Sign up to see the future, today<\/p>\n
Can\u2019t-miss innovations from the bleeding edge of science and tech<\/p>\n
Google\u2019s AI Overviews<\/a> are peddling misinformation on a scale that may be virtually unprecedented in human history.<\/p>\n A recent analysis conducted by the AI startup Oumi at the behest of The New York Times<\/a> found that the AI-generated summaries, which appear above Google search results, are accurate around 91 percent of the time.\u00a0<\/p>\n In a sense, that may sound like an impressive figure. But here\u2019s an even more impressive one: five trillion. That\u2019s roughly the number of search queries that Google processes every year, translating to tens of millions of wrong answers that the AI Overviews are providing every hour \u2014 and hundreds of thousands every minute, the analysis calculated.<\/p>\n In other words, Google has created a misinformation crisis. Studies have shown that people tend to trust what an AI tells them without question, with one report finding<\/a> that only 8 percent of users actually double checked an AI\u2019s answer. Another experiment found<\/a> that users still listened to AI when it gave them the wrong answer nearly 80 percent of the time \u2014 a grim trend the researchers dubbed \u201ccognitive surrender.\u201d<\/p>\n Large language models adopt an authoritative tone and can confidently present fabricated information as fact when it can\u2019t immediately glean a straight answer. Add the convenience that Google\u2019s AI Overviews offer, and it\u2019s easy to imagine untold numbers of users taking its summaries at their word.<\/p>\n Oumi conducted the analysis using a test called SimpleQA, a widely used benchmark for AI accuracy in the industry which was designed by OpenAI. The first round of tests, conducted in October, used a version of the AI Overviews powered by Google\u2019s Gemini 2 model. A follow-up conducted in February tested the feature after it was switched to Gemini 3, its much-hyped upgrade<\/a>.<\/p>\n Each round of tests involved 4,326 Google searches. Gemini 3 came out the more accurate model, giving a factually sound response 91 percent of the time. Gemini 2 performed significantly worse, at just 85 percent accurate.<\/p>\n On the one hand, it shows that the models are improving. On the other, it shows that Google was willing to foist a model on its userbase that was even more prone to hallucinating, in an ongoing experiment that\u2019s still misinforming hundreds of millions of people.<\/p>\n Google called the analysis flawed. \u201cThis study has serious holes,\u201d Ned Adriance, a Google spokesman, told the NYT in a statement. \u201cIt doesn\u2019t reflect what people are actually searching on Google.\u201d<\/p>\n Yet Google\u2019s own tests paint a no less damning picture, the reporting notes. In an internal analysis of Gemini 3, it found that the AI model produced incorrect information 28 percent of the time. Google claims, however, that AI Overviews are more accurate because they draw on Google search results before answering.<\/p>\n The improvement between Gemini 2 and Gemini 3 may be papering over a more serious flaw. In the Oumi analysis, Gemini 2 provided answers that were \u201cungrounded\u201d 37 percent of the time, meaning the AI Overviews cited websites that didn\u2019t support the information they provided. But with Gemini 3, this jumped to 56 percent. On top of suggesting that the AI is pulling facts out of thin air, ungrounded responses make it difficult for users to verify the AI\u2019s claims.<\/p>\n