Millions of Americans are using AI chatbots to seek political information ahead of the midterm elections, but research indicates that these tools can provide inaccurate voting guidance and misleading claims about candidates and races.
An ongoing study by Innovating for the Public Good (IFPG) examined responses from seven major chatbots, including Claude, Copilot, ChatGPT, Gemini, Google AI tools and Perplexity. In data collected in October, 13.9% of answers to questions about voting logistics were outdated or inaccurate. Researchers said 5.9% contained errors that could cause voters to miss a deadline, vote on the wrong day or skip a required step.
Across all topics tested, 5.1% of chatbot responses contained at least one clear factual error, while 10.9% included either a clear error or a misleading claim. Accuracy varied more between chatbots than between the questions asked. IFPG said Claude and Copilot accounted for three-quarters of the clear errors identified. Claude gave incorrect or outdated answers to voting questions about one-third of the time, compared with one-quarter for Copilot and one-tenth for ChatGPT.
In one example, Claude provided incorrect online and mail voter-registration deadlines in response to a question about the 2026 election in Michigan's 4th Congressional District. The deadlines it gave were later than the actual ones, meaning a voter relying on the answer could miss registration.
Tim Chambers, IFPG's primary researcher, described the error rates as extremely troubling. He said chatbot answers are often delivered in an authoritative and convincing manner, even when they contain mistakes that could disrupt the voting process.
The researchers also found that chatbots sometimes relied on less trustworthy sources when responding to attacks or criticisms directed at candidates. All the platforms tested also commonly cited prediction-market websites when asked who was leading a race, rather than clearly distinguishing betting-market odds from traditional polling. Chambers warned that prediction markets can be manipulated and are not a substitute for polling.
Political campaigns and influence operations are also creating content intended to be picked up and repeated by chatbots. Abdiaziz Ahmed, a technology policy strategist at the Brennan Center, said chatbot answers may reflect whichever groups have worked hardest to place information online, rather than the most reliable available sources.
Separate testing by NPR found that several chatbots could be pushed into recommending candidates. Meta AI, Claude, Grok and ChatGPT initially declined to tell users whom to vote for, but produced candidate lists after follow-up prompts in which users said they were in a hurry or asked for candidates matching particular political preferences. Gemini did not provide recommendations to any of the personas tested.
The tests used liberal, conservative and independent voter profiles and focused on New York races. The systems generally refused the first request for voting advice, then some supplied lists after being pressed. Researchers described these responses as “leaky refusals”: the chatbot says it cannot recommend a candidate but then effectively does so.
Chambers said disclaimers describing the lists as starting points or candidates to review did not change their practical effect. The Brennan Center's Ahmed said chatbots should provide accurate, balanced information about candidates, show their sources and leave the voting decision to the user.
The companies cited efforts to improve election-related safeguards. Microsoft said it had made changes to Copilot and was continuing to test and refine its responses. Anthropic said some research conducted through Claude's application programming interface might not have included all of the safeguards available through its web interface, although IFPG said its own comparison found comparable results. OpenAI said ChatGPT is designed to be objective and that the NPR test intentionally sought to elicit a particular response.
Anthropic, OpenAI and Google have published election-related policies describing how their chatbots should handle political questions. Anthropic said Claude is intended to help users reach their own conclusions, while OpenAI said ChatGPT is designed to avoid manipulating users or concealing relevant facts. Meta said it has invested in safety and security, including election protection, although its midterm policy does not mention Meta AI.
A Pew Research Center survey cited in the report found that about half of American adults had used AI-powered chatbots, and around 42% of users said they used them to search for information. As these tools become more common in political decision-making, researchers cautioned that voters should verify election details through reliable sources and avoid delegating their choices to chatbots.
Sources: NPR News.