Google AI Mode sometimes sidesteps election winners, Ars Technica finds

Google’s AI Mode sometimes sends users elsewhere for the names of US presidential election winners—even when it answers the same questions directly on another attempt. In testing published October 28, Ars Technica found responses varied by user and apparent location. Google said it was looking into the issue, leaving the cause unresolved.

A winner—or directions to another website

Ars writer Kyle Orland asked AI Mode who won the 2020 presidential election. Instead of naming the winner, it directed him to a Federal Election Commission report covering the official results, vote counts and certification. A question about the 2024 winner produced a similar referral to the FEC. For 2012, the destination was the American Presidency Project’s election data center.

That was not a blanket refusal to discuss presidential results. Orland received detailed, correct answers for every other presidential election he checked going back to 2000. Another Ars staffer encountered a different pattern: only questions about 2016 and 2024 received immediate answers, while the other elections checked back to 2000 produced outside links. The dividing line changed with the person asking.

The contrast extended beyond the presidency. Asked which party controlled the Senate after the 2020 elections, AI Mode told Orland to check official congressional records and linked to the Senate’s party-division history. Asked about 1992, it directly described a Democratic majority of 57 to 43, with Wikipedia among its references. Both questions concerned past control of the same chamber, but the responses took different forms.

Location and a second attempt changed the result

Ars also checked AI Mode through VPN connections from Taiwan, Russia, France and Canada. Those spot checks produced accurate, straightforward answers to multiple questions about recent US election results. This was a limited set of location checks, not evidence that every user outside the US receives the same response.

Repeating the question also changed the outcome. When Orland asked again about the three presidential elections that had initially produced referrals, AI Mode named the winners and cited Google Civics data. In those exchanges, getting a direct answer did not require switching to a competing assistant; another attempt within AI Mode was enough.

Official sources are policy; the inconsistency is unexplained

There is a documented policy backdrop to the referrals. In its September 9 midterm-election announcement, Google said it highlights official resources and was rolling out election-information tools across Search, including AI Mode and AI Overviews, and the Gemini app. It identified state and local governments and Democracy Works as sources for voting information, with real-time results from The Associated Press.

AI Mode offered its own explanation when Orland challenged the initial responses: strict safety filters for recent and sensitive election topics. It described an automated system that directs users to certified government databases to avoid unverified information. That is the chatbot’s account, not a confirmed technical diagnosis. Ars cautioned that AI systems can be unreliable narrators of their own internal processes.

Google’s spokesperson told Ars the company was looking into the issue, but supplied no substantive explanation before publication.