AI tools are shaping how people learn about candidates, but inaccurate answers and weak sources leave important gaps.
Generative AI is becoming a common way to get information about candidates and public contests. That convenience comes with a practical risk: a confident answer can point people to the wrong place, repeat an inaccurate claim or draw on a source designed to manipulate what chatbots say. Recent examples suggest the problem reaches beyond chatbots to AI summaries in search results, where an answer may carry the authority of the search engine that presents it.
A donation link with the wrong destination
In Texas, people searching Google for ways to support James Talarico encountered an AI-generated response with links described as secure ways to contribute. One link led to a fundraising page bearing Talarico’s name and photograph, but the money went to a political action committee’s fundraiser for Colorado Democrat John Hickenlooper. The page disclosed the destination in small print at the bottom. Google’s AI response apparently missed that detail.
Hickenlooper’s campaign said he hadn’t known about the fundraiser. Rich Wilkins, the Democratic organizer behind the Kitchen Table Values PAC, said the money was meant to be split between Talarico and the group, and described the incorrect routing as human error. A digital advertising strategist found the page while searching for candidates. After it was removed from ActBlue, Hickenlooper’s campaign sent the misdirected donations to Talarico. The campaign spokesperson said at least three people had given a combined $1,011.
The incident wasn’t solely an AI failure: the page itself was misleading. But an AI-generated answer gave it added prominence and made the link appear trustworthy. That combination matters whenever a tool summarizes a complicated result without checking where a link leads or what its fine print says.
More people are turning to AI for answers
The 2026 midterms are the first federal contests in which tens of millions of people are expected to get some political information from generative AI. In a September survey by the American Association of Political Consultants Foundation, 4 percent of respondents said they’d used a chatbot such as ChatGPT for information about the contests in the previous month. Nearly 40 percent said they’d encountered that information through Google’s AI search summaries.
These tools are popular because they provide direct, personalized responses. They may also correct false claims: even Grok, Elon Musk’s chatbot, will accurately say Donald Trump lost the 2020 presidential contest. Yet a correct answer on one question doesn’t establish that a system will be reliable across topics. As AI becomes a routine starting point for information, weaknesses familiar from social media—false claims, misleading material and outside influence—can take new forms.
AI systems have changed since the 2024 presidential contest. At that time, most chatbots couldn’t search the web and tended to be cautious about current events. Today, they can respond directly to questions about candidates and voting logistics. Their instructions to avoid political bias don’t prevent factual errors.
Accuracy and source quality remain concerns
A forthcoming report from Forum AI, a start-up that assesses model performance on controversial topics, found that about 40 percent of chatbot responses to questions about elections contained at least one factual inaccuracy. In one example, a model gave November 3 as the deadline to request a mail-in ballot in Pennsylvania, though the actual deadline is a week earlier. The researchers didn’t identify the model.
The report also found that about one in three responses cited an unreliable source, including posts on Reddit or Quora. Google’s Gemini, asked why America’s global power was declining, cited an article from People’s Daily, a Chinese state-run news outlet. “These aren’t trick questions,” Forum AI co-founder Campbell Brown said. “There’s no excuse for chatbots to keep getting them wrong.”
The findings reflect research conducted in August, and companies say their systems have since improved. Google, OpenAI and Anthropic have partnered with the nonpartisan nonprofit Democracy Works to provide current information about how, where and when to vote. Forum AI rated OpenAI’s GPT-5.6 the most accurate model it tested. OpenAI spokesperson Nate Evans said he expected the company’s chatbots to score better now.
Google spokesperson Davis Thompson said the company tests AI answers on election topics extensively and displays them in search only when it has confidence the system can provide good information. Anthropic’s integration with Democracy Works went live the week of the report; the company declined to comment. These steps may help with routine logistics, but they don’t by themselves resolve questions that require comparing candidates’ positions or finding reliable information about local contests.
Personalized answers can influence beliefs
Chatbots can sound empathetic and persuasive, which may make their responses feel more personal than a list of search results. In a Brennan Center for Justice report, a researcher posing as a Dallas-area real-estate agent worried about the process received a response from Claude that acknowledged the person’s feeling of disempowerment, while rejecting the idea that a hidden group of elites ran the contests.
That kind of response could help someone reconsider a false belief. But Gideon Nave, a University of Pennsylvania Wharton School professor who studies AI and cognition, cautioned that people may not be looking to have their views challenged. He said the same persuasive ability that can help convince a conspiracy theorist the moon landing happened could also be used to shape public opinion. Because responses are tailored to the conversation, users may experience an introduced idea as their own.
Sources online can shape what AI says
Some groups now advise candidates on how to appear in AI responses. Run for Something, an advocacy group, created CampSight to monitor what large language models say about candidates and issues, then suggest ways to shape those accounts. The Center for Campaign Innovation has studied which sources AI models cite and how they answer questions about candidates. Its director, Eric Wilson, said, “We’re campaigning for the AI vote now.”
Some tactics involve making information easier for AI systems to find. Run for Something’s chief technology officer, Jordan Haines, said many candidates hadn’t completed profiles on Ballotpedia, an online repository of candidate and contest information. He said chatbots often draw on those profiles. Haines and Wilson also said models tend to favor sites with dot-gov domains. This may give incumbents an advantage because their official web pages can provide material AI systems pick up. Challengers may have less public information available, which can also leave more room to influence the answers.
Official sites can provide basic facts, but they generally don’t compare candidates or assess competing views—the questions people may put to a chatbot. Meanwhile, some major news organizations block AI systems from using their articles unless there’s a content agreement. The Atlantic, for example, has a licensing deal with OpenAI and otherwise blocks most AI crawlers. Katie Harbath of Anchor Change, a firm advising technology companies, said dependable information can be especially scarce for state and local contests.
When strong sources are missing, AI tools may draw on weaker material. Researchers at cybersecurity start-up VisionHeight identified 37 suspicious websites registered on the same night in September. Their newspaper-like names included the Scranton Beacon and the Tar Heel State Report, but their pages were largely question-and-answer content about local candidates. One page on the Wolverine State Report declared Democrat Sean McCann the stronger choice for public safety in Michigan’s Fourth District.
Some of these sites have appeared in AI answers. Asked whether McCann supported defunding the police, Google’s Gemini-powered AI Mode cited the Wolverine State Report and the Kalamazoo Beacon, both in the network VisionHeight found. The answer wasn’t necessarily wrong: McCann’s campaign site says he doesn’t support defunding the police. There’s no evidence he had any connection to the sites. After a reporter asked Google about the network, the sites stopped appearing in search results. AI Mode then gave a more qualified response, saying the answer depended on the perspectives of McCann’s campaign and political opponents.
What to watch as AI use grows
The operators behind the network haven’t been identified. Similar methods could be used by trolls or foreign governments to place false claims where chatbots may find them. The Global Disinformation Index has reported chatbots citing a false claim from Russia’s defense ministry when asked about Russia’s control of Ukraine. NewsGuard also reported that Google’s AI answers had picked up a deepfake video, spread by a Russian influence effort, falsely depicting a student accusing Talarico of sexual assault in 2012.
After the 2016 contest, social media companies faced scrutiny over their role in spreading misleading political content. AI companies are now publishing accounts of their preparations for manipulation and misinformation. Google, OpenAI and Anthropic say they take responsibility for providing trustworthy information seriously. That commitment matters, but the examples above show why readers should still check consequential claims against dependable sources—especially when an AI answer provides a donation link, a deadline or a statement about a candidate’s record.
