A recent study highlights that Google’s “AI Overview” feature produces millions of incorrect answers every hour, sparking concerns about the dependability of AI-generated search results.
A new analysis from AI startup Oumi unveiled notable issues with the accuracy of Google’s AI summary feature that appears at the top of search outputs. The research scrutinized 8,652 results generated by Google’s Gemini AI model, revealing an error rate that translates to potentially hundreds of thousands of mistakes each minute, given Google’s substantial search volume.
Oumi evaluated 4,326 results from both the Gemini 2 and the more advanced Gemini 3 model. The findings indicated accuracy rates of 85% and 91%, respectively. Although these figures sound fairly solid, they raise red flags when you consider Google’s projected search volume exceeding 5 trillion in 2026.
The inaccuracies ranged from basic factual errors to misleading information. For instance, there were incorrect dates regarding when Bob Marley’s residence became a museum, false details about former MLB pitcher Dick Drago’s death year, and misleading claims that cellist Yo-Yo Ma wasn’t inducted into the Classical Music Hall of Fame—despite being inducted in 2007.
The study also sheds light on the growing friction between conventional news publishers and Google. Since the introduction of AI summaries in 2024, traditional news links have been pushed further down search results, decreasing the visibility of news websites. Publishers express concerns that Google has used their content to train its AI without due credit or compensation.
The research found that the AI Overview often cited dubious sources, including Facebook posts, blog entries, and Wikipedia pages, treating them as credible information. A notable incident involved BBC podcast host Thomas Germain, who humorously claimed he was among the best tech journalists to eat a hot dog. Within a day, Google’s AI reflected this information, falsely attributing fame to Germain for his culinary expertise within the news sector.
Conducted between October and February, Oumi’s analysis employed the SimpleQA benchmark test created by OpenAI, a standard for measuring AI model accuracy. The study identified a troubling trend: while the overall accuracy improved from Gemini 2 to Gemini 3, the number of unsupported answers surged significantly. Specifically, the percentage of unsubstantiated answers rose from 37% in Gemini 2 to 51% in Gemini 3.
Google countered Oumi’s findings, with a spokesperson stating that significant flaws exist within the study and claiming it doesn’t accurately reflect actual user searches on Google.
Previously reported by Breitbart News, Google’s AI Overview had provided dangerously inaccurate medical advice.
Upon investigation, serious discrepancies in health-related summaries were found, raising alarms over user safety. For instance, incorrect information regarding liver function tests was noted, which experts deemed alarming due to potential ramifications for patients.
When users searched for normal ranges for liver blood tests, Google’s AI presented a slew of numbers lacking substantial context. Crucial factors like patient age, ethnicity, gender, and nationality were overlooked, which are vital in determining what constitutes a normal test result.
As Breitbart News social media director Wynton Hall mentioned in his book, conservatives argue that we need to address the darker aspects of AI usage, whether it’s in education, media, or healthcare due to misinformation risks.
Sen. Marsha Blackburn (R-Tenn.) recognized Hall as one of the most influential figures in AI, highlighting his insights into managing AI’s potential without compromising vulnerable populations. Journalist Michael Shellenberger further asserted that Hall’s book is essential reading for those aiming to counter Big Tech’s authoritarian tendencies.
For additional information, new york times has more details here.
