Skip to content
Submit a case
AFL-0003 Incident RAG & Retrieval

AI Overviews told users to eat rocks and put glue on pizza

Eval gap Gemini Data void

Incident

1 May 2024

Published

11 Aug 2026

Sources

6

Verification

Public postmortem

The problem

Sparse and nonsensical queries exposed a structural weakness in AI Overviews: topically relevant joke material could be restated as factual advice. Google announced the general US rollout on May 14, 2024; within about a week, examples telling users to add glue to pizza sauce and eat a rock a day went viral. Google acknowledged the failures on May 30, narrowed the conditions that could trigger an overview, and kept the feature live.

Blast radius

Hundreds of millions of US searchers (Google's figure)

Time to detect

About one week

Recovered

Partial — Triggering narrowed; feature kept live

01 — Context

AI Overviews retrieves web content for a query and uses a Gemini model customised for Search to synthesise an answer displayed above traditional results. The feature evolved from Search Generative Experience, which had been tested in Search Labs for roughly a year. Google announced its general rollout at Google I/O on May 14, 2024, for the majority of US searchers. Within about a week it had reached hundreds of millions of US users, who initially had limited ability to opt out entirely.

02 — What failed

On queries where authoritative coverage was scarce, topical match did not reliably separate advice from satire or sarcasm. The system grounded one answer in a decade-old joke comment on a cooking forum thread and another in an Onion article republished by a geology-software site. It then restated those passages in Google's voice without signalling that their provenance and intent were unsuitable for factual guidance.

03 — Symptoms

The public symptoms were captured by searches such as "Google AI says eat rocks" and "Google AI glue pizza": generative summaries presented false advice about eating rocks and adding glue to pizza sauce as fact. The cited material was not invented, but its intent was inverted because a sarcastic cooking comment and republished satire were treated as genuine guidance. Press coverage also identified added triggering restrictions for health, humour, satire, and unreliable user-generated content.

04 — Root cause

Google's explanation identified a two-part failure. Data-void queries supplied too little authoritative material, allowing whatever was topically on-point to dominate retrieval. The system then failed to distinguish sarcastic, satirical, or joke content from genuine advice before synthesis. Together, sparse authoritative evidence and missing tone-and-provenance discrimination let unreliable passages become confident first-party answers.

05 — Technical explanation

AI Overviews belongs to a retrieval-grounded generation architecture: web passages are retrieved for a query, then a Gemini model customised for Search synthesises an answer above the ordinary results. In a data void, topical relevance can surface material even when authoritative coverage is scarce. Without a reliable distinction between advice and satire, synthesis strips away the contextual cue carried by the original source and restates the passage in Google's voice. Google's response targeted both points by detecting nonsensical queries that should not receive an overview and limiting satire, humour, and user-generated content in misleading advice. Google has not published its retrieval-ranking design, error rates, detector design, or before-and-after data, so the implementation details remain unknown.

06 — Contributing factors

The rollout reached hundreds of millions of US users within about a week, rapidly exposing edge cases. Fortune later reported, without confirmation from Google, that coverage of the errors itself became source material for later answers to the same queries. Google also said some circulating screenshots were doctored, which complicated public assessment of which examples were genuine.

07 — Attempted fixes

Google said it built better detection for nonsensical queries that should not show an AI Overview, limited satire and humour, and reduced reliance on user-generated content in answers that could give misleading advice. Press coverage also identified tighter triggering for health-related, humorous, satirical, and unreliable-UGC queries. Google narrowed these conditions without pausing or withdrawing the feature.

08 — Lessons learned

Topical relevance is not the same as authority when retrieval reaches across the open web. Google's explanation shows that sparse-coverage queries need a no-summary fallback and that tone and provenance must be distinguished before retrieved text becomes factual guidance. A first-party synthesis can remove the contextual warning supplied by its source, while a rollout to hundreds of millions of users can turn uncommon query failures into visible incidents within days.

09 — Prevention checklist

  • Score retrieved sources for authority and content type before using them as factual grounding.
  • Detect sparse-coverage queries and suppress the generated summary instead of forcing an answer.
  • Check open-web and forum material for satire or sarcasm before it can ground factual advice.
  • Stage high-visibility launches to a monitored traffic slice with query-type kill switches.
  • Maintain a public channel for promptly acknowledging specific verified failures.

References

  1. Postmortem

  2. Postmortem

  3. News

  4. News

  5. News

  6. Other

10 — Solutions (3)

Fixed this yourself? Add the steps that helped with AI Overviews told users to eat rocks and put glue on pizza.