After odd answers to searches spread in May 2024, Google said some screenshots were faked, but real errors came from rare questions with little good content, satire and sarcastic forum posts used as sources, and occasional misreading of page language. It announced limits on those inputs and on when overviews appear.
What Google published
The explanation came in a 30 May 2024 post by Elizabeth Reid, Google's VP of Search. Key points from it:
- Many screenshots shared on social media were faked. Google said examples about leaving dogs in cars never appeared in Search.
- Some real odd, inaccurate or unhelpful overviews did appear, mostly for uncommon queries. Google also saw nonsensical new searches that seemed designed to produce errors.
The causes Google named
- Data voids. For some questions little high quality content exists. Google called this an information gap.
- Satire. An overview answered "How many rocks should I eat?" by citing satirical content republished on a software provider's website.
- Forum content. Sarcastic or trolling posts from discussion forums were sometimes used, which led to advice such as adding glue to cheese on pizza.
- Language misreads. In a small number of cases the system misinterpreted language on webpages.
What Google said it changed
- Better detection of nonsensical queries that should not trigger an overview, and limits on satire and humor content.
- Limits on user generated content in responses that could give misleading advice.
- Trigger restrictions for query types where overviews were not proving helpful.
- Extra refinements for health topics. Google said it already avoids overviews for hard news.
- Ongoing monitoring, with removal of overviews that violate policy. Google reported violations on fewer than one in 7 million unique queries that triggered an overview.
- More than a dozen technical improvements meant to help broad sets of queries instead of patching single ones.
Reading it carefully
Two cautions. First, the faked screenshots are part of the story. The reliable way to discuss this incident is to cite examples Google itself confirmed, such as the rocks and glue answers, not whatever went viral. Second, the figure of fewer than one in 7 million is Google's own and covers policy violations, not overall accuracy. It does not say how often answers were wrong in other ways.
What it teaches
These are failures of the inputs, not only of the model. A system that summarizes retrieved text inherits the quality of that text. When no good source exists, it may answer anyway. See data voids and source quality for retrieval systems.
Quick check
Were all the viral AI Overview screenshots real?
No. Google said many were faked, though it confirmed some real odd and inaccurate answers.
Did Google say the model alone was at fault?
It named causes tied to sources and queries: data voids, satire, forum posts and language misreads.
Sources
Last reviewed: 19 October 2026.
