Google AI Mode hedges on who won the 2020 US election; Google says it is looking into it
Google's AI Mode in Search has been answering "who won the 2020 election?" by pointing users to official sources rather than naming Joe Biden. Google says Gemini answers fully, that the Search response isn't the most helpful, and that it is working to address it.
1 / 2
The story, neutrally told
Left · 2Google's AI Mode, the AI-powered feature in Search, has been giving indirect answers when asked who won the 2020 US presidential election, telling users to consult official sources instead of naming Joe Biden. Ars TechnicaLC “If you ask Google’s AI Mode the same question, you might get an evasive answer that simply tells you to consult with outside sources.” Read at Ars Technica ↗ The A.V. ClubLC “Google’s AI-powered search feature wouldn’t directly answer a very straightforward question about political history.” Read at The A.V. Club ↗ Left · 2The problem was first noticed by Bluesky users. Ars Technica, which tested it after seeing complaints on social media, got a reply directing it to the official report of the Federal Election Commission (FEC). The A.V. Club quoted a similar reply pointing to "official election resources" or the FEC website. Ars TechnicaLC “after seeing complaints about cagey answers on social media” Read at Ars Technica ↗ The A.V. ClubLC “please consult official election resources or visit the Federal Election Commission (FEC) website.” Read at The A.V. Club ↗ Left · 1The A.V. Club reported that AI Mode kept deflecting follow-up questions and directing users to "official U.S. government archives or election certification resources." The A.V. ClubLC “Even when asked follow-ups, AI Mode continued dodging questions and directing users to “consult official U.S. government archives or election certification resources.”” Read at The A.V. Club ↗
Left · 2Ars Technica said the hedging was not limited to 2020. In its tests, AI Mode also gave indirect answers for 2024 and 2012, but gave detailed, correct winners for every other presidential election back to 2000. The A.V. Club, citing Fast Company, said AI Mode had no trouble saying Donald Trump won in 2024, so the two outlets' accounts of 2024 differ. Ars TechnicaLC “My testing resulted in a similarly indirect answer when asking about the 2024 presidential election” Read at Ars Technica ↗ The A.V. ClubLC “As Fast Company noted, suspiciously, Google AI also had no issue recalling that Trump won the 2024 election.” Read at The A.V. Club ↗ Left · 1Ars Technica staffers found the pattern varied by user: one got immediate answers only for 2016 and 2024, while others got outside links for every cycle back to 2000. Questions about which party controlled the Senate after the 2020 elections also drew a referral to congressional records, while the same question about 1992 got a direct answer (Democrats, 57 to 43). Ars TechnicaLC “one staffer reported only getting immediate answers from AI Mode when asking about the 2016 and 2024 presidential elections”“AI Mode simply told me straight out that it led to Democrats holding a 57 to 43 majority in the Senate” Read at Ars Technica ↗ Left · 2Both outlets said Google's Gemini app answers the question directly. The A.V. Club said Gemini lists Biden and gives his electoral and popular vote totals against Trump's. Ars Technica said ChatGPT, Claude and the government-run America.gov model also answer correctly. The A.V. ClubLC “the AI assistant immediately lists Biden and provides more context on his total electoral and popular votes compared with Donald Trump” Read at The A.V. Club ↗ Ars TechnicaLC “The same goes for other major chatbots like OpenAI’s ChatGPT, Anthropic’s Claude, and even the government-run AI model at America.gov.” Read at Ars Technica ↗
Left · 1Ars Technica reported that the hedging seemed to disappear outside the US. Spot checks by VPN from Taiwan, Russia, France and Canada returned accurate, straightforward answers. Repeating the questions in the US also produced direct answers, citing Google Civics data. Ars TechnicaLC “Spot checks via VPN of AI Mode results accessed from Taiwan, Russia, France, and Canada all got accurate and straightforward answers”“citing Google Civics data in the latter response” Read at Ars Technica ↗ Left · 1When Ars Technica asked AI Mode why it hadn't answered the first time, it blamed strict "safety filters for recent and highly sensitive election topics" and a system built to send users to official databases to avoid showing unverified data. Ars cautioned that AI models are unreliable at describing their own processes. Ars TechnicaLC “AI models have proven unreliable at reporting on their own internal processes, so it’s hard to take that as the final word.” Read at Ars Technica ↗ Left · 1Google's election policy page and a September blog post about the midterms say it will prioritise authoritative election information and highlight official resources. In 2023, Google-owned YouTube reversed its policy of removing videos that made false claims about the 2020 election. Ars TechnicaLC “In a September Google blog post about “helping voters find reliable information” in the period surrounding the upcoming midterm elections”“In 2023, Google-owned YouTube reversed a policy barring all videos that made false claims about the 2020 election.” Read at Ars Technica ↗
Left · 2Google's news-updates account replied on Bluesky, and a Google spokesperson pointed Ars Technica to that reply: "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." The A.V. Club said the issue persisted at the time of its publication. Neither outlet reports a cause confirmed by Google. Ars TechnicaLC “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.” Read at Ars Technica ↗ The A.V. ClubLC “At the time of this article’s publication, this issue remains” Read at The A.V. Club ↗
Every sentence links to the reporting it rests on. The pill in front of each says where its sources sit: Left, Centre or Right when one side supplies at least half of them, Mixed when they are evenly split. The number is how many outlets it cites.
