Four in ten chatbot answers about the American midterm elections contain an error

Nearly 40 percent of the answers AI chatbots give to questions about the American midterms contain at least one factual error. That is the finding of a study by Forum AI, a start-up that evaluates AI models, reported by the AFP news agency. The report appeared in outlets including France 24, The Japan Times and The Star. The researchers tested eight models, which made more mistakes about voting procedures than about candidates.

Google's Gemini 3.6 Flash, for example, gave the wrong deadline for requesting a mail-in ballot in Pennsylvania: the real deadline was a week earlier. Anyone following the answer would have requested their ballot too late, Forum AI warns. Anthropic's Claude wrongly described North Carolina congressman Don Davis as a former school superintendent. The chatbots also misrepresented statements by candidates and observers and relied on weak sources, such as discussions on Reddit.

According to a survey by the National Conference on Citizenship, two in five Americans have already looked up election information through AI, but only 40 percent trust these tools for practical questions about where, when and how to vote. The watchdog NewsGuard found that Google's AI Overview, Gemini and Microsoft's Copilot repeated a baseless allegation against a Texas Senate candidate that probably came from a Russian disinformation campaign. Google and Microsoft say their systems have since been adjusted. Anthropic did not respond to AFP's questions.

There is also a more positive finding. The Brennan Center for Justice found that the major chatbots rejected well-known conspiracy theories about rigged voting machines, even when users kept pushing. The Institute for Strategic Dialogue calls Tuesday's election on 3 November the first in the United States in which AI is built into every major tech platform. Transparency: this edition was compiled by Claude, made by Anthropic.