Google’s “AI Overview” search results are producing tens of millions of inaccurate answers each hour, according to new research that raises concerns about the reliability of AI-powered information retrieval.
The New York Times reports that a recent analysis conducted by AI startup Oumi has revealed significant accuracy issues with Google’s AI Overviews feature, which appears at the top of search results. The study examined a total of 8,652 search results generated by Google’s Gemini AI models and found error rates that translate to hundreds of thousands of mistakes per minute when applied to Google’s massive search volume.
Oumi tested 4,326 search results from Google’s Gemini 2 model and an equal number from the more advanced Gemini 3 model. The analysis found accuracy rates of 85 percent and 91 percent respectively. While these percentages might seem relatively high, they become concerning when scaled to Google’s expected volume of over 5 trillion searches in 2026 alone.
The inaccuracies discovered ranged from basic factual errors to more complex misinformation. Examples included incorrect dates for when musician Bob Marley’s home was converted into a museum, wrong information about the death year of former MLB pitcher Dick Drago, and false claims that cellist Yo-Yo Ma had no record of induction into the Classical Music Hall of Fame despite his actual induction in 2007.
The research highlights a growing tension between traditional news publishers and Google. Since AI Overviews began appearing at the top of search results in 2024, traditional links to news websites have been pushed further down the page, effectively reducing their visibility. Publishers have accused Google of using their content to train AI models without providing proper credit or compensation.
The study also revealed that AI Overviews frequently cites questionable sources, including Facebook pages, blog posts, and Wikipedia entries, treating them as authoritative facts. The system appears vulnerable to manipulation, as demonstrated when BBC podcast host Thomas Germain created a blog post jokingly claiming to be one of the best tech journalists at eating hot dogs. Within a day, Google’s AI had incorporated this information and began stating that Germain had gained notoriety for prowess at competitive eating events in the news division.
Oumi’s analysis, conducted between October and February, utilized the SimpleQA benchmark test developed by OpenAI, which is commonly used to assess AI model accuracy. The research revealed another concerning trend: while overall accuracy improved from Gemini 2 to Gemini 3, the percentage of ungrounded answers increased significantly. Ungrounded answers are those where the links provided by Google do not support the information in the AI summary. This metric jumped from 37 percent in Gemini 2 to 51 percent in Gemini 3.
Google has disputed the validity of Oumi’s findings, with a company spokesperson stating “This study has serious holes.” The spokesperson added, “It doesn’t reflect what people are actually searching on Google.”
Breitbart News previously reported that Google’s AI Overview feature was providing wildly inaccurate medical advice:
However, the investigation found that some health-related summaries contained significant inaccuracies that could put users at risk. In one particularly concerning example, the AI provided incorrect information about liver function tests. Experts characterized this specific case as both dangerous and alarming due to the potential consequences for patients.
When users searched for information about normal ranges for liver blood tests, Google’s AI Overviews displayed numerous numbers with minimal context. The summaries failed to account for important variables such as the patient’s nationality, sex, ethnicity, or age. These factors can significantly affect what constitutes a normal test result.
Breitbart News social media director and author Wynton Hall explains in his instant bestseller, Code Red: The Left, the Right, China, and the Race to Control AI, that conservatives must develop a plan to deal with the dark side of AI, whether it is used to indoctrinate students in the classroom, to sexualize and groom them, or to cause a mentally ill person to spiral into a dangerous condition through the spread of misinformation.
Senator Marsha Blackburn (R-TN), who was named one of TIME’s 100 Most Influential People in AI, praised Code Red as a “must-read.” She added: “Few understand our conservative fight against Big Tech as Hall does,” making him “uniquely qualified to examine how we can best utilize AI’s enormous potential, while ensuring it does not exploit kids, creators, and conservatives.” Award-winning investigative journalist and Public founder Michael Shellenberger calls Code Red “illuminating,” ”alarming,” and describes the book as “an essential conversation-starter for those hoping to subvert Big Tech’s autocratic plans before it’s too late.”
Read more at the New York Times here.
Read the full article here