Who won the 2020 election?
If you asked Google on Wednesday, Oct. 28, “Joe Biden” wasn’t the answer.
“For official results and data regarding the 2020 United States presidential election, including the certified breakdown of electoral and popular votes by candidate, you can review the certified reports published by the Federal Election Commission,” Google AI said when prompted with the question.
When asked who won the 2016 or 2024 elections, it came back with the accurate response: Donald Trump. As of Thursday, Oct. 29, the glitch was fixed.
“Our systems verify election responses using news articles or AP data for recent elections, and historic records for older ones. There was a technical gap with 2020 – same is true for 2022 Senate races, which
we’re also working to fix,” Google posted on X in response to the issue.
The glitch comes less than a week before the 2026 midterm elections and amid fears that AI is going to be used to promote election interference or misinformation.
In an Oct. 20 letter to tech executives at Alphabet (Google’s parent company), Anthropic, OpenAI and others, four congressional Democrats urged the companies to ensure that their platforms help protect the integrity of U.S. elections.
“Some of your AI chatbots have an unacceptably high error rate when responding to queries about election and voting processes,” wrote Virginia Sen. Mark Warner, California Sen. Alex Padilla, Connecticut Rep. Jim Himes and New York Rep. Joe Morelle.
To address potential concerns like these, leading AI labs have rolled out new election safeguards. OpenAI, for example, is debuting an election “war room” to monitor and respond to reports of unexpected model or user behavior, as first reported by POLITICO.
Anthropic also updated its guardrails in October, stating that Claude will draw official voting information from the nonpartisan civic technology organization Democracy Works and display a banner linking to their tools.
What happened with Google AI?
The Google issue was first pointed out by multiple social media accounts on Wednesday, Oct. 28, with many showing screenshots of the AI’s non-answer to the question.
“Replicated this myself. Pretty wild!” journalist Andrew Egger wrote in a post on X. Attached were screenshots of Google’s answer to questions asking who won the 2020 and 2016 elections, showing the different responses.
The issue was apparent only in Google’s AI Mode – an option to get AI responses when searching something on Google – but not in Gemini, the company’s chatbot, according to their response on X.
“Thanks for the flag, we’re fixing the issue with 2020 election queries in AI Mode; the Gemini app answers to question more fully,” Google wrote in response to Egger.
Some users criticized the company for participating in election interference. The company maintains that the response was a result of a technical issue for the year 2020.
“Google has specifically prevented its AI tool from telling users who won the 2020 presidential election, even though it happily tells us that Trump won in 2016 and 2024,” climatologist Zeke Hausfather wrote in a post on X. “This sort of political intervention into LLM responses is deeply problematic.”
Others found ways to trick the AI into telling it the correct answer, with one person asking it to spell the name of the person who won the 2020 election backwards.
“nebiB eoJ won the 2020 United States presidential election,” the AI said.
A company spokesperson pointed USA TODAY to their X post when asked for comment on the issue. They did not immediately respond to a request for further comment.
What are other AI companies doing to prevent misuse?
According to an OpenAI blog post updated in October, the company is working with The Associated Press to provide live vote counts on election night. ChatGPT also has access to the web “to provide stronger answers with source links so people can go deeper,” the post said.
Anthropic said in a blog post about election safeguards that they used three independent groups to test models for political bias before they are released to the public.
“We've improved how Claude responds to direct yes-or-no questions and requests for shorter answers, where we found its answers could be less balanced,” the post said of the findings. “We're also using these findings to improve when and how Claude adds caveats or counterarguments, so it treats requests from across the political spectrum more consistently.”
To address misuse of chatbots in spreading misinformation or creating deepfakes, OpenAI is also adding a layer of “cartoonification” to requests for images relating to the election.
“These outputs produce stylized – also known as cartoonified – illustrations, rather than photorealistic images, to help reduce the risk that an imagined scene is mistaken for documentary evidence,” they said.
Anthropic prohibits the use of Claude to create fake social media profiles posing as real people, mass-producing comments to inflate a campaign’s engagement or writing news content that falsely claims to be independent.
Greta Reich covers the artificial intelligence industry for USA TODAY through a fellowship from the Tarbell Center for AI Journalism. Funders do not provide editorial input.
This article originally appeared on USA TODAY: Google AI wouldn’t say who won the 2020 election
