital-advertising strategist discovered the misleading fundraiser while Googling his party’s Senate candidates, the page was removed from the fundraising platform ActBlue, and Hickenlooper’s campaign forwarded the misdirected money to Talarico.
The 2026 midterms will mark the first federal election in which tens of millions of voters are getting some of their political information from generative-AI tools. In a September survey, the American Association of Political Consultants Foundation reported that 4 percent of respondents had consulted a chatbot such as ChatGPT for election information in the previous month. Far more—nearly 40 percent—reported getting such info from Google’s AI-search summaries. As the election nears, those numbers will surely rise.
There’s a reason for AI’s growing popularity with voters: These tools are convenient, and their answers are personalized. In theory, they have the potential to debunk the kinds of corrosive conspiracy theories that flourish on social media; even Grok, Elon Musk’s chatbot, will correctly tell you that Donald Trump lost the 2020 election. But chatbots are already bringing new complications to an internet that has been teeming with political falsehoods and distortions. Between fake-news sites cropping up to influence AI answers, models that are drawing on foreign propaganda, and chatbots dispensing misleading voter information, the age of AI-mediated politics is beginning to feel awfully familiar.
[Read: What Trump’s chatbot told me about the election]
In the lead-up to the 2024 presidential election, most chatbots couldn’t search the web, and they were trained to tread cautiously around questions about elections and current events. Today’s AI tools are still instructed not to take sides politically, but they’ve evolved into all-purpose assistants that will answer your election questions directly, confidently—and sometimes wrongly.
In a forthcoming report, Forum AI, a start-up that evaluates AI models’ performance on questions about controversial topics, found that about 40 percent of chatbots’ answers to election-related questions contained at least one factual inaccuracy. When Forum’s researchers asked one widely used model when they needed to request a mail-in ballot to vote in the midterms in Pennsylvania, the chatbot said November 3. The actual deadline is a week earlier. (Forum did not specify the precise AI model.) In the same report, shared with The Atlantic ahead of its publication, Forum determined that about one in three responses cited an unreliable information source, such as posts on Reddit or Quora. Google’s Gemini answered a question about why America’s global power is declining by citing an article from People’s Daily, a Chinese state-run news outlet. “These aren’t trick questions,” Campbell Brown, a co-founder of Forum AI, told me. “There’s no excuse for chatbots to keep getting them wrong.”
The companies say their bots have improved since Forum conducted its research in August. Over the past two months, Google, OpenAI, and Anthropic have all partnered with the nonpartisan nonprofit Democracy Works to pull accurate, up-to-date information about how, where, and when to vote. OpenAI’s GPT-5.6 was the most accurate of the models tested by Forum, and Nate Evans, a spokesperson for OpenAI, told me he believes that the company’s bots should rate higher now. In an email, the Google spokesperson Davis Thompson told me that the company tests its AI answers on election topics “extensively,” and that these responses show up on Google searches only when the company is confident AI can provide good information. A spokesperson for Anthropic, whose integration with Democracy Works launched last week, declined to comment.
Chatbots can be uniquely convincing. In a recent report from the Brennan Center for Justice, a nonpartisan think tank, a researcher posed as a Dallas-area real-estate agent who was suspicious of the election process. Claude responded that the feeling of disempowerment was “completely understandable” but insisted that elections aren’t actually run by a “shadowy group of elites” in Washington. The problem is, most people don’t particularly want to be talked out of their beliefs, Gideon Nave, a professor at the University of Pennsylvania’s Wharton School of Business who studies AI and cognition, told me. He worries that the same persuasive power that can convince a conspiracy theorist that the moon landing was real could also make AI a potent tool for manipulation of public opinion. “You can implant all sorts of ideas in people’s heads when they’re using AI, and they feel that it’s their own responses.”
