A new study found that Google’s AI Overviews feature provides accurate, reputable responses roughly 91% of the time. This sounds great, but actually, it also means that tens of millions of answers churned out every hour are erroneous or at least questionable.

True, the fact that the AI startup Oumi found Google’s AI-generated summaries, which appear above search results, to be accurate more than 9 times out of 10, seems impressive.

However, another figure is even more stunning – 5 trillion. That’s the number of search queries Google, a tech giant transforming from a curator of information into a publisher, processes every year.

And that means that 9% of those AI-powered responses spit out inaccurate or plainly wrong results, resulting in what some would call an unprecedented misinformation-peddling campaign.

The math is quite straightforward, even if we round up the numbers, say Oumi researchers whose study is shared by The New York Times.

Google is now processing 5 trillion searches a year. This means it provides tens of millions of incorrect answers every hour, or hundreds of thousands of inaccuracies every minute.

Oumi analyzed the accuracy of AI Overviews using a benchmark test called SimpleQA, which is used across the industry to measure AI systems. The startup tested the system last October when Gemini 2 was providing results, and then again in February, after the model was upgraded to Gemini 3.

Oumi’s analysis focused on 4,326 complex Google searches. The company found that the results were accurate 85% of the time with Gemini 2 and 91% of the time with Gemini 3, widely considered the least hallucinatory AI model.

According to Oumi, it’s also concerning that more than half of the accurate responses were “ungrounded,” meaning they linked to websites that didn’t completely support the information they provided. This makes it difficult to check AI Overviews’ accuracy.

Adding to the confusion, Google’s system may generate a new response to each query. If Google Search receives the same query at separate times – even seconds apart – it may produce one answer that’s accurate and another that’s not.

“Even when the answer is true, how can you know it is true? How can you check?” asked Oumi CEO Manos Koukoumidis.

This could be the beginning of a huge misinformation crisis. That’s because, as multiple studies have sadly shown, people tend to trust what an AI tells them without question.

Yes, Google has added a fine-print disclaimer to its AI Overviews feature, saying “AI can make mistakes, so double-check responses.”

But at least one report found last year that only 8% of users actually double-checked an AI’s answer. Another more recent study found that people still listened to AI 80% of the time when the model gave them the wrong answer.

In short, many people simply don’t understand the sad reality that AI chatbots, whether integrated into search engines or not, are still deeply flawed.

---

Unlock more exclusive Cybernews content on YouTube.