Google has finally addressed the issues surrounding its AI Overviews feature.
For those who may not be up to speed, AI Overviews made their debut on Google’s search engine on May 14, transitioning from the beta Search Generative Experience to a live feature accessible to all users in the U.S. Initially intended to provide AI-generated responses at the top of nearly every search query, the feature quickly garnered attention for suggesting peculiar advice, such as adding glue to pizzas or offering potentially harmful health recommendations. While still active, AI Overviews appear to have diminished in prominence on the platform, with fewer searches yielding responses from Google’s AI algorithms, according to the Lifehacker team’s observations.
Google Search VP Liz Reid shed light on the situation, explaining that although the feature underwent extensive testing, “there’s nothing quite like having millions of people using the feature with many novel searches.” While acknowledging the less-than-stellar reputation AI Overviews has acquired, the company also acknowledged the breakdowns and is actively working to rectify them.
Reid clarified that AI Overviews operate differently from chatbots and other language model products, emphasizing that they do not merely generate responses based on training data. Instead, they perform traditional search tasks, pulling information from top web results. Consequently, errors are attributed to the model misinterpreting existing web content rather than experiencing hallucinations.
“We observed AI Overviews that featured sarcastic or troll-like content from discussion forums,” Reid elaborated. While forums can provide valuable firsthand information, they can also be sources of less-than-helpful advice. The challenge lies in the robot’s inability to distinguish sarcasm from genuine assistance, leading to occasional misinterpretations.
Furthermore, Reid highlighted the issue of “data voids” on certain topics, where insufficient serious content exists. In such cases, AI Overviews inadvertently sourced from satirical sources instead of reputable ones. To address these challenges, Google claims to have implemented improvements to AI Overviews.