AI tools that had repeated a false allegation about a US election candidate later cited professional fact-checkers when AFP tried them again. Google and Microsoft attributed the change to newer information. The retest concerns one claim; separate research still found factual errors across election questions, particularly those asking how and when to vote.
Google AI Overview, Gemini and Microsoft Copilot turned to fact-checking coverage when AFP revisited a false allegation about a US election candidate. Earlier, the same tools had repeated the allegation in a NewsGuard evaluation of ten AI services. Professional fact-checkers had since widely debunked it. AFP's follow-up tested the tools' treatment of that claim rather than repeating the earlier studies' full set of questions.[1]
Google attributed the changed answers to its systems surfacing recent reporting. Microsoft said reliable information had been scarce during the initial, rapidly changing situation, and that Copilot incorporated later context identifying the allegation as invented. In a separate Forum AI examination covering eight models, almost 40 percent of election answers contained a factual mistake. Errors appeared more often in questions about voting procedures than in candidate questions.[1]
Among the errors, Gemini gave Pennsylvania's mail-ballot application deadline a week too late. Researchers also found quotations assigned to the wrong people and unreliable references. Google said it applies election-specific quality and policy safeguards; Anthropic gave AFP no response to its comment request. Separate Brennan Center research found leading chatbots resisted familiar election conspiracies even when repeatedly challenged, while still making mistakes on other election information.[1]
