Google AI Overviews: 1 in 10 Answers Wrong, Spreading 'Hundreds of Thousands of Lies' Per Minute

Google's AI-powered search feature is generating a staggering volume of misinformation. A new analysis, conducted by The New York Times with the help of AI startup Oumi, suggests that while Google's AI Overviews provides the correct answer about 90% of the time, the remaining 10% translates into hundreds of thousands of incorrect responses being delivered to users every single minute. This systemic error rate, even if seemingly small, represents a massive scale of potential falsehoods being disseminated by the world's dominant search engine.

The analysis used the SimpleQA evaluation, a common benchmark for testing the factuality of generative AI models. This test, originally released by OpenAI, consists of over 4,000 questions with verifiable answers. By probing Google's Gemini-powered AI Overviews with this dataset, the assessment found a persistent failure rate that, when extrapolated to Google's immense daily query volume, results in millions of erroneous outputs per hour. The feature, which appears at the top of search results, has faced consistent criticism for its 'scattershot accuracy' since its 2024 launch, despite Google's claims of gradual improvement.

This revelation places intense scrutiny on Google's core product integrity and its aggressive push to integrate generative AI into search. A 90% accuracy rate may sound high, but for a platform serving billions, the absolute number of errors is colossal, raising profound questions about the reliability of automated information at scale. The involvement of another AI firm, Oumi, in the testing underscores the industry's own focus on benchmarking these systems, even as they are deployed to the public. The pressure is now on Google to explain how it will contain this flood of algorithmic 'lies' before it further erodes user trust in search results.