Why did Google's AI say you should eat rocks?

Updated 2026-07-151,900 searches/moRanked #201 of 519· AI explained
Short answer

In May 2024, Google's AI Overviews told users to eat at least one small rock per day. It pulled the claim from a satirical article in The Onion headlined "Geologists Recommend Eating At Least One Small Rock Per Day," which had been republished on a geology software company's site. Google's Head of Search, Liz Reid, confirmed this. Do not eat rocks.

Why — the first-principles explanation

The rock answer was not a hallucination in the usual sense. The AI did not invent anything — it faithfully summarized a real web page. The failure was that it could not tell a joke from a fact.

Here is the mechanism. AI Overviews works by retrieving web pages that match your query, then having a language model summarize what they say. This design is meant to reduce made-up answers by grounding them in real sources. But grounding only helps if the sources are good, and the system has a specific blind spot: when almost nothing on the internet seriously addresses a question, whatever does address it dominates. Google's own explanation named this a "data void." Nearly nobody writes earnest articles about the correct daily allowance of rocks. The Onion did write a joke about it. So the joke was effectively the entire corpus.

The second half of the failure is laundering through republication. The Onion is recognizable as satire on its own site. But the piece had been reposted on a geological software provider's website, stripped of the context that flags it as comedy. To a retrieval system scoring relevance, a rock-related claim on a geology company's domain looks like a topically authoritative match. The satire signal was lost in transit; the topical signal survived.

The deeper lesson is about what these systems actually optimize. A retrieval model ranks pages by relevance and apparent authority, not by sincerity. Sincerity is not a feature you can measure from text statistics — sarcasm, parody, and deadpan reporting look identical to earnest writing at the surface level. Humans catch satire using context the model does not have: knowing The Onion's reputation, sensing that the premise is absurd, noticing the joke's rhythm. Google's fix, per Reid, was to add explicit detection and handling of satirical content — a patch on a specific failure mode, not a general solution to the machine's inability to get a joke.

An example that makes it click

Imagine you ask a very fast, very literal research assistant: "How many rocks should I eat per day?" He sprints to the library. There is no medical textbook on rock consumption, because the question is ridiculous. But he finds exactly one page mentioning it — a humor column, photocopied and stapled into a geology company's binder with the funny letterhead cut off. He returns in four seconds and reports, confidently: "Geologists recommend one small rock per day."

He did not lie. He read the only document that existed and summarized it correctly. He just cannot laugh. Every joke he reads is a fact he files. The problem was never that the library was empty — it was that the only book on the shelf was a comedy book, and nobody told him.

Key facts

Infographic: Why did Google's AI say you should eat rocks — short answer and key facts
Visual summary — Why did Google's AI say you should eat rocks?
▶ The 60-second explainer (script)

In May 2024, Google's AI told people to eat at least one small rock per day. Here's why — and it's more interesting than just "the AI made it up." Because it didn't make it up. The AI was quoting The Onion. The satirical newspaper had run a piece headlined "Geologists Recommend Eating At Least One Small Rock Per Day." That article got republished on a geology software company's website, stripped of the context that marks it as a joke. Google's Head of Search, Liz Reid, explained exactly this. Two things went wrong. First, a data void. Almost nobody on the entire internet writes seriously about how many rocks to eat, because the question is absurd. So when the AI went looking, the joke wasn't one source among many — it was basically the only source. It won by default. Second, satire laundering. On The Onion's own site, you'd know it's comedy. Reposted on a geology domain, it just looks like a rock article on a rock website. The AI ranks pages by relevance and authority. It cannot measure sincerity. Sarcasm and sincerity look identical in raw text. That same week, AI Overviews also suggested glue to keep pizza cheese from sliding — that one came from a joke Reddit comment. Google's fix was to teach it to spot satire. To be clear: eat zero rocks per day.

What authoritative sources say

The Register — Google finally addresses those bizarre AI search resultsmedia — Google's Head of Search Liz Reid explained that the rock-eating answer originated from The Onion's satirical article republished on a geological software provider's site, in a query area with little serious web content. source ↗
The Daily Beast — Google Explains Why Its AI Overviews Told Users to Eat Rocks and Glue Pizzasmedia — Google publicly explained why AI Overviews told users to eat rocks and to use glue on pizza, and said it would better identify and handle satirical content. source ↗
MedlinePlus (U.S. National Library of Medicine) — Foreign object - swallowedgov — Swallowed foreign objects can lodge anywhere from the esophagus to the colon and cause a blockage or tear in the GI tract. source ↗
NIH/PMC — Management of foreign body ingestion in adultsedu — About 80% of ingested foreign bodies pass asymptomatically, roughly 20% require endoscopic removal, and perforation occurs in under 1% but can lead to peritonitis and sepsis. source ↗

People also ask

Did Google's AI invent the rock advice?

No. It accurately summarized a real article — a satirical piece from The Onion that had been republished elsewhere without the context signaling it was a joke.

What is a "data void"?

A query where almost no serious content exists on the web. Whatever content does exist — even a joke — then dominates the AI's summary because there is nothing to outweigh it.

Has Google fixed this?

Google added specific detection for satirical and joke content, and limited AI Overviews on some nonsensical queries. The underlying limitation — models cannot reliably detect sincerity from text alone — has not been solved.

Should you actually eat rocks?

No. Zero rocks per day. Swallowing rocks risks cracked teeth, choking, and intestinal blockage or perforation.

What was the glue-on-pizza answer?

The same May 2024 wave. AI Overviews suggested adding non-toxic glue to help cheese stick to pizza, tracing back to a joking Reddit comment that the system read as genuine advice.

Related questions