A startup called Oumi found that Google’s AI assistant Google Gemini generates millions of factual errors every day. According to the analysis, one in ten responses contains inaccuracies, and at Google’s scale of five trillion queries per year, this results in more than 57 million incorrect answers per hour—or nearly one million per minute.

## The scale of Gemini’s accuracy problem

The startup Oumi tested Google AI Overviews using the SimpleQA benchmark, which includes 4,326 queries. In October 2025, when the Gemini 2 model was used for complex questions, accuracy was 85 percent. In February 2026, after updating to Gemini 3, the figure increased to 91 percent. Thus, despite improvements, one in ten responses is still wrong.

With an annual processing volume of about five trillion queries, even this accuracy rate means a massive flow of inaccurate information. Users receive tens of millions of incorrect answers per hour—hundreds of thousands of errors every minute.