Tens of Millions of Errors Per Hour: Investigation Reveals the 'Accuracy Illusion' of Google AI Search

marsbitPublicado em 2026-04-13Última atualização em 2026-04-13

Resumo

A New York Times investigation, in collaboration with AI startup Oumi, reveals significant accuracy and reliability issues with Google's AI Overviews search feature. Testing over 4,300 queries showed the accuracy rate improved from 85% (powered by Gemini 2) to 91% (Gemini 3). However, given Google's scale of ~5 trillion annual searches, this 9% error rate translates to nearly 57 million incorrect answers generated hourly. A critical finding is the prevalence of "unsubstantiated citations." For correct answers, the rate of citations that do not support the AI's summary surged from 37% to 56% with the Gemini 3 upgrade, making it difficult for users to verify information. The AI heavily relies on low-quality sources, with Facebook and Reddit being among its top-cited websites. Furthermore, the system is highly manipulable. A BBC journalist successfully "poisoned" it by publishing a fabricated article; Google's AI began presenting the false information as fact within 24 hours. Google disputed the study's methodology, criticizing its use of the SimpleQA benchmark and an AI model (Oumi's own) to evaluate another AI. The company maintains its AI Overviews, combined with its search ranking systems, perform better than the underlying model alone. Critics note this defense does little to bolster user confidence in the feature's reliability.

Author: Claude, Deep Tide TechFlow

Deep Tide Guide: A recent test conducted by The New York Times in collaboration with AI startup Oumi shows that the accuracy rate of Google Search's AI Overviews feature is approximately 91%. However, given Google's scale of processing 5 trillion searches annually, this translates to tens of millions of incorrect answers generated every hour. More troublingly, even when the answers are correct, over half of the cited links fail to support their conclusions.

Google is disseminating misinformation on an unprecedented scale, and most people are completely unaware.

According to The New York Times, AI startup Oumi, commissioned by the publication, used the industry-standard test SimpleQA, developed by OpenAI, to evaluate the accuracy of Google's AI Overviews feature. The test covered 4,326 search queries, conducted in two rounds: one in October last year (powered by Gemini 2) and another in February this year (upgraded to Gemini 3). The results showed that Gemini 2's accuracy was about 85%, which improved to 91% with Gemini 3.

91% sounds good, but it's a different story when considering Google's massive scale. Google processes approximately 5 trillion search queries annually. With a 9% error rate, AI Overviews generates over 57 million inaccurate answers per hour, nearly 1 million per minute.

Correct Answers, Wrong Sources

More alarming than the accuracy rate is the issue of "unsubstantiated citations."

Oumi's data shows that in the Gemini 2 era, 37% of correct answers had the problem of "unsubstantiated citations," meaning the links attached to the AI summary did not support the information provided. After upgrading to Gemini 3, this proportion increased instead of decreasing, jumping to 56%. In other words, while the model gives correct answers, it is increasingly failing to "show its work."

Oumi CEO Manos Koukoumidis pointedly questioned: "Even if the answer is correct, how do you know it's correct? How do you verify it?"

The heavy reliance on low-quality sources by AI Overviews exacerbates this problem. Oumi found that Facebook and Reddit are the second and fourth most cited sources for AI Overviews, respectively. In inaccurate answers, Facebook was cited 7% of the time, higher than the 5% rate in accurate answers.

BBC Journalist's Fake Article "Poisons" Results Within 24 Hours

Another serious flaw of AI Overviews is its susceptibility to manipulation.

A BBC journalist tested the system with a deliberately fabricated false article. In less than 24 hours, Google's AI Overview presented the false information from the article as fact to users.

This means anyone who understands how the system works could potentially "poison" AI search results by publishing false content and boosting its traffic. Google spokesperson Ned Adriance responded by stating that the search AI feature is built on the same ranking and security mechanisms used to block spam, and claimed that "most examples in the test are unrealistic queries that people wouldn't actually search for."

Google's Rebuttal: The Test Itself Is Flawed

Google raised several concerns about Oumi's study. A Google spokesperson called the research "seriously flawed," citing reasons including: the SimpleQA benchmark itself contains inaccurate information; Oumi used its own AI model, HallOumi, to judge another AI's performance, potentially introducing additional errors; and the test content does not reflect real user search behavior.

Google's internal tests also showed that when Gemini 3 operates independently outside the Google Search framework, it produces false outputs at a rate as high as 28%. However, Google emphasized that AI Overviews, leveraging the search ranking system, performs better in accuracy than the model alone.

Nevertheless, as PCMag pointed out in a logical paradox: If your defense is that "the report pointing out our AI's inaccuracies itself uses potentially inaccurate AI," this likely does not enhance user confidence in your product's accuracy.

Perguntas relacionadas

QWhat was the accuracy rate of Google's AI Overviews feature as tested by Oumi, and how many errors does this translate to per hour given Google's search volume?

AThe accuracy rate of Google's AI Overviews was found to be 91% in the test. Given Google's annual volume of 5 trillion searches, this 9% error rate translates to over 57 million inaccurate answers generated every hour.

