Skip to main content

Google defends AI search results after they told us to put glue on pizza

Google defends AI search results after they told us to put glue on pizza

/

Google blames ‘data voids’ and edge cases for its bad AI-generated search results.

Share this story

Illustration of a robot brain.
Image: The Verge

Last week, Google rolled out its AI search results for millions of users to tinker with. The goal was to deliver a better search experience. Instead, AI delivered all sorts of weird results, like saying people should put glue on their pizza to help the cheese stick and eat rocks.

Google worked quickly to remove some inaccurate AI results, which it calls AI Overviews, but the damage — and meme-ification — was already done.

Now, in a company blog post published Thursday by Google’s head of search, Liz Reid, the tech giant blames “data voids” for the inaccurate results, along with people making up odd questions, and doubled down by claiming that AI results are leading to “higher satisfaction” with Search. Reid argues that AI Overviews generally don’t “hallucinate;” they just sometimes misinterpret what’s already on the web.

“There’s nothing quite like having millions of people using the feature with many novel searches,” she writes. “We’ve also seen nonsensical new searches, seemingly aimed at producing erroneous results.” She also correctly notes that a “very large number of faked screenshots” of AI Overviews have been spreading online.

First, I’d like to point out that “what mammal has the most bones?” is a fair query if you spend any amount of time with an inquisitive toddler. Second, Google opted millions of people into this feature, which resulted in a lot of backlash and even spurred articles explaining how to improvise a way to disable it.

Reid’s blog also explains how Google is fixing AI Overviews by limiting when they appear for “nonsensical” queries and satire. This is good to address because a lot of people were surprised that the well-known satirical website The Onion and a Reddit user named “fucksmith” were not filtered out of AI results in the first place.

Part of Reid’s blog also compares AI Overviews to another longstanding Search feature called featured snippets, which highlight information from a relevant webpage without using generative AI. According to Reid, the “accuracy rate” for featured snippets is “on par” with AI Overviews.

Ultimately, Google’s introduction of AI Overviews has been yet another PR blunder for the company. Meanwhile, Google is racing to compete against OpenAