On May 14, 2024, at its I/O developer conference, Google announced that AI Overviews, AI-generated summaries that sit at the very top of a results page, would “begin rolling out to everyone in the U.S., with more countries coming soon.” Google framed it as a win: “with AI Overviews, people use Search more, and are more satisfied with their results,” and said people had already used the feature “billions of times.” For the company that defined web search, putting a generative answer above the ten blue links was a statement that AI was now the front door to the world’s information.
Within days, the front door was producing nonsense. Screenshots spread of AI Overviews advising users to add “non-toxic glue” to pizza sauce so the cheese would stick, and telling someone asking how many rocks to eat that geologists recommend eating one per day. The failures were funny, viral, and pointed: they showed that an AI summary trained to sound authoritative could confidently surface a joke from a forum or a satirical article as fact, with Google’s brand stamped on top.
On May 30, 2024, Google published a follow-up post in which it acknowledged that “odd, inaccurate or unhelpful AI Overviews certainly did show up.” The post is notable for naming the failures directly: it explained the rock-eating answer as a “data void” where the only content matching the query was satirical, and it conceded that web forums “can lead to less-than-helpful advice, like using glue to get cheese to stick to pizza.” Google said it had made “more than a dozen technical improvements,” including better detection of nonsensical queries, reduced reliance on user-generated content for advice, and tighter restrictions on showing AI Overviews for health and news topics. Because the failure is documented by Google’s own response, this episode keeps a primary-sourced record of both the launch and the correction.
The lesson runs deeper than a few bad answers. The same mechanism that makes a language model fluent, predicting plausible text, makes it unable to tell a recipe from a prank when the underlying web data is thin or poisoned. Placing that mechanism at the top of the most-used search engine on Earth meant every weakness became a public headline instantly, and it showed how thin the margin is between a helpful summary and an authoritative-sounding mistake.