What Happened
In May 2024, shortly after Google rolled out AI Overviews (formerly Search Generative Experience) to all US users, the feature produced a series of spectacularly wrong answers that went viral. The most infamous examples included suggesting that users should eat rocks because they are a good source of vitamins and minerals, recommending adding non-toxic glue to pizza sauce to prevent cheese from sliding off (drawing from a joke Reddit post), and providing other absurd health and safety advice.
The AI Overviews feature placed AI-generated summaries at the top of Google Search results, giving them the highest-visibility position in the world’s most-used information retrieval system.
Timeline
Google AI Overviews was enabled for all US users on May 14, 2024. Within days, users began sharing screenshots of bizarre outputs. By May 23, the examples had gone viral. Google responded by pulling AI Overviews from many query types by late May and implementing additional content filters. The company reduced the frequency of AI Overviews appearing in search results and added restrictions for health, safety, and other sensitive topics.
Impact
The incident was a significant embarrassment for Google’s AI strategy. The company had staked its competitive response to ChatGPT on AI integration into Search, and the viral failures undermined confidence in that approach. The placement of AI-generated answers at the top of search results meant that the hallucinations had maximum visibility and potential to mislead users.
The “eat rocks” and “glue pizza” examples became shorthand for AI unreliability and were cited in media coverage, congressional discussions, and competitor marketing for months afterward.
Response
Google stated that the viral examples represented edge cases and that AI Overviews were accurate for the vast majority of queries. The company implemented technical restrictions including reducing AI Overviews for health queries, adding detection for queries likely to produce nonsensical results, and limiting the feature’s reliance on user-generated content like Reddit posts.
Google VP Liz Reid published a blog post acknowledging the errors while emphasizing the overall utility of the feature.
Lessons Learned
The AI Overviews incident demonstrated the unique danger of AI hallucinations in high-trust information environments. When an AI generates bad advice in a chatbot, users may treat it skeptically. When the same bad advice appears at the top of Google Search, it carries the implicit authority of Google’s brand and its position as the world’s primary information gateway.
The episode also showed that AI systems trained on internet content will inevitably encounter and sometimes amplify satirical, joke, or deliberately misleading content. The glue-on-pizza recommendation came from a clearly humorous Reddit comment, but the AI could not distinguish satire from genuine advice.