# HD-INC-029: Google's AI Overviews Told People to Put Glue on Pizza and Eat a Small Rock Every Day

> Days after Google put an AI-written answer above its search results for US users, the feature was advising glue on pizza and a daily small rock. The bad answers were faithful summaries of a joke and a satire.

- Source: https://useheadlights.com/library/HD-INC-029/
- Publisher: Headlights (useheadlights.com), author Ellie Harris
- Published: 2026-06-11
- Company: Google
- Sector: Consumer AI
- Country: United States
- Year: 2024
- Behaviour in the record (Headlights reading): Overconfidence
- Governance domains: Consumer protection, Algorithmic accountability
- Licence: CC BY-NC 4.0. Cite as: Harris, E. (2026). Google's AI Overviews Told People to Put Glue on Pizza and Eat a Small Rock Every Day. Headlights AI Incident Library, HD-INC-029. https://useheadlights.com/library/HD-INC-029/

## What happened

It was reported that in May 2024, Google launched AI Overviews across the United States, placing AI-generated answers above its traditional search results. Instead of giving people a list of websites to explore, Google began presenting information gathered from across the internet as a single, confident answer.

Within days, some of those answers were making headlines for all the wrong reasons, including advice to add glue to pizza sauce to stop cheese sliding off and a suggestion that people should eat a small rock every day. Other bizarre screenshots circulated online, although Google later explained that some were fabricated and not every example being shared was genuine.

The pizza advice came from an old joke posted on Reddit more than a decade earlier, while the recommendation to eat rocks originated from satirical content that had been republished on another website. Neither was intended as serious advice, but Google's AI presented both as factual information without recognising the humour behind them.

The problem was not simply that the AI made mistakes, but that it turned unreliable material into answers that appeared to carry Google's authority. People searching for information were no longer just being shown questionable sources; they were being given questionable information as though it had already been checked.

Google acknowledged the errors and introduced more than a dozen technical improvements, including restrictions on using satire and certain user-generated content in AI Overviews. The company also reduced when the feature appeared, particularly for searches where reliable information was limited.

## What an auditable version would have shown

The examples became public because people shared screenshots, but understanding why the answers appeared required tracing them back through the websites and material the system had relied upon.

An auditable system would have recorded which sources contributed to each answer, whether those sources were credible and what checks were performed before the information appeared at the top of Google's search results.

That would have made it easier to identify patterns, including how often AI-generated answers relied on satire, jokes or unverified material rather than independent sources.

## Where the gap was

Headlights identifies **overconfidence and a rules-behaviour gap**, where a system presents information with greater certainty than the evidence supporting it deserves.

There is a familiar human behaviour behind this failure because people often associate a confident explanation with knowledge, particularly when it comes from a source they already trust.

Google had spent decades becoming a place people turned to for answers, and its AI Overviews inherited that credibility simply by appearing at the top of the search page. When the system repeated a joke as factual advice, the authority of the platform made the mistake appear more believable than the original source ever would have been.

The technology was capable of finding and summarising information, but it did not consistently distinguish between something written to inform people and something written to make them laugh.

## What governance should have looked like

An AI system presenting information as fact needs to establish whether the material supporting its answer is reliable, particularly when the subject involves health, safety or other decisions that could affect someone's wellbeing.

A VerificationGate could require important claims to be checked against credible sources, while a ConstraintGate could restrict the use of satire or unverified material when generating factual advice. A ConductRecord could preserve evidence of those checks, making it possible to understand why a particular answer was produced and whether the safeguards worked.

Where reliable information cannot be established, showing people the original search results is a better outcome than generating an answer that sounds certain but may be wrong.

The lesson is that making information easier to access also increases the responsibility to get it right, because people are less likely to question an answer when a trusted platform has already appeared to do the thinking for them.

**Failure Pattern:** Google AI Overviews presented jokes and satirical material as factual advice, giving unreliable information the appearance of authority through its position at the top of search results.

**Governance Principle:** AI-generated answers should be supported by credible evidence, with safeguards that distinguish reliable information from satire, jokes and unsupported claims before presenting them as fact.

The reference implementation of VerificationGate, ConstraintGate and ConductRecord is open source. It lives at [github.com/saffronandindia/headlights-oss](https://github.com/saffronandindia/headlights-oss), Apache 2.0 licensed and free to install. The repository is public now.

## Sources

- [AI Overviews: About last week (Elizabeth Reid, VP of Search, Google blog)](https://blog.google/products/search/ai-overviews-update-may-2024/)
- [Google makes fixes to AI-generated search summaries after outlandish answers went viral (PBS NewsHour)](https://www.pbs.org/newshour/politics/google-makes-fixes-to-ai-generated-search-summaries-after-outlandish-answers-went-viral)
- [Pizza Glue? Small Rocks? Google AI Overview Answers Are a Mess (Bloomberg Opinion)](https://www.bloomberg.com/opinion/articles/2024-05-30/pizza-glue-small-rocks-google-ai-overview-answers-are-a-mess)

## Related

- [HD-INC-001](/library/hd-inc