Google Launches AI Overviews in Search
Summary
Google began rolling out AI Overviews — AI-generated summary answers at the top of search results — to all US users, representing the most significant change to Google Search in its history. The launch was marred by widely publicized errors, including telling users to eat rocks and put glue on pizza.
What Happened
At Google I/O on May 14, 2024, Google announced the broad rollout of AI Overviews (formerly Search Generative Experience/SGE) to all US search users. AI Overviews used Gemini to generate summary answers displayed prominently above traditional search results, fundamentally changing the search experience for billions of users.
The rollout quickly encountered problems. Within days, screenshots of bizarre and dangerous AI-generated answers went viral: the system suggested adding glue to pizza sauce to help cheese stick (based on a satirical Reddit post), recommended eating rocks for their mineral content, and provided other confidently incorrect information. The errors highlighted the challenge of deploying generative AI at search scale, where even a small error rate translates to millions of problematic responses.
Google responded by reducing the scope of AI Overviews, limiting them for queries where they were less reliable, and implementing additional safeguards. The company emphasized that the viral errors represented edge cases that had been quickly addressed.
Why It Matters
AI Overviews represented the most consequential commercial deployment of generative AI by any company, affecting billions of daily searches. The feature's errors — while individually trivial — collectively demonstrated the fundamental tension between generative AI's tendency to produce plausible-sounding but incorrect information and search's requirement for reliable, accurate answers.
For publishers and content creators, AI Overviews represented an existential threat. By providing AI-generated answers directly in search results, Google reduced the incentive for users to click through to the underlying websites that provided the source information. This "zero-click" dynamic threatened the web's economic model, where publishers created content in exchange for search traffic. It also fueled the copyright debate, as publishers saw Google using their content to train AI that then reduced their traffic.
The episode also demonstrated that hallucination — a fundamental limitation of current generative AI — was not merely a technical inconvenience but a barrier to deployment in high-stakes, high-volume applications where accuracy is expected by default.