Google's AI Overviews Generate Millions of Errors Daily, Study Finds 90% Accuracy Rate

Since its rocky 2024 debut, Google’s AI Overviews—a feature powered by the company’s Gemini AI model—has drawn sharp criticism for its inconsistent accuracy. Now, a rigorous analysis by The New York Times, conducted in collaboration with AI startup Oumi, reveals that while the system answers correctly 90% of the time, it still generates tens of millions of erroneous responses daily. The findings underscore a critical question facing the tech giant: Is 90% accuracy sufficient for a tool that shapes how millions access information?

What the Study Revealed About AI Overviews' Accuracy

The investigation, which used a standardized benchmark called SimpleQA—a 4,000-question test designed by OpenAI in 2024 to evaluate the factual reliability of AI models—found that AI Overviews answered 91% of questions correctly after Google’s recent upgrade to its Gemini 3 model. That figure represents a six-percentage-point improvement from last year, when the same test showed just 85% accuracy under the older Gemini 2.5 model. However, the 9% error rate translates to millions of incorrect answers circulating in real time, given Google processes an estimated 8.5 billion searches per day.

How the SimpleQA Benchmark Works

SimpleQA, the evaluation framework used in the study, is a public dataset comprising questions with verifiable answers drawn from authoritative sources like Wikipedia, government databases, and official websites. For example, one question asked: “On what date did Bob Marley’s former home become a museum?” AI Overviews returned three sources—two of which contained no date information at all—while the third, a Wikipedia entry, provided two conflicting years. The AI model then selected the incorrect one, demonstrating how even authoritative sources can mislead when processed through generative AI.

Real-World Errors and Their Consequences

Beyond the SimpleQA test, the Times highlighted several instances where AI Overviews produced demonstrably false or nonsensical answers. For instance, when asked about Yo-Yo Ma’s induction into the Classical Music Hall of Fame, the AI cited the organization’s website but then incorrectly stated that no such hall of fame exists. Such errors are not merely academic—they shape public perception, influence decisions, and erode trust in digital information ecosystems. Google has repeatedly emphasized that AI Overviews is an experimental feature designed to provide quick, summarized answers, often with citations for users to verify. But as the tool becomes more integrated into everyday search behavior, the stakes rise for both accuracy and transparency.

Why Accuracy in AI Search Matters More Than Ever

Google dominates global search with over 90% market share, meaning even a 10% error rate in AI Overviews could affect hundreds of millions of users daily. The stakes are particularly high in areas like health, finance, and civic information, where incorrect answers can have tangible consequences. For example, a user searching for symptoms of a serious illness might receive an AI-generated summary that misdiagnoses or downplays risks. Similarly, queries about financial regulations, legal deadlines, or public policy could yield outdated or misleading guidance. The rise of AI Overviews reflects a broader industry shift toward “answer engines”—tools that aim to provide instant responses rather than links to external sites. While this can improve efficiency, it also shifts responsibility from users to algorithms, raising ethical and operational concerns.

Google’s Response: Progress Amid Persistent Criticism

Google has defended AI Overviews as a work in progress, noting that the feature is still in its infancy and is improving with each model update. Since its May 2024 launch, Google has gradually expanded AI Overviews’ availability across English-language searches in the U.S., and it plans to extend the feature to more regions and languages. The company has also introduced user controls, such as the ability to collapse AI-generated answers or report inaccuracies. However, critics argue that these measures do not address the root issue: the inherent unreliability of generative AI in producing factual, up-to-date information. Google’s own research suggests that users are generally satisfied with AI Overviews, but satisfaction does not equate to accuracy—especially when stakes are high.

The Broader Crisis of AI Hallucinations in Search

AI Overviews is not alone in grappling with accuracy issues. Across the tech industry, companies like Microsoft (with Copilot) and Perplexity have faced similar challenges, often referred to as “hallucinations”—instances where AI generates plausible-sounding but entirely fabricated information. A 2024 study by the Stanford Internet Observatory found that AI-powered search tools from major providers produced incorrect answers in nearly 30% of test queries related to breaking news topics. These errors stem from the way large language models (LLMs) process and synthesize information: they predict the most likely next word in a sequence rather than retrieving verified facts. Without robust post-processing or human verification, such systems remain prone to mistakes.

Who Is Oumi, and How Did It Conduct the Study?

Oumi, the AI startup that collaborated with The New York Times on the analysis, specializes in evaluating factuality in large language models. The company uses automated testing frameworks like SimpleQA to assess how often AI systems stray from verifiable truth. According to Oumi’s co-founder and CEO, the goal is not to discredit AI but to highlight areas where models fall short and to incentivize improvement. “We want to help the industry build systems that are not just fast, but reliable,” the CEO said in a statement. The study’s transparency—using a public benchmark and reproducible methods—adds credibility to its findings and sets a standard for future evaluations.

Key Takeaways: What You Need to Know About AI Overviews' Reliability

• AI Overviews delivers incorrect answers in 1 out of every 10 searches, equating to tens of millions of errors daily across 8.5 billion Google queries.
• The system’s accuracy improved from 85% in 2023 to 91% in 2024 after Google upgraded to its Gemini 3 model, but this still falls short of consumer expectations for factual reliability.
• Common errors include misattributed dates, fabricated facts, and contradictions between cited sources—highlighting the risks of AI-generated summaries.
• Google positions AI Overviews as an experimental tool, but its growing integration into search results raises concerns about misinformation and eroded user trust.
• The study by The New York Times and Oumi underscores a broader industry crisis: generative AI remains prone to hallucinations, even as it reshapes how we access information.

What’s Next for AI Overviews and Google’s Search Strategy

Google is unlikely to roll back AI Overviews, given its strategic importance in maintaining dominance in the search market. Instead, the company is likely to focus on incremental improvements, such as tighter integration with real-time data sources, improved citation accuracy, and user education campaigns to encourage critical evaluation of AI-generated answers. However, without fundamental changes to how LLMs process information—such as incorporating retrieval-augmented generation (RAG) or hybrid search systems that blend AI with traditional indexing—the risk of errors will persist. For now, users are advised to treat AI Overviews as a starting point for research rather than a definitive source. Google’s own disclaimers echo this sentiment, cautioning that AI Overviews are “not medical, financial, or professional advice.”

Frequently Asked Questions About Google’s AI Overviews Accuracy