Joe Biden won the 2020 US presidential election. That is a complete sentence, a settled fact, available free of charge from the Federal Election Commission and approximately ten thousand news outlets. It is also, Kyle Orland at Ars Technica discovered, a fact that Google’s AI Mode — the AI answer engine layered onto the most-used search engine on Earth — often would not produce on its own, at least in the United States. Ask it about recent presidential results and it offered not answers but errands: links to outside sources, an invitation to go look things up yourself.
The reticence is notable because the rest of the industry has no such affliction. Google’s own Gemini app answered readily. So did OpenAI’s ChatGPT, Anthropic’s Claude, and even the government-run AI model at America.gov. When the federal government’s chatbot is more forthcoming about election results than Google’s flagship search product, something has gotten inverted.
The behavior also got stranger the longer you poked at it. Ask the same questions a second time and AI Mode relented, naming the winner in all three elections where it had previously only provided links, this time citing Google Civics data. Access AI Mode from outside the country and the bashfulness disappeared entirely: VPN spot checks from Taiwan, Russia, France and Canada all got straightforward, accurate answers about recent US election results. Which is to say, a person in Russia could get a straight answer from Google’s AI about who won the American presidency, while a person in America got a hyperlink.
Asked to explain itself, AI Mode blamed “safety filters for recent and highly sensitive election topics [that] are very strict,” explaining 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 sounds plausible, with one large caveat, also flagged by Ars: AI models are notoriously unreliable narrators of their own internal machinery. Asking a chatbot why it did something is a bit like asking a magic 8-ball about its decision process.
Google’s official posture is real enough. Its Election Public Policy page says it “prioritize[s] surfacing authoritative election information directly within results,” including “direct links to official resources,” and a September blog post about “helping voters find reliable information” around the upcoming midterms pledged to “highlight official resources” wherever possible. On paper, the AI was doing policy. In practice, the policy decided that “who won an election six years ago” is live ammunition.
I think the interesting question is why that answer is treated as dangerous at all, and the timeline offers a hint. In 2023, Google-owned YouTube reversed its blanket ban on videos making false claims about the 2020 election, writing: “In the current environment, we find that while removing this content does curb some misinformation, it could also have the unintended effect of curtailing political speech without meaningfully reducing the risk of violence or other real-world harm.” In 2023 the corporate fear was that moderation would look like censorship. By 2026, the fear apparently extends to stating arithmetic without a chaperone from the FEC. A company that has been yelled at for years over election misinformation has converged on the safest possible position: know the answer, and decline to give it.
A Google spokesperson pointed Ars to a Bluesky post conceding that “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.”
There is a defensible version of all this — when in doubt, send people to the primary source. But “who won in 2020” is not edging up against doubt. A search engine that knows the answer, and hands it to anyone in France, Taiwan or Russia on the first try, but makes an American ask twice, has built a product that treats its own users as a hazard to be managed. The AI knows who won. It would just like the answer to come from literally anyone else.