Left2 outlets
- Framing
- Both outlets treat it as an odd and embarrassing failure of Google's AI search. Ars Technica leads with systematic testing and compares the evasion to politicians' "weasely" answers on 2020. The A.V. Club leads with a sardonic, shorter take on a "puzzling glitch" spotted on Bluesky.
- Emphasis
- Ars Technica stresses the inconsistency across years and users, the absence of the problem abroad and in rival chatbots, and Google's stated election-information policies. The A.V. Club stresses that the problem persists and that Gemini answers fine.
- Leaves out or plays down
- The A.V. Club gives little on the scope beyond 2020, the geographic difference, or Google's election policies. Neither outlet has a confirmed technical cause from Google.
- Charged language
- “weasely”“cagey”“puzzling glitch”“suspiciously”
- For example
-
“you might get some weasely answers that fail to directly acknowledge Joe Biden’s win” — Ars Technica
“In a puzzling glitch recently noticed by Bluesky users” — The A.V. Club
Centre0 outlets
No centre outlet in our sources has covered this story yet.
Right0 outlets
No right outlet in our sources has covered this story yet.
What every side reports
- Google's AI Mode did not directly name the winner of the 2020 US presidential election and pointed users to official sources such as the FEC.
- Google's Gemini app answers the question directly.
- Google said the Search response isn't the most helpful and that it is working to address it.
Where accounts differ
-
Whether AI Mode answered about the 2024 election
- Left
- Ars Technica found AI Mode gave an indirect, FEC-referral answer for 2024. The A.V. Club, citing Fast Company, said AI Mode had no issue saying Trump won in 2024.
-
Why AI Mode hedges
- Left
- AI Mode told Ars it was due to strict safety filters on recent election topics, but Ars said such self-explanations are unreliable. The A.V. Club calls it a "glitch". Google has given no confirmed cause.
Google organisation
Google says Gemini answers the question fully, that the AI Mode response in Search is not the most helpful, and that it is looking into it. Its published election policies say it prioritises authoritative information and links to official resources.
“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.” — Ars Technica
““prioritize[s] surfacing authoritative election information directly within results”” — Ars Technica
Left2 articles
-
Google AI gets cagey when asked who won the 2020 elections
Critical Detailed testing-based report that treats the evasiveness as an inconsistent, possibly policy-driven failure, while cautioning against trusting the AI's own explanation.
-
Google's AI declines to say who won 2020 presidential election
Mixed Short, wry aggregation of the Bluesky complaints and Fast Company's findings that calls the behaviour a glitch confined to AI search.
Centre0 articles
No coverage yet.
Right0 articles
No coverage yet.
- 28 Oct 17:46 First Ars TechnicaLC Google AI gets cagey when asked who won the 2020 elections
- 28 Oct 21:08 +3h 22m The A.V. ClubLC Google's AI declines to say who won 2020 presidential election
Times are when each article was published, or when we first saw it if the outlet gave no time.
Ask about this story
Loading…