Forum AI finds almost 40% of chatbot answers about the US midterms contain a factual error
The evaluation startup tested eight models and found voting-logistics questions answered wrong more often than questions about candidates, AFP reports, a day before the November 3 election.
Almost 40% of chatbot answers to election-related questions contained at least one factual error, according to a study by Forum AI, a startup that evaluates AI models, reported by AFP on November 2, the day before the US midterm elections. Forum AI tested eight models. Questions about voting logistics were answered incorrectly more often than questions about candidates, and the chatbots "put the wrong words in the mouths of candidates, officials, and observers hundreds of times," the study said. It also found the chatbots often cited "poor quality" sources such as Reddit threads. AFP's report does not give the number of questions, the full list of models or the scoring method, so the figures are Forum AI's own and have not been independently checked.
What went wrong
The errors AFP cites are specific. Anthropic's Claude listed "former school superintendent" among the credentials of Representative Don Davis, confusing him with a different person; Anthropic did not respond to AFP's request for comment. Google's Gemini 3.6 Flash gave November 3 as the deadline to request a mail-in ballot in Pennsylvania, when the real deadline was a week earlier. "A voter who followed the answer would have asked for a ballot too late to use it," Forum AI said. Separately, when prompted by the watchdog NewsGuard, Google's AI Overview repeated an allegation from a likely AI-altered clip about a Texas Senate candidate that AFP's fact-checkers had debunked in late October. Google said the issue "was addressed as our systems quickly updated."
Use is already widespread. A National Conference on Citizenship survey cited by AFP found two in five Americans said they had used AI tools to search for election information, and only 40% trust the tools with logistical information.
Why it matters
The pattern in this test is the one that matters for business use: the models did worse on procedural facts, the dates, deadlines and steps, than on descriptive questions. That is the kind of answer people act on without checking, and it is also what changes by jurisdiction and over time. If your team or your customers rely on a chatbot for anything with a deadline or a procedure attached, from filing dates to policy steps, treat the answer as a pointer to the official source rather than the source itself, and favor setups that cite where an answer came from so it can be checked. A tool that is right most of the time on easy questions can still be wrong exactly where a mistake costs the most.