Since its launch in 2024, the AI Overviews feature at the top of Google Search has been plagued by concerns over accuracy. Despite continuous upgrades to the underlying models, the latest analysis shows that the tool still suffers from significant "hallucinations."
According to an assessment conducted by The New York Times in collaboration with AI startup Oumi, Google’s AI Overviews are accurate about 90% of the time. While that figure may sound high, the sheer scale of Google Search traffic means that one in every ten answers is incorrect, resulting in tens of millions of pieces of false information being pushed to users daily.
Accuracy Bottlenecks and Testing Details
The assessment utilized OpenAI’s SimpleQA benchmark, which includes over 4,000 questions with verifiable answers. Testers evaluated AI Overviews following the Gemini 3 model update and found an accuracy rate of 91%, an improvement over the 85% seen during the Gemini 2.5 era.
Nevertheless, errors remain frequent. When asked about the founding date of the Bob Marley Museum, AI Overviews confidently provided an incorrect year, despite the Wikipedia page it cited containing conflicting information. In another instance, while referencing a webpage about Yo-Yo Ma’s induction into the Classical Music Hall of Fame, the AI claimed that the "Classical Music Hall of Fame" does not exist.
Addressing these concerns, Google spokesperson Ned Adriance stated that the company believes the SimpleQA benchmark itself contains errors. Google maintains that its AI models are constantly evolving and are designed to provide users with an increasingly precise search experience.
Although Google is attempting to improve reliability through technical iterations, the inherent uncertainty of AI-generated content remains difficult to eliminate. For users who rely on search engines to find factual information, this "one-in-ten" error rate remains a risk that cannot be ignored.
