Google AI Overviews: The Illusion of Precision

Google AI Overviews: The Illusion of Accuracy in an Ocean of Errors

Nearly two years ago, Google introduced a new feature called AI Overviews. This innovation relies on artificial intelligence to provide summaries of web search results. The engine behind this technology is Gemini, among other systems, which compiles and synthesizes multiple sources of information to offer a concise summary accessible to users.

Today, AI Overviews are available in over 200 countries and cover 40 languages. However, France still does not have access to this feature. Despite this absence, stories are already circulating about notable errors produced by these summaries, such as the infamous anecdote about the glue pizza. These amusing tales conceal a deeper and more serious issue. A study conducted by Oumi, a company specializing in open-source AI, has quantified what many suspected: the error rate of AI Overviews is alarming, especially when considering the volume of searches processed by Google.

One Wrong Answer Every Ten Seconds

Oumi's study reveals that 90% of the summaries generated by AI Overviews are correct and well-sourced. Google has not hesitated to highlight this flattering figure. However, what the company does not emphasize is the significance of the remaining 10%. On a global scale, this represents a colossal number of errors. Projections indicate that Google will process over five trillion searches by 2026. This translates to tens of millions of potentially erroneous responses, or several hundred thousand per minute.

Another major issue lies in the very design of AI Overviews. A question asked twice can generate two different answers, one correct and the other incorrect. The user, faced with these variations, does not always know which answer is reliable. Moreover, these summaries appear at the top of the results page, even before links and sources, capturing the attention of internet users.

Increased Risk of Misinformation

AI Overviews do not prioritize sources based on their reliability. They aggregate content from across the web indiscriminately. For example, Facebook and Reddit rank among the most frequently cited sources, in second and fourth positions, respectively. This means that information from unreliable sources can be placed on equal footing with recognized scientific publications.

This approach opens the door to manipulation by malicious actors. Since AI Overviews choose sources based on criteria of popularity and accessibility rather than reliability, it is possible to deceive the system. By creating a website, publishing false information, and artificially generating traffic, an individual can influence the summaries generated by the AI.

Anyone skilled in SEO techniques and equipped with resources to buy traffic can thus create misleading summaries. These boxes, viewed by billions of users, then become sources of authority, even if the information they contain is false.

Google's Responses

In response to the criticism, Google reacted through its spokesperson, Ned Adriance. In a statement to the New York Times, he asserted that "the vast majority of spam [i.e., false content]" is blocked by Google's systems and that "most of these examples correspond to searches that people wouldn't actually make." In other words, Google downplays the significance of the errors by labeling them as unrepresentative.

This response raises questions about how Google perceives errors. If they exist, they are not statistically significant; if they are serious, they are considered unlikely.

Although Google includes a warning below the AI Overviews stating, "AI can make mistakes, verify the answers," few users actually take this caution into account. The majority of internet users do not read the terms and conditions or legal notices and spend an average of 15 seconds on a results page to quickly obtain the information they seek.

This warning seems to exist primarily for Google to protect itself legally in case of litigation related to false information that has caused harm. Since its inception in 1998, Google has always emphasized the accurate indexing of the web, but these recent errors from AI Overviews call this fundamental mission into question.