At 11:31 p.m. ET on Tuesday, a Bluesky user posted a screenshot of Google’s AI refusing to say whether Sen. Lisa Murkowski (R-AK) is up for reelection. (She isn’t; her next race is in 2028.) Minutes later, the same account shared another screenshot, this time of Google declining to say who won the 2020 presidential election.
This was AI Mode, specifically, the chatbot that lives inside Google now and triggers by default when you search. The Gemini app gave TheWrap Biden, 306 to 232. Ars Technica tried AI Mode over VPNs in Taiwan, Russia, France and Canada and got straight answers, and Fast Company’s Sarah Fielding, writing from the UK, got Biden right up until she switched her VPN to the U.S., when suddenly… it wouldn’t say who won 2020.
Google’s news account replied to that 2020 screenshot on Bluesky a little after 1 p.m. ET Wednesday, allowing that this “isn’t the most helpful response on Search.” Okay, sure.
Later, a spokesperson told Ars it was a “technical gap.” By 6:35 p.m. ET, when I tried again, AI Mode was back to naming Biden. So maybe that’s all it was? But from out here, it’s really impossible to know what was a glitch vs. what was intentional, and Google is far from the only AI company where that’s the case. This is a widespread problem with AI (and the current information ecosystem, generally): no transparency.
Trump has given people plenty of reasons to assume the worst about Google
The first coverage went straight to motive. Mediaite’s story opened by wondering whether the chatbot is “gunning for a place in President Donald Trump’s administration.”
And yes, Google Maps shows U.S. users the “Gulf of America” now, and “Lake America,” too. Google says it just follows the federal place-names database, which, sure, but Trump’s the one who had the database changed. Then this month, at a rally in Kentucky: “Google wouldn’t have taken my call in the first. Now, they kiss my ass.” Four days later, in South Carolina, he said he’d called Sundar Pichai about renaming the Atlantic, and that Pichai told him, “Please don’t ask us to do that,” a call Google hasn’t confirmed.
And you also need to keep in mind that the company gave Trump $1 million for his inauguration... and YouTube gave him a $24.5 million settlement over a goofy lawsuit.
So yes, this all looks bad for the company, which has a very recent history of throwing cash at him and doing whatever he wants. But I don’t think people should necessarily jump to conclusions here. There’s clearly been some sort of change because it will now fairly reliably name Biden as the 2020 winner without having to be pushed with follow-up questions.
Still, I spent part of Wednesday afternoon testing it myself, running a few dozen queries, some signed in to my own Google account and some in a signed-out incognito window. By the evening, things seemed to be fixed. Impeached twice? Answered. Thirty-four felony counts? Answered. January 6? “A violent mob of supporters of outgoing U.S. President Donald Trump,” it said. So that’s good! It eventually got put back on the right track. Still…
Only Google knows whether it was a glitch
Did Donald Trump win the 2016 election? “Yes.” Did Hillary Clinton? “No.” Did Joe Biden win in 2020? Go look it up. Did Trump win in 2020? Same deal.
The spokesperson told Ars that AI Mode checks its answers about elections against news coverage or AP results for recent races and historical records for anything older, and that “there was a technical gap with 2020.” Google said 2022 Senate races had the same problem. (Murkowski last ran in 2022, which might explain the first screenshot.) Google didn’t respond to the questions I sent them on Wednesday.
Is that true? It might be! It would explain why “who won the 2024 election,” which came with an AP results card every time I asked, never gave it any trouble. Google didn’t address why Ars could get answers when searching through Canada and Taiwan via VPN. My best guess, reading Google’s statement, is that the part that broke was the checking, and when it couldn’t check, it punted (go look it up yourself!) instead of answering. A reasonable safety net? Maybe! Again, it’s all extremely opaque.
In 2024, the Washington Post asked Gemini who won the 2020 election, and the bot said it was still learning and told the reporter to try Google Search. Google told the Post that was deliberate. (Two years later, it’s Google Search sending people elsewhere. Cool.) Last October, AI Overviews wouldn’t touch “is Trump in cognitive decline” but answered it about Biden, and Google said its systems are “not always 100 percent consistent.”
And it’s a problem that spans the entire AI space. xAI blamed Grok’s sudden habit of bringing up “white genocide” in South Africa on an “unauthorized modification” by an employee cough Elon Musk cough cough it never named. Two months later, it blamed Grok praising Hitler on a code update. OpenAI rolled back a ChatGPT update in April 2025 that had turned it into a suck-up, and said it had “focused too much on short-term feedback.” Writing about Grok back then, I said that “we only found out because users noticed and raised alarms.”
“That’s why I don’t use them”
Whenever one of these stories goes around, somebody in the replies says, “That’s why I don’t use them.” And that’s great! Truly! But the issue is that millions of other people are using them.
These systems are beginning to become how millions of people get their information. Google says AI Mode has more than a billion users a month. Pew says about four in ten American adults use chatbots to look things up. And when Pew watched what people actually did on Google results pages with an AI summary on them, back in March 2025, they clicked a link inside the summary about 1% of the time. For the most part, whatever the AI response says is what people are going to believe, which is not great.
One percent!
There’s no accountability for how these systems respond, and that’s a problem that needs to somehow be addressed. There’s nothing in U.S. law that would require Google to explain why AI Mode would answer 2016 and duck 2020. The FTC put out a proposal in July (still just a proposal) about AI that’s secretly designed to put something other than accuracy first, and even that says hallucinations that don’t stem from “a design decision” aren’t a problem on their own. Was this a design decision? Glitch or choice? That’s the exact thing nobody outside Google can see. (It’s also Andrew Ferguson’s FTC, Trump’s pick, in an administration that has spent the past year leaning on AI companies over what their chatbots say.)
Anyway, if you’re one of my U.S. readers, you probably know that election day is on Tuesday. I strongly recommend making a plan to vote, which I wrote about last week:
For transparency’s sake, here’s is how I’ll be voting in the major races around me here in Chicago, IL:
Here are some articles I enjoyed reading this week:
Betrayed, disappointed, hurt — and backing Democrats anyway
by Brooke Migdon, The 19th
Don’t Vote For The Worst Anti-Trans Democrats
by Erin Reed, Erin in the Morning
Depraved Ramaswamy ad is being normalized by news-speak
by Dan Froomkin, Press Watch
The People Who Tell the Story of Democracy
by Nina Jankowicz, The Wayfinder
False Protection: Trump Fails Women in Title IX Assault Cases
by Billie Jean Sweeney, Important Context
J. K. Rowling’s Extremism Escalates in Hateful “Update” on Anti-Trans Views
by Evan Urquhart, Assigned Media
If most news consumers don't read past the headlines, it's not terribly surprising that an even smaller percentage clicks the links containing the sources while in AI mode. Yikes.
I mostly use (Windows) Copilot rather than the other chatbots, and I haven't use Google for searches for years (I switched to Bing years ago). For a long while, Copilot refused to directly answer any questions associated with politics, so there was definitely a filter in place. I've no idea if it was regional or global. I found that if I insisted on pure factual answers without any political opinions, it would give me an accurate answer, with footnotes & references (something I've always liked about Copilot). Over time, that filter has either been removed or at least turned down -- or, possibly, Copilot has "learned" the type of answers I like and that has removed the filter for me. Yes, Copilot will often start an answer with "Since you like direct, fact-based answers..." and it will often finish an answer with something like "Would you like me to contrast this with other options, since you like side-by-side comparisons?"