Google's AI Overviews tells users to eat rocks and put glue on pizza

Screenshots began circulating on social media showing that AI Overviews, the generated-answer feature Google had put above search results at its I/O conference a little over a week earlier, was presenting satire and old forum posts as settled fact. One widely shared answer told users to eat at least one small rock a day for vitamins and minerals — traced back to an article published by the satirical site The Onion. Another recommended adding non-toxic glue to pizza sauce to help cheese stick, sourced from an eleven-year-old joke comment on Reddit.

Other examples spread alongside the two viral cases, including an AI Overview reportedly telling a user with kidney stones to drink litres of urine, and one repeating the false claim that Barack Obama is Muslim. Not every screenshot in circulation was genuine — some were later shown to be fabricated — but enough were verified by reporters that Google did not dispute the underlying pattern: the system was surfacing a single scraped source, including jokes and satire, and displaying it as a direct answer rather than a search result to be evaluated.

The episode became one of the most widely shared examples of large-language-model hallucination reaching a mass audience rather than a technical one, arriving as Google, Microsoft and AI-native competitors such as Perplexity were racing to put generated answers in front of searchers by default. It also renewed a specific concern about “data voids” — queries with little genuine content online, where a single satirical or joke source can dominate what a model retrieves. Google restricted the feature and published a list of fixes within the week.