One in ten Google AI answers is wrong. At scale, that's tens of millions of lies per day.
Google AI Overviews wrong 10 percent of the time, generating millions of incorrect answers daily
The New York Times, working with AI startup Oumi, tested Google's AI Overviews using SimpleQA — a benchmark of more than 4,000 questions with verifiable answers. The result: AI Overviews is right about 91% of the time. The other 9-to-10% are wrong. At Google's scale, that works out to hundreds of thousands of incorrect AI answers going out every minute of the day.
The test caught some specific failures. When asked for the date Bob Marley's former home became a museum, AI Overviews found Wikipedia, which listed two contradictory years, and confidently picked the wrong one. When asked about Yo-Yo Ma's induction into the Classical Music Hall of Fame, AI Overviews stated that there is no such thing as the Classical Music Hall of Fame. There is.
Google spokesperson Ned Adriance told the Times the study "has serious holes" and doesn't reflect real search behavior. Google also pointed out that AI Overviews typically runs on faster, cheaper Gemini Flash models rather than its best model — which answers more queries more quickly but is more prone to error. The accuracy did improve after Google upgraded from Gemini 2.5 to Gemini 3, jumping from 85% to 91%. Getting better is not the same as being reliable.
At the bottom of every AI Overview, Google includes a small disclaimer: "AI can make mistakes, so double-check responses." That note is doing a lot of work.