QAccording to the Oumi study, what was the trend in 'unsubstantiated citations' between the Gemini 2 and Gemini 3 versions of the AI Overviews?

AThe problem of 'unsubstantiated citations' (where the provided links did not support the AI's answer) increased from 37% with Gemini 2 to 56% with the upgraded Gemini 3.

QWhich low-quality websites were identified as major sources frequently cited by Google's AI Overviews?

AFacebook and Reddit were identified as the second and fourth most frequently cited sources by the AI Overviews feature.

QHow did a BBC journalist demonstrate the vulnerability of Google's AI Overviews to manipulation?

AA BBC journalist tested the system by publishing a deliberately fabricated article. Within 24 hours, Google's AI Overviews began presenting the false information from that article as a factual answer to user queries.

QWhat were Google's main criticisms of the Oumi study's methodology?

AGoogle criticized the study for having 'serious flaws,' stating that the SimpleQA benchmark itself contains inaccuracies, that using Oumi's own AI model to judge another AI could introduce errors, and that the test queries did not reflect real user search behavior.

Leituras Relacionadas

CPU Makes a Comeback to the Table, A $170 Billion "Power Seizure" Drama Begins

A new era is dawning for the server CPU (Central Processing Unit), driven by the shift from AI model training to large-scale reasoning and the rise of Agentic AI. This article explores how the CPU is reclaiming a central role in the AI data center. For years, the focus has been on the GPU (Graphics Processing Unit) for AI training. However, as AI moves to the inference and Agent phase—where tasks involve complex, multi-step reasoning, tool calls, and data management—the workload balance is flipping. Studies show CPUs now handle over 70% of the workload in Agentic AI, up from 10-30% in training. This is because Agent tasks generate massive intermediate data (KV Cache) that exceeds GPU memory, forcing it to be offloaded to the CPU's larger, more scalable memory pools. This increased importance is translating into market changes. Major players are taking note: NVIDIA launched its first standalone CPU line, Vera, based on ARM architecture and optimized for Agent performance. AMD doubled its server CPU market forecast to over $1200 billion by 2030. Analyst reports project the total server CPU market could reach $1700 billion by 2030, with AI-driven demand being a primary driver. Furthermore, the classic ratio of CPUs to GPUs in AI servers is rapidly changing, converging from 1:8 toward 1:1 for Agent deployments. This surge in demand has led to a rare industry-wide price increase of 10-15% for server CPUs from Intel and AMD, breaking a decade-long trend of "more performance for the same price." Demand is bifurcating into high-core-count CPUs for in-rack GPU support and moderate-core CPUs for standalone Agent task orchestration. In China, this global trend presents an opportunity for domestic CPU manufacturers like Hygon (海光信息) and Huawei Kunpeng, who are bolstered by both growing AI infrastructure needs and national policies promoting technological self-reliance ("xin chuang"). The maturity of their software ecosystems is also accelerating, evidenced by faster adaptation to new AI models. In conclusion, the narrative is shifting from a GPU-centric view to one where CPU-GPU synergy is critical. The CPU is no longer a peripheral component but a performance-defining bottleneck and a key growth driver in the AI hardware stack, opening a massive new market estimated in the hundreds of billions of dollars.

marsbitHá 7h

CPU Makes a Comeback to the Table, A $170 Billion "Power Seizure" Drama Begins

marsbitHá 7h

TechFlow Intelligence: AMD AI Director Publicly Criticizes Claude Code for "Becoming Dumber and Lazier", Trump Claims Full Ceasefire in Hormuz But Strait Still Has 80 Unexploded Mines

TechFlow Intelligence Report: This daily digest covers key developments in AI, crypto, hardware, and geopolitics. In AI, SK Telecom faces US export control scrutiny over its partnership with Anthropic, while a Gemini user reports being misled in a scam scenario, sparking safety debates. China's Z.AI launches the GLM-5.2 model, rivaling Claude Opus without NVIDIA chips. In crypto, Bithumb lists ReProtocol, and Upbit delists KernelDAO. On the hardware front, MIT researchers build a custom OS to study chips, ASML denies US claims its advanced lithography machines are in China, and Amazon considers selling its in-house AI chips. Apple's future A21 Pro chip may use TSMC's latest N2P process. Major tech issues include 10,000 GitHub repositories distributing malware and Apple patching a critical eavesdropping flaw in Beats earbuds. US stocks rise, led by semiconductors, with Intel surging 10.6%, while SpaceX falls 3.5%. Geopolitically, despite a US-Iran deal, the Strait of Hormuz remains risky with ~80 uncleared mines, stalling 80M barrels of oil on standby tankers. Iran postpones Switzerland talks, and Trump calls the agreement an "unconditional surrender." The report highlights a contrast: temporary geopolitical calm versus the ongoing, fundamental restructuring of tech supply chains and chip independence.

marsbitHá 7h

TechFlow Intelligence: AMD AI Director Publicly Criticizes Claude Code for "Becoming Dumber and Lazier", Trump Claims Full Ceasefire in Hormuz But Strait Still Has 80 Unexploded Mines

marsbitHá 7h

Trading

Spot
Futuros
活动图片