Google scales back AI Overviews after viral bad-advice answers

Google published a response to the viral bad-advice answers generated by AI Overviews, the search feature it had rolled out to US users at its I/O conference two weeks earlier. In a blog post titled “About last week,” search head Liz Reid acknowledged the eat-rocks and glue-on-pizza answers, along with other examples such as advice on smoking during pregnancy, while also arguing that some of the screenshots circulating on social media were faked rather than genuine.

Reid described “more than a dozen technical improvements” made in response. These included better detection of queries that should not trigger an AI Overview at all, restricting how much the system could draw on user-generated content such as forum posts and satire, adding stronger safeguards around health-related queries specifically, and limiting the inclusion of joke or satirical content more broadly. Google attributed part of the problem to “data voids” — topics where genuine, high-quality content online is sparse, leaving satirical or joke posts as some of the only material available for the system to draw on and, in these cases, mistakenly present as fact.

Reid also defended the feature on the terms Google had launched it: she said internal data showed AI Overviews increased user satisfaction with search results and prompted people to ask longer, more complex questions than they had before. The company did not withdraw the feature.

The episode set a template that recurred across the industry over the following two years: a generative feature launched at scale, a handful of viral failures disproportionate to the actual error rate, and a public list of narrow technical patches rather than a change of approach. It remained one of the most widely cited examples of a major AI product hallucinating in front of a general audience rather than a technical one.