Google AI gets cagey when asked who won the 2020 elections

Google’s AI Mode sometimes refuses to directly name winners of recent US presidential elections, instead directing users to official government records. Asked who won in 2020, it pointed to the Federal Election Commission; it did the same for 2024, and sent the user to a data archive for 2012. It answered questions about older elections - including the 1992 Senate result - more directly. Ars staff found similar behavior, though which election questions triggered links rather than answers varied by user. Google’s Gemini app and chatbots from OpenAI, Anthropic, and America.gov answered the election questions directly.

The behavior appeared specific to Google Search in the US: tests from Taiwan, Russia, France, and Canada produced straightforward, accurate answers. Repeating a question could also prompt AI Mode to name the winner, citing Google Civics data. Asked why it initially withheld the answer, the system attributed its caution to strict safety filters for sensitive elections and a policy of steering users to certified sources. That explanation remains unconfirmed, since AI systems can misrepresent their own processes. Google’s election guidance emphasizes authoritative information, while its spokesperson said the Gemini app gives fuller answers and that the company was investigating how to improve Search’s response.