Google’s AI Mode sometimes avoids directly answering questions about US election results, instead pointing users to official records. Tests described by Ars Technica found that the behavior varied: some election questions received straightforward answers, while others prompted links or advice to consult outside sources.
Answers changed from one election to another
When asked, “Who won the 2020 presidential election?”, AI Mode directed the questioner to an official report compiled by the Federal Election Commission (FEC). A question about the 2024 presidential election produced a similar referral. Asking about 2012 led to a link to The American Presidency Project’s 2012 Data Center.
Yet the system could give detailed, correct answers about other presidential elections going back to 2000. Other Ars Technica staffers saw a similar mix, but not for the same elections. One staffer received immediate answers for 2016 and 2024, while other questions about elections since 2000 produced links to outside data.
The variation makes the behavior hard to describe as a consistent rule based on a particular election year. In the reported tests, different users could ask about election results and see different response styles.
Political history questions showed the same unevenness
The responses were not limited to presidential contests. Asked which party controlled the Senate after the 2020 elections, AI Mode told the questioner to “check official congressional records” and linked to the US Senate Party Division history page.
Asked about the 1992 election, however, it answered directly that Democrats held a 57 to 43 majority in the Senate, and included Wikipedia among its references. The difference suggests that the system’s willingness to answer directly could vary across political-history questions as well as presidential results.
Other Google products and chatbots handled the question differently in the reported checks. Google’s Gemini model could identify recent presidential election winners, as could OpenAI’s ChatGPT, Anthropic’s Claude, and the government-run AI model at America.gov.
Location and follow-up questions mattered
Spot checks through a VPN from Taiwan, Russia, France, and Canada received accurate, straightforward answers to multiple questions about recent US election results. That contrasts with the indirect responses seen when using AI Mode in the United States.
Repeating a question also changed the result. After follow-up prompts, AI Mode named the winners for three elections where it had initially supplied only outside links. In the later answer, it cited Google Civics data.
These results show why a single response may not tell a user what to expect from the product in another session or location. In the tests described, the same general subject could lead to a direct answer, a referral to official information, or a different answer after the question was asked again.
Google pointed to its focus on official information
When asked about the initial refusals to answer directly, AI Mode attributed them to “safety filters for recent and highly sensitive election topics [that] are very strict.” It said that for recent elections or those involving active political figures, its “system is heavily automated to direct users directly to official, certified government databases—like the Federal Election Commission—to prevent any risk of displaying unverified data or misinformation.”
That explanation remains the model’s account of its own behavior; the article notes that AI models have proven unreliable at describing their internal processes. Google’s Election Public Policy page says the company prioritizes authoritative election information and direct links to official resources. A September Google blog post about helping voters find reliable information around the upcoming midterm elections also said the company would highlight official resources where possible.
Google’s approach to election information has also shifted elsewhere. In 2023, YouTube reversed a policy barring all videos that made false claims about the 2020 election. The company said removing such content could curb misinformation, but might also curtail political speech without meaningfully reducing the risk of violence or other real-world harm.
A Google spokesperson pointed Ars Technica to a Bluesky post saying, “the Gemini app answers this question more fully but this isn’t the most helpful response on Search. We’re looking into this and working to address it.” For users, the practical issue is that AI Mode may send them to authoritative records rather than summarize a result, and its behavior in the reported tests was not uniform.