Google reins in wild AI Overviews after viral blunders

If you saw screenshots last week of Google suggesting people eat rocks or put glue on pizza, you were not alone.

Following its splashy rollout of AI Overviews to all U.S. users at Google I/O, the tech giant spent the week watching bizarre outputs go viral. On May 30, Google published a detailed postmortem explaining why the search feature stumbled and what it changed to fix the mess.

How satire broke the search engine

Google explained that AI Overviews do not operate like typical generative chatbots. Instead of inventing text out of raw training data, the model pulls facts directly from top-ranked search results.

That architecture worked fine until users hit it with absurd queries. Take the viral query about how many rocks to eat. Because sensible people do not ask that, Google ran into an information gap where almost no high-quality web content existed.

The only page seriously matching the text was a satirical piece republished on a geological software company website. The AI read the page, took the joke literally, and served it up as an answer.

A similar breakdown happened with forum posts. Sarcastic comments on Reddit about gluing cheese to pizza were pulled into summaries as if they were genuine culinary tips.

Google also pushed back against the viral mob, noting that many widely shared screenshots were completely faked. Dangerous examples circulating around smoking while pregnant or leaving dogs in hot cars never actually appeared, according to the company.

Putting guardrails on the machine

To plug the leaks, Google rolled out more than a dozen technical updates aimed at restricting when AI steps in.

The system now has stricter filters for nonsensical queries and limits humor and satire from appearing in summaries. Google also scaled back reliance on user-generated content from discussion boards for queries where misleading advice could do harm.

For sensitive topics like breaking news, Google will largely avoid showing AI Overviews altogether. Health queries received additional constraints to prevent dubious medical advice from surfacing.

Google noted that actual content policy violations occurred on less than one in every 7 million unique queries where AI Overviews appeared. But across billions of daily searches, that tiny fraction was enough to dominate headlines.

Why it matters

This rollout underscores the massive risk of pasting generative AI onto the world's most trusted information utility.

For everyday users, AI Overviews remain unreliable whenever a query veers away from standard, well-documented facts. If you are doing basic research, it might save a click. If you are asking anything obscure or safety-critical, verify the underlying links yourself.

We think Google got caught in an awkward trap. Pressured by rivals like Perplexity and OpenAI, it rushed to ship an AI-first search experience before solving a fundamental flaw: language models cannot reliably detect sarcasm or bad forum humor.

By dialing back triggers and cutting out user forums, Google is making the feature safer, but also less useful. It proves that scaling AI to billions of people is vastly harder than building a clever demo.