Google AI Overviews Found Spreading Millions of Lies Hourly
▼ Summary
– Google’s AI Overviews feature, powered by Gemini, now appears at the top of search results but has faced criticism for accuracy issues since its 2024 launch.
– A New York Times analysis, conducted with startup Oumi, found AI Overviews provides correct answers about 90% of the time using the SimpleQA evaluation benchmark.
– Despite the high accuracy rate, the error rate means the system generates tens of millions of incorrect answers daily when scaled to all Google searches.
– The analysis showed accuracy improved from 85% with the Gemini 2.5 model to 91% following the Gemini 3 update.
– The report highlights specific errors, such as providing a wrong date for a museum opening and incorrectly stating a hall of fame does not exist.
When users search for information on Google, they now encounter AI Overviews at the top of the results page. This Gemini-powered feature has faced significant criticism since its 2024 debut for its inconsistent accuracy. While it generally provides correct answers and is improving, a recent analysis reveals a concerning scale of error. According to a study by The New York Times, AI Overviews delivers accurate information roughly 90 percent of the time. The critical takeaway is the remaining 10 percent, which translates to hundreds of thousands of incorrect responses generated every minute across the platform.
The Times collaborated with the AI startup Oumi for this evaluation. Using the SimpleQA evaluation framework, a standard benchmark for assessing generative AI factuality, Oumi tested the system’s performance. This framework, introduced by OpenAI in 2024, consists of over 4,000 questions with definitive, verifiable answers. Oumi first ran the test when Google’s Gemini 2.5 model was active, recording an 85 percent accuracy rate. After the release of the Gemini 3 update, a rerun of the test showed improvement, with AI Overviews answering 91 percent of questions correctly. However, applying this error rate to Google’s immense search volume indicates the feature produces tens of millions of false statements daily.
Specific examples from the report illustrate how these errors manifest. One query asked for the date Bob Marley’s former home was converted into a museum. The AI Overview cited three sources, two of which contained no date information. The third source was a Wikipedia page listing two conflicting years; the AI confidently selected the incorrect one. In another test, the system was prompted to provide the date of cellist Yo Yo Ma’s induction into the Classical Music Hall of Fame. While it correctly cited the hall’s official website listing Ma’s induction, it simultaneously claimed no such institution exists, creating a direct contradiction. These instances highlight the system’s ongoing struggle with source verification and logical consistency, even as its overall accuracy climbs.
(Source: Ars Technica)
