New Delhi: Google’s AI Overviews, launched in 2024 to deliver AI-generated summaries at the top of search results, have shown a high degree of accuracy but still produce a considerable number of incorrect responses due to the sheer volume of queries they process. An analysis by The New York Times, conducted in partnership with AI start-up Oumi, indicates that while AI Overviews are accurate approximately 90% of the time, Google’s handling of nearly five trillion searches annually results in tens of millions of incorrect answers each hour.
According to the study, about one in ten AI-generated responses may include false information. Furthermore, more than half of the accurate answers were categorized as “ungrounded,” meaning the sources cited did not completely support the information provided, complicating efforts for users to verify the responses. Oumi utilized the SimpleQA benchmark to evaluate thousands of queries, finding that accuracy improved from 85% with the Gemini 2 model to 91% following the rollout of Gemini 3. However, this increased accuracy has been accompanied by a rise in the proportion of ungrounded yet correct answers, highlighting ongoing challenges in AI systems’ interpretation and attribution of information.
Specific examples from the analysis illustrate how errors can arise even when sources are referenced. In one case, an AI Overview inaccurately claimed that Bob Marley’s home became a museum in 1987, despite records indicating it opened in 1986. Another example involved the system misidentifying the river that borders a city in North Carolina, deriving incorrect geographical information from a linked source. Instances also occurred where the AI provided partially correct answers but included misleading additional details or failed to recognize information even when linking to the correct source.
Google has acknowledged that its AI-generated summaries are not infallible and includes a disclaimer urging users to verify responses. However, the company has contested the findings of the analysis. “This study has serious holes,” said Ned Adriance, a Google spokesperson, asserting that the benchmark used in the evaluation contained inaccuracies and did not accurately reflect typical user searches.
These AI-generated Overviews have previously attracted scrutiny, particularly when incorrect information has appeared prominently in the search results. Following the Air India crash in Ahmedabad last year, an AI Overview mistakenly identified the aircraft involved, leading to public backlash before the response was removed.
As AI continues to evolve, the accuracy of these tools remains a critical focus for both developers and users. Google’s ongoing attempts to refine its AI capabilities underscore the broader industry challenge of balancing speed and accuracy in information delivery. While advancements like Gemini 3 have shown improvements, the persistent presence of ungrounded content and misinformation continues to pose significant risks to users relying on these summaries for accurate information.
See also
Anthropic Launches Claude Mythos Preview, Identifying Thousands of Critical Vulnerabilities
Meta Launches Muse Spark AI Model to Enhance Reasoning and Multimodal Interactions
Google DeepMind Accelerates AI Development by Merging Resources and Embracing Startup Culture
Large Language Models Market to Surge from $3.5B in 2025 to $25B by 2033 with 28% CAGR
Florida AG Launches Investigation into OpenAI Over FSU Shooter Chatlogs
