Google scaled back AI Overviews after it told users to put glue on pizza and eat rocks

The problem

AI Overviews placed an AI-generated answer, produced by Google’s Gemini model, at the very top of the search results page, above the ranked links it summarised. When the system drew on satire or low-quality user posts and presented the result as a direct answer, it could surface a confident, plainly false statement in the position users trust most.

What was built

The feature under scrutiny was Google AI Overviews, rolled out to all users in the United States in May 2024 source. No numeric outcome (failure rate, query volume) was ever published, so this page documents the erroneous outputs, Google’s public acknowledgment, and the restrictions it shipped, not a measured error rate.

The outcome

Within days the feature produced answers that were widely reported and mocked. The BBC reported that AI Overviews told some users that to keep cheese from sliding off pizza they could use “non-toxic glue,” and that its responses said “geologists recommend humans eat one rock per day” source. Google’s own follow-up post acknowledged advice “like using glue to get cheese to stick to pizza” source, and the Associated Press quoted the “how many rocks should I eat?” example directly source. As reported, the errors were traced to the system surfacing satire and user content, the BBC noted answers based on “articles written by satirical site, The Onion” source, and AP said the pizza-glue answer was “pulled from a satirical reddit comment” source. A Google spokesperson told the BBC the cases were “isolated examples” source.

On 30 May 2024, Google’s head of Search, Liz Reid, published a post acknowledging the failures and describing a set of scale-backs. Google said it “made more than a dozen technical improvements to our systems”; it “added triggering restrictions for queries where AI Overviews were not proving to be as helpful”; and it “updated our systems to limit the use of user-generated content in responses that could offer misleading advice” source. The Associated Press independently reported the “more than a dozen technical improvements” figure source, and ABC News reported that Google now aims “to not show AI Overviews for hard news topics, where freshness and factuality are important” source.

A note on the weakest load-bearing source. The scope and count of the fixes, “more than a dozen technical improvements” and the specific restrictions, originate with Google’s own blog post (Tier 1, but a first-party statement against interest, not an independent audit). The Associated Press and ABC News (both Tier 2) re-report the same figures firsthand rather than measuring them; no independent party quantified the failure. The load-bearing facts are therefore the public ones: the viral erroneous outputs, quoted by the BBC and AP, and Google’s own on-record acknowledgment that it restricted the feature.

How this was verified

Method: Google’s own blog post (head of Search Liz Reid) acknowledges the failures and the restrictions it shipped, a first-party statement against interest. Two mutually independent newsrooms carry the same changes: the Associated Press and ABC News (ABC by its own byline, Michael Dobuski, not AP syndication). The erroneous outputs are carried firsthand by the BBC and AP, and admitted in Google’s own post. All 17 quotes were checked verbatim (tools/quotecheck.py, 17/17 OK) against local Wayback captures that byte-tie to their cited CDX digests. Verified 2026-08-07. Scope is honest: Google restricted / scaled back the feature; it did not fully disable it.

Sources

1. Google · “What happened with AI Overviews and next steps” (Liz Reid) · 2024-05-30 · Tier 1 (first-party, statement against interest) · https://blog.google/products/search/ai-overviews-update-may-2024/
2. BBC · “Google AI search tells users to glue pizza and eat rocks” · 2024-05-24 · Tier 2 (independent press) · https://www.bbc.com/news/articles/cd11gzejgz4o
3. Associated Press · “Google makes fixes to AI-generated search summaries after outlandish answers went viral” · 2024-05-31 · Tier 2 (independent press) · https://apnews.com/article/google-ai-overviews-hallucination-33060569d6cc01abe6c63d21665330d8
4. ABC News · “Google makes adjustments to AI Overviews after a rocky rollout” (Michael Dobuski) · 2024-06-03 · Tier 2 (independent press) · https://abcnews.go.com/Technology/google-makes-adjustments-ai-overviews-after-rocky-rollout/story?id=110710227

Related case files

• Microsoft capped Bing AI chat at 5 turns after unhinged responses, the closest sibling: another AI answer product restricted, not withdrawn, days after launch when its outputs went off the rails.
• Microsoft took its Tay chatbot offline after 16 hours, the same failure mode a decade earlier: a public AI system surfacing the worst of user-generated content and pulled in response.
• Character.AI removed open-ended companion chat for under-18 users, another consumer AI product restricted, not withdrawn, after its outputs drew public and safety pressure.

Path to green

This is an AI-failure / accountability story, not a client-outcome case study, so a green badge would require a human to confirm the incident record on the record with Google, whose blog post is already public and against interest. No numeric outcome is claimed; the verifiable facts are the erroneous outputs, Google’s public acknowledgment, and the restrictions Google shipped. Until a human confirms, the story stays in the checker’s hands.