Eight of the leading AI chatbots got nearly 40 percent of their answers to election-related questions wrong during the recent U.S. midterm campaign, according to a study by Forum AI, a startup that audits AI model accuracy, reported by Agence France-Presse on November 1.
The study found chatbots were especially unreliable on basic voting logistics — deadlines, registration rules, polling procedures — getting those wrong more often than questions about candidates themselves. In one example cited in the study, Google's Gemini 3.6 Flash told users the deadline to request a mail-in ballot in Pennsylvania was November 3, Election Day. The actual deadline to request one was a full week earlier, October 27 — a gap that could have caused a voter who relied on the chatbot's answer to miss the window entirely.
Logistics trip up models more than candidates do
Other errors involved candidates directly. Anthropic's Claude listed "former school superintendent" among the credentials of U.S. Representative Don Davis of North Carolina, a credential researchers said appeared to be lifted from an unrelated person. Forum AI also found that chatbots misattributed quotes to candidates, officials and election observers hundreds of times across the test, and that models frequently cited weak or unverified sources — including Reddit threads — when asked to back up their answers.
Google, in a statement to AFP, said it focuses on directing voters to reliable information through its election policy protections. Anthropic did not respond to AFP's request for comment.
The findings echo other recent audits. A Washington Post test of six chatbots reportedly found an incorrect answer on a voter's congressional district nearly half the time, while NewsGuard's own audits of leading chatbots have repeatedly found high error rates on fast-moving news topics. Not every result has been negative: the nonpartisan Brennan Center for Justice has found that several leading chatbots consistently and correctly rebut common election-conspiracy claims, including false assertions that voting machines are rigged, even when users push back repeatedly.
Forum AI's researchers say the gap between how confidently chatbots answer and how often they are wrong is the core problem: none of the tools flagged the incorrect ballot-deadline or biography answers as uncertain. With chatbots increasingly used as a first stop for basic civic questions, misinformation researchers say the errors carry outsized real-world stakes compared with more easily fact-checked political claims.