Google's AI answers miss 10%
An analysis found Google’s AI-generated search overviews are wrong about 10% of the time, which — given Google’s scale — translates into millions of incorrect answers delivered hourly. Reports also say correct summaries often lack clear source support, raising product-level tradeoffs between convenience and trust. (popsci.com) (newsbytesapp.com)
Google now puts an artificial intelligence summary above many search results, so instead of 10 blue links you often get one polished paragraph first. The problem is that a polished paragraph can be wrong in a way a list of links usually is not. (nytimes.com) A New York Times analysis with the startup Oumi found Google’s search summaries answered about 90 percent of tested questions correctly, which leaves about 10 percent wrong. At Google’s scale, that small-looking miss rate can turn into tens of millions of bad answers in a single hour. (nytimes.com) (searchengineland.com) (blog.google) The test used SimpleQA, a benchmark built to check short factual answers with clear right-or-wrong outcomes. Oumi also compared October 2025 results with February 2026 results and found the newer system improved, but still missed roughly 1 in 10. (arstechnica.com) (newsbytesapp.com) The harder part is not only wrong answers. The Times reported that even correct summaries often did not clearly show where a claim came from, which means readers can see a confident answer without an easy way to check the receipt. (nytimes.com) That source problem shows up in what the system cites. Popular Science reported that Facebook and Reddit were among the most-cited sources in the analysis, and inaccurate answers cited Facebook more often than accurate ones did. (popsci.com) Google has warned users about this from the start. Its own help page says artificial intelligence responses in Search may include mistakes, even as the company keeps expanding the feature to more countries, more languages, and more people. (support.google.com) (blog.google) The rollout has been huge. Google said in May 2025 that AI Overviews had expanded to more than 200 countries and territories, and by its own count the feature had scaled to more than 1.5 billion users. (blog.google 1) (blog.google 2) Google is also giving the summaries a bigger job. In February 2026, it said Gemini 3 became the default model for AI Overviews and added follow-up questions, which moves Search another step from “find a page” toward “trust this answer.” (blog.google) That shift changes the economics of the web as much as the interface. If a summary answers the question on the page, fewer people need to click through to the site that did the reporting, even when that site is the reason the answer exists. (searchenginejournal.com) (blog.google) Google says people like these summaries, search more often with them, and see more links on the page. Critics are looking at the same product and seeing a machine that can flatten careful reporting, shaky forum posts, and half-checked facts into one voice that sounds equally sure about all of them. (blog.google) (mittechnologyreview.com)
