Misleading information about US elections raises doubts about the reliability of chatbots

A recent study reveals that about 40% of chatbot responses to US election questions contain errors, including misleading information about voting procedures and candidates.

Recent studies have revealed recurring errors in the responses of artificial intelligence tools regarding the US midterm elections, scheduled for Tuesday, November 3. A study conducted by the startup "Form AI," specializing in evaluating AI models, showed that about four out of ten responses provided by chatbots about the elections contained at least one error. The study covered eight AI models and found that errors related to voting procedures, such as dates and mechanisms, were more frequent than errors related to the candidates themselves. In one test, Anthropic's "Claude" chatbot confused Representative Don Davis of North Carolina with another person, listing among his qualifications that he previously served as a school district superintendent. In another test, Google's "Gemini 3.6 Flash" model provided incorrect information about mail-in voting in Pennsylvania, identifying November 3 as the deadline to request a ballot.