Testing suggests Google’s AI Overviews tell millions of lies per hour (arstechnica.com)

Google's AI Overviews, powered by the Gemini model, has been found to provide accurate answers only 90 percent of the time, equating to hundreds of thousands of incorrect answers issued every minute. An analysis by The New York Times, conducted in partnership with AI startup Oumi, employed the SimpleQA evaluation—a rigorous test created by OpenAI to evaluate the factuality of generative models. The analysis revealed a notable improvement in accuracy from 85 percent with Gemini 2.5 to 91 percent with Gemini 3, yet the remaining 10 percent error rate implies that AI Overviews could generate tens of millions of incorrect responses daily. These inaccuracies raise significant concerns for the AI/ML community, as they challenge the reliability of AI-generated information in critical search engine applications. Several examples highlighted the model's tendency to furnish misleading or outright erroneous details, such as misidentifying the date a museum was established or fabricating nonexistent entities. As users often rely on AI Overviews for quick facts, the potential consequences of disseminating false information could undermine user trust in AI systems and prompt calls for heightened scrutiny and improvements in verification protocols within generative AI.