Testing suggests Google's AI Overviews tell millions of lies per hour

2026-04-07 · Source: AI - Ars Technica · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Emerging Technologies & Innovation · Depth: Novice, short

Summary

A new analysis by The New York Times, in collaboration with the startup Oumi, assessed the accuracy of Google's Gemini-powered AI Overviews, finding it to be correct 90 percent of the time. This evaluation utilized the SimpleQA benchmark, a list of over 4,000 questions with verifiable answers. Initial tests with Gemini 2.5 showed an 85 percent accuracy rate, which improved to 91 percent after the Gemini 3 update. Extrapolating this 10 percent error rate to all Google searches suggests AI Overviews generates tens of millions of incorrect answers daily. Google disputes the findings, with spokesperson Ned Adriance claiming SimpleQA contains incorrect information and that the study "doesn't reflect what people are actually searching on Google," preferring its own SimpleQA Verified test.

Key takeaway

For CTOs and VPs of Engineering evaluating AI search integration, your teams should recognize that even 90% accuracy in AI Overviews means a significant volume of errors at scale. You must implement robust verification layers or clearly communicate the potential for inaccuracies to end-users, especially for critical information, rather than relying solely on the AI's summary.

Key insights

Google's AI Overviews achieve 90% accuracy on SimpleQA, but its 10% error rate translates to millions of daily inaccuracies.

Principles

• AI model evaluation is complex and often subjective.
• Factuality benchmarks vary significantly across companies.
• Grounding AI with web data improves accuracy over naked models.

Method

The analysis used the SimpleQA evaluation, feeding over 4,000 verifiable questions to AI Overviews, and then comparing its answers to known facts to determine accuracy percentages.

In practice

• Use SimpleQA for generative model factuality testing.
• Consider Gemini Flash for faster, less accurate AI Overviews.
• Cross-reference AI Overviews with source links.

Topics

• AI Overviews
• Gemini Models
• SimpleQA Evaluation
• AI Accuracy Assessment
• Generative AI Factuality

Best for: CTO, VP of Engineering/Data, Director of AI/ML, Tech Journalist, AI Product Manager, MLOps Engineer

Related on AIssential

• Google’s AI Can’t Spell Google. That’s Not a Joke. — LLMs fundamentally struggle with character-level tasks like spelling due to tokenization, impacting their reliability in factual search.
• From Exams to Escape Rooms: How We Learned to Test AI — AI evaluation is a dynamic process, constantly adapting to measure increasingly complex model capabilities beyond narrow task performance.
• Did Google’s AI agents really build an operating system for $916? — Google's AI agent OS claim lacks transparency, highlighting the need for rigorous "open-world evaluation" methodologies for complex AI tasks.
• Google’s AI Overviews Can Scam You. Here’s How to Stay Safe — Google's AI Overviews, while appearing authoritative, can generate misleading content from scraped web data.
• DailyReport: An Open-ended Benchmark for Evaluating Search Agents on Daily Search Tasks — DailyReport offers a user-centric benchmark for Search Agents, evaluating real-world tasks via cascade rubrics and user preference scores.
• GPT-5.5 tops benchmarks but still hallucinates frequently at a 20 percent higher API cost — Higher compute in LLMs does not automatically reduce hallucinations or improve reasoning against nonsense.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI - Ars Technica.