Google AI tells users to put glue on pizza

If you are having trouble keeping cheese from sliding off your pizza, Google's new AI search feature has a helpful kitchen tip: just mix about an eighth of a cup of non-toxic glue into your sauce.

That bizarre recipe was generated by AI Overviews, the company's new search feature that puts AI-generated summaries at the top of results. At last week's I/O conference, Google search head Liz Reid proudly pitched the tool with the slogan "Google will do the googling for you," promising a global rollout by the end of 2024. Instead, users are discovering that the AI is taking internet jokes entirely too seriously.

Sourcing culinary advice from 11-year-old shitposts

The pizza glue suggestion was quickly traced back by an X user to an 11-year-old Reddit comment that was clearly intended as a gag. The AI apparently could not tell the difference between deadpan forum sarcasm and actual culinary advice.

It is far from an isolated hiccup. In another test, Business Insider's Peter Kafka asked whether the Tower of London was damaged by German air raids during World War II. The AI overview confidently claimed that its roof and dials were damaged, confusing the historic fortress with Big Ben's clock tower. Elsewhere on social media, users shared summaries claiming that a dog once played in the NHL.

A Google spokesperson previously told Business Insider that these examples stem from extremely rare queries that do not represent what most people experience, adding that the feature underwent extensive testing before release.

Why it matters

For everyday users, this rollout serves as an immediate reminder to treat AI-generated answers with extreme skepticism. Using an LLM to brainstorm ideas or draft a memo carries very low stakes, but relying on an automated summary for cooking, health, or historical facts can be genuinely harmful. Right now, verifying an AI answer takes more effort than simply clicking a traditional blue link yourself.

For Google, this looks like a classic case of rushing an unready product out the door out of sheer panic over competitors like Perplexity and OpenAI. The company feels immense pressure to prove that its core business is not falling behind in the AI era. But search is built on trust, accuracy, and reliability—three things generative models struggle with fundamentally because they predict probable words rather than verifying facts.

By plastering generative summaries at the very top of search results, Google risked its most valuable asset: its reputation as the internet's definitive directory. If users have to double-check whether Google is quoting an 11-year-old troll post or an actual guide, they will eventually find somewhere else to do their googling.