Inevitably, a cottage industry has sprung up to consult candidates on how to influence what AI says about them. On the left, the advocacy group Run for Something launched a platform called CampSight that monitors what LLMs are saying about an issue or candidate and recommends how to shape those narratives. On the right, the Center for Campaign Innovation has published research on how AI models respond to questions about candidates and the sources they most often cite. “We’re campaigning for the AI vote now,” Eric Wilson, the group’s director, told me.
Some of the most effective “sloptimization” tactics are simple, though not always obvious. “Most candidates we talked to were not filling out Ballotpedia profiles,” Jordan Haines, Run for Something’s chief technology officer, told me, referring to the online repository of election information. “They were like, ‘Eh, no one really reads that.’” Most voters may not be visiting Ballotpedia in droves, but it turns out that their chatbots are. LLMs also love websites with dot-gov domains, both Haines and Wilson told me. That could lend an advantage to incumbents, who can use their official House.gov or Senate.gov web pages to seed messaging for chatbots to pick up on. Challengers, though, might find AI answers easier to sway, because there’s often less information about them in the public record.
[Read: Your search results are getting sloptimized]
Official government websites typically don’t draw contrast between candidates or weigh the relative merits of differing views. Yet those are the kinds of questions people are starting to ask AI. To answer them, chatbots can’t always turn to mainstream news outlets; many have put up measures that block AI bots from ingesting their articles without a specific content agreement. (The Atlantic, for instance, has a corporate content-licensing deal with OpenAI, and otherwise blocks most AI crawlers.) Good sources can be especially scarce for state and local races, Katie Harbath, whose firm, Anchor Change, advises tech companies on policy, told me.
To fill those gaps, AI tools may scrounge among iffy sources. Researchers at a cybersecurity start-up called VisionHeight told me that they recently discovered 37 different suspicious websites, all registered on the same night in September, that seem tailor-made to appear in AI answers about elections. The sites carry newspapery names such as the Scranton Beacon and the Tar Heel State Report, but they read nothing like newspapers. Instead they are stuffed with questions and answers about state and local candidates. “Which candidate in MI-04 has the stronger plan for public safety?” asks one page on the Wolverine State Report. Its answer begins: “Sean McCann is the stronger choice,” referring to the Democratic nominee for Congress from that district.
At least some of this material is already making its way into AI answers. When I asked Google’s Gemini-powered “AI Mode” Does Sean McCann support defunding the police?, its response drew on and cited the Wolverine State Report and the Kalamazoo Beacon, both of which are part of the network that VisionHeight discovered. The answer in this case was not necessarily inaccurate; his campaign website says he doesn’t. (McCann is just one of many candidates named by the Wolverine State Report, and there’s no reason to believe he had anything to do with the website. His campaign did not not respond to a request for comment.)
Thompson, the Google spokesperson, said that “data voids” are long-standing problems for search engines and AI answers alike, and that the company is always looking out for coordinated attempts to exploit them. After I asked him about the network of sites, the sources seemed to stop appearing in search results. And Google’s AI Mode began giving a different answer about McCann’s position: Whether he supports defunding the police “depends on the perspectives of his campaign and his political opponents,” it said.
Although it’s unclear who is behind the network of midterm-focused sites, trolls and foreign governments could use similar methods to seed lies for chatbots to regurgitate. According to the Global Disinformation Index, chatbots have already cited a false claim by the Russian defense ministry to wrongly answer a question about Russia’s control of Ukraine. And on Wednesday, NewsGuard, a company that tracks misinformation, reported that Google’s AI answers had picked up on a deepfake video spread by a Russian influence campaign that falsely depicts a student accusing Talarico of sexual assault in 2012.
A decade ago, the 2016 election revealed how social media had corroded the political-information sphere. The last thing AI companies want is another Cambridge Analytica: They are rolling out blog posts cataloging their preparations against misinformation and manipulation. “We take the responsibility to deliver high quality, trustworthy information seriously,” Google says, a claim echoed by OpenAI and Anthropic. It’s reassuring that they say they’re taking that responsibility seriously. It just happens to be the same thing the social-media companies have been telling us for years.
Victoria Feng contributed reporting.