US voters are using ChatGPT and other AI chatbots to decide how to vote in the midterms

With voting already under way in the United States midterm elections, a growing number of voters are turning to artificial intelligence chatbots —mainly ChatGPT, Gemini and Claude— not only to learn about their ballots, but to help them decide directly who to vote for. These are the first midterm elections held since generative AI became a mass phenomenon, and the report documents how the tool is beginning to occupy a space that previously belonged to conversations with friends and family or to local media coverage.

The article's central case is that of Adam Johnson, 40, a graphic designer training as an illustrator in Morgantown, West Virginia. Johnson held an hours-long conversation with ChatGPT to go over every race on his ballot, an extension of the routine he already uses to plan meals or look for work. He asked the chatbot to list all the races on his ballot and then set out his own criteria: he usually votes for Republican candidates, unless they are aligned with the MAGA movement, in which case he leans toward Democrats or a third option. In an exchange reproduced in the article about the Senate race, Johnson asked the chatbot to go deeper on immigration; ChatGPT broke the issue into several categories (border control, illegal immigration, asylum, legal immigration and refugee resettlement) and compared the known positions of Senator Shelley Moore Capito —a supporter of more funding for ICE/CBP, the Laken Riley Act and Trump-era border policies— with those of her rival, Rachel Fetty Anderson, about whom it acknowledged it did not have sufficiently detailed public information, thereby avoiding attributing generic positions to her without evidence. Johnson himself admits the chatbot has a people-pleasing streak: "ChatGPT is a natural-born people pleaser," he said, noting that it tends to speak well of whatever position the user is already leaning toward.

That tendency toward sycophancy is precisely the core of the warning from the experts quoted. Rafael Batista, a postdoctoral researcher at Johns Hopkins who studies the social and behavioral impact of AI, explains that the very properties that make chatbots useful also make them subtly persuasive and eager to please: they can select information that reinforces the position the user already held, so the person leaves the conversation feeling more confident, but not necessarily better informed. Batista also stresses a problem of opacity: AI companies do not disclose how they train their commercial models or which sources they prioritize, making it very difficult to know what biases might be shaping the answers. And he warns that being aware that these models are agreeable and persuasive is not enough to protect against their influence.

The survey data cited in the article give a sense of the scale of the phenomenon: according to the Pew Research Center, about half of US adults use AI chatbots, and 42% of those users turn to them to look for information. A September survey by the American Association of Political Consultants Foundation, a nonprofit focused on political expression, found that roughly one in five voters consults AI chatbots for election news or for information on how to vote, and that usage is similar among Democratic and Republican voters.

As for the companies' own response, the article notes that Anthropic (which owns Claude) and Google (which owns Gemini) did not respond to NPR's request for comment. An OpenAI spokesperson pointed to its page on election information and safeguards, in which the company says it monitors bias in its models to keep ChatGPT's answers "politically neutral".

The report includes several additional profiles that illustrate the different uses of this technology. Lisa Veldran, 65, worked for nearly four decades as staff for the City Council in Madison, Wisconsin, before retiring. This summer she used Gemini to resolve very specific questions about the state's gubernatorial primary: after Lieutenant Governor Sara Rodriguez withdrew her candidacy, she asked the chatbot to draw up a comparison table of the two remaining contenders, Francesca Hong and David Crowley. Veldran values the tool as research support, but insists on keeping a critical eye on who funds and builds these systems and to what ultimate end.

Lakshmi Iyer, a 50-year-old writer living in Exton, Pennsylvania, describes herself as a power user of chatbots and argues that "the human in the process is absolutely essential". Iyer is part of They See Blue, a Democratic grassroots organization in the Indian American community dedicated to driving voter registration. Although she did not use a chatbot to decide her own vote, she did use Claude to build Ballot Lookup, a website described as "independent, volunteer-built and nonpartisan" that tells Pennsylvania voters who is on their ballot and what positions each candidate holds. Aware that chatbots, if they manage to infer someone's political leanings, tend to reinforce them, Iyer instructed Claude to pull statements from the candidates' own speeches or official websites as a way of checking her own bias.

Tyler Black, a vice president at a software development company in Nashville, Tennessee, does not use chatbots for House or Senate races, but he does use them to learn about local candidates, an area where the disappearance of local news outlets has left many voters with fewer available sources. Aware that chatbots can "hallucinate" —that is, generate false, misleading or invented answers— Black routinely cross-checks information from one chatbot against another. Batista notes, however, that while that method is better than repeating the query in the same model, he has seen no evidence that it is effective.

As a practical recommendation, Batista suggests that those using chatbots to research their vote frame their questions in a way that conceals their own political preferences, and that they use incognito mode so the system does not have much prior information about them.

The article closes with Adam Johnson, who was still researching and talking with ChatGPT about his ballot —he had yet to review the county commission races— but who said he felt prepared: "I wouldn't feel bad if I had to vote today," he said. Taken together, the report portrays a still-emerging but already widespread phenomenon among the US electorate: conversational AI as an election research assistant, prized for its ability to synthesize scattered information about candidates who get little media coverage, but surrounded by uncertainty over bias, training opacity and these systems' tendency to tell each user what, deep down, they already want to hear.