Google AI gets cagey when asked who won the 2020 elections

Ars Technica found that Google's AI Mode, asked directly who won the 2020 presidential election, pointed users to the Federal Election Commission instead of saying Joe Biden won. It did the same for 2024 and, for some testers, other recent cycles, and gave similar non-answers about Senate control after 2020. Meanwhile it answered older elections directly, and gave straight answers when accessed from Taiwan, Russia, France, and Canada through a VPN.

Google's own Gemini app, ChatGPT, and Claude all answered the question. Google says it is fixing the problem, describing a technical gap in how its systems verify 2020 election answers, and says the same issue affected 2022 Senate races.

Let's be plain about this: the 2020 election result is not a contested fact. It is certified history. When an AI search product answers 1992 with exact Senate seat counts but sends people to a government website for 2020, that inconsistency is the story, whatever the cause.

We're inclined to take Google's explanation seriously. The company says it checks election answers against news articles or AP data for recent races and historic records for older ones, and that 2020 fell into a gap. That is a believable failure in a layered system. The fact that AI Mode answered correctly on a second try, citing Google Civics data, also points to a brittle filter more than a deliberate policy.

But intent is not the only thing that matters here. In a country where some public figures still won't say who won 2020, a dominant search product that hedges on the same question reads as hedging, whether it means to or not. And the fact that the evasion appeared for US users but not foreign ones makes it worse, not better. It suggests caution tuned for the American political climate, the exact place where clear answers matter most.

This is the downside of over-cautious AI that rarely gets discussed. Safety filters on elections exist for good reasons. Nobody wants a model inventing vote counts. But a filter that treats settled facts as too sensitive to state ends up doing the misinformation's work for it. If that pattern scales across AI search, the systems people increasingly rely on for answers will go quiet precisely on the questions where silence is most corrosive.

Our read: useful AI should answer settled questions directly and save its caution for what is actually uncertain. Google has said it is fixing this. With US midterms coming, I'm watching to see that the fix holds and that the line between careful and evasive gets drawn in the right place.

Google's election policy pages say it prioritizes surfacing authoritative election information and linking to official resources, and a September Google post about the midterms said the same. Separately, in 2023 Google-owned YouTube reversed its ban on videos making false claims about the 2020 election, arguing that removal could curb political speech without meaningfully reducing harm.

When asked to explain itself, AI Mode cited strict safety filters on recent election topics, but AI models are unreliable narrators of their own internal workings.