AI Explained

Google's AI Told People to Put Glue on Pizza: The Story (and a 3-Question Quiz)

In May 2024, Google's AI Overviews suggested non-toxic glue in pizza sauce and eating rocks. Where those answers came from, and what Google changed.

Watch the Short

The 60-second version is on YouTube. This article is the full story.

New Shorts every week on tech trivia, AI and gadgets.

Subscribe to TheTechPrana

Key takeaways

  • In May 2024, Google's new AI Overviews told a user to mix about ⅛ cup of non-toxic glue into pizza sauce so the cheese wouldn't slide off.
  • The idea came from a joke Reddit comment written about 11 years earlier – and the AI repeated it almost word for word. A separate answer about eating rocks traced back to a satire website.
  • Google said it limited forum posts and satire in AI Overviews, improved detection of nonsense questions, and showed AI answers for fewer kinds of searches.

In May 2024, Google’s brand-new AI search feature gave one of the most famous wrong answers in AI history: put glue on your pizza.

The Short above turns it into a 3-question quiz. Try it first – then read the full story below.

The quick quiz (answers further down)

  1. How much glue did Google’s AI suggest? A few drops · ⅛ of a cup · A whole bottle
  2. Where did the AI get the idea? A cookbook · A joke Reddit comment · A famous chef
  3. What else did it tell people to eat? Rocks · Soap · Sand

What happened

On 14 May 2024, at its I/O conference, Google rolled out AI Overviews to everyone in the US. Instead of just a list of links, Google Search now showed an AI-written summary at the top of the results.

Within days, screenshots went viral. One user asked how to stop cheese sliding off pizza. The AI’s answer included this tip: mix about ⅛ cup of non-toxic glue into the sauce to give it more “tackiness.”

Answer 1: ⅛ of a cup. (Please don’t.)

Where did the glue idea come from?

The AI didn’t invent it. It found it.

The tip matched a joke comment on Reddit posted about 11 years earlier – and the AI repeated it almost word for word, as if it were genuine cooking advice.

Answer 2: a joke Reddit comment.

This is the key lesson. An AI summary is only as good as what it reads. The model couldn’t tell that a decade-old forum joke wasn’t a real kitchen tip.

And then there were the rocks

Another viral answer told people that eating rocks could be good for you, because rocks contain minerals – suggesting at least one small rock a day.

Google’s head of Search, Liz Reid, later explained that hardly anyone had ever asked “how many rocks should I eat?” – so for that rare question, the AI linked to the main page that had written about it: a satire website. Reports traced it to an article from The Onion.

Answer 3: rocks.

Google calls this kind of gap a “data void”: when very few reliable pages exist for an unusual question, the AI may lean on whatever it can find – even jokes.

Not everything viral was real

Reid also pointed out that some of the screenshots shared at the time were fakes that Google couldn’t reproduce. But the glue-on-pizza and eat-rocks answers were real.

What Google changed

In a blog post on 30 May 2024, Reid said Google had made more than a dozen technical improvements, including:

  • Better detection of nonsense questions that shouldn’t get an AI Overview at all
  • Limiting satire and humour content in AI answers
  • Limiting user-generated content – like social media and forum posts – that could give misleading advice
  • Showing AI Overviews for fewer kinds of searches where they weren’t helpful, including some health topics

How did you score?

  • 3/3: AI chef. You can spot an AI fail from a mile away. 👨‍🍳
  • 2/3: Close – the cheese only slid off a little. 🧀
  • 0–1/3: Don’t worry – Google’s AI didn’t do much better. 🧴

The takeaway for using AI answers

  • AI summaries can sound confident and still be wrong. They’re stitched together from web pages – including jokes, old forum posts and satire.
  • Check the source link. If an AI answer points to a forum joke or a comedy site, that’s your clue.
  • Be extra careful with food, health and money. These are exactly the areas where a confident wrong answer can hurt.

If you’d like to understand what these AI tools actually are, our explainer on AI vs agentic AI vs AGI is a good next read.

Your turn: what was your quiz score – 0, 1, 2 or 3? Tell us in the YouTube comments.

Sources

  1. Fortune: Google's AI search gives terrible advice (23 May 2024)
  2. Engadget: Google is putting more restrictions on AI Overviews after it told people to put glue on pizza (30 May 2024)
  3. TechCrunch: Google admits its AI Overviews need work (31 May 2024)
  4. Forbes Australia: Google restricts AI search tool after 'nonsensical' answers
GoogleAI OverviewsAI mistakesAI hallucinationsQuiz