THE ESSENTIALS IN 30 SECONDS

  • Almost 40% of answers from eight AI models to election-related questions contained at least one factual error, according to Forum AI as reported by AFP.
  • Questions on voting logistics, such as deadlines, produced errors more often than questions about candidates.
  • In May, Forum AI's NewsBench audit of four chatbots also reported a left lean in ChatGPT, Claude and Gemini failures and a right lean in Grok's.

As Americans prepare to vote in the November 3 midterm elections, an evaluation suggests that AI chatbots remain an unreliable guide to the basics of casting a ballot. Forum AI, a start-up that evaluates AI models, found that almost 40% of the answers eight models gave to election-related questions contained at least one factual error, according to an AFP report published on November 1.

What the study found

The errors were more frequent on practical questions about how to vote than on questions about candidates. In one example cited by AFP, Google's Gemini 3.6 Flash, asked about voting in Pennsylvania, gave November 3 as the deadline to request a mail-in ballot, while the real deadline was a week earlier. Forum AI noted that a voter relying on that answer would have requested a ballot too late to use it. In another case, Anthropic's Claude listed "former school superintendent" among the credentials of North Carolina Representative Don Davis, apparently mixing him up with someone else.

The study also found that chatbots attributed words to candidates, officials and observers incorrectly hundreds of times, and often relied on weak sources such as Reddit threads. The AFP report names only two of the eight models tested and does not detail the number of questions, the testing period or the scoring method, so per-model results could not be checked. Google told AFP it focuses on helping voters find reliable information and applies strict protections to election topics; Anthropic did not respond.

Earlier findings on bias and sourcing

These figures follow NewsBench, a benchmark Forum AI launched in May. That audit used 3,136 prompts and 12,542 responses from ChatGPT, Gemini, Claude and Grok, scored for accuracy, neutrality and source quality by AI judges calibrated against a bipartisan expert panel. About 30% of all responses contained at least one verifiable error. Nearly a quarter failed the neutrality check: failures by ChatGPT, Claude and Gemini leaned left far more often than right, while Grok's leaned right. Forum AI is led by Campbell Brown, a former Meta executive, who has argued that model makers are effectively grading their own work.

The stakes are not trivial. A National Conference on Citizenship survey cited by AFP found that two in five Americans had used AI tools to look for election information. Official state election websites remain the reference for voting rules.