Google: More Accurate AI but Verification in Question

Google Under the Microscope: A Study on the Accuracy of Its AI

The startup Oumi recently conducted a study for the New York Times, analyzing the performance of Google's artificial intelligence. By examining 4,326 searches performed on Google, the study revealed that Google's AI previews provided correct answers 85% of the time with the Gemini 2 model, and this figure rose to 91% with the transition to Gemini 3.

However, even with an error rate of only 9%, this translates to millions of incorrect answers every hour, given the enormous volume of searches processed by Google. Google responded to the study by highlighting that it presents "serious gaps."

Increased Accuracy but Declining Verifiability

While the study shows an improvement in the accuracy of responses with the shift to Gemini 3, it also highlights a decline in verifiability. Indeed, 56% of the correct answers provided by Gemini 3 could not be verified by the cited sources, compared to 37% with Gemini 2.

Google accompanies each AI-generated response with a warning stating that "AI responses may contain errors," but the exact frequency of these errors remains largely unexplored.

Study Methodology: A Contested Benchmark

To conduct this study, Oumi used the SimpleQA benchmark, recognized in the industry. The tests were conducted in two distinct phases: the first in October with Gemini 2, followed by a second in February after the update to Gemini 3.

The results showed a notable improvement in accuracy, but Google criticized the methodology, claiming that SimpleQA contains incorrect information and does not reflect true user queries. SimpleQA, developed by OpenAI, focuses on particularly difficult questions, those for which at least one AI model failed during a pre-selection process, and is designed for scenarios without internet access.

The Issue of Source Verifiability

Another concerning aspect of the study is the quality of the sources cited by Google. Among the 5,380 sources used, platforms like Facebook and Reddit were among the most common. Facebook appeared in 5% of the correct answers and 7% of the incorrect answers, raising questions about the reliability of these sources. Google may have an interest in favoring sources less likely to pursue legal action for content use.

The New York Times noted several examples of errors, even when Google identified the correct source. For instance, for a question about the Classical Music Hall of Fame, although Google identified the correct site, it incorrectly stated that there was no record of Yo-Yo Ma's induction.

Reactions and Criticism from Google

To verify answers on a large scale, Oumi used its own AI verification model, HallOumi. However, this method has limitations, as the AI used for verification can also make mistakes. Additionally, AI previews can generate different responses for identical searches, even when the queries are separated by just a few seconds.

Ned Adriance, a spokesperson for Google, criticized the study, calling it flawed and highlighting "serious gaps." He also mentioned that the SimpleQA benchmark is designed for difficult questions, naturally increasing the failure rate.

In the Artificial Analysis Intelligence Index, Google's latest model, Gemini 3.1 Pro, shows a 38 percentage point decrease in the hallucination rate compared to the previous Gemini 3. Google claims that results with web search are more accurate than those based solely on the model's knowledge.

The Impact on the Open Web and Publishers

The debate over Google's AI previews extends beyond the accuracy of responses, touching on the impact on the open web. By providing direct answers, Google reduces traffic to external sites, thereby affecting their business model.

The open web is losing its role as a freely linked information network, increasingly replaced by a centralized AI interface under Google's control. While Google denies that its AI previews harm web traffic, it has yet to share its own data to support this claim. In contrast, OpenAI has been more transparent when launching web features for ChatGPT, although this transparency has diminished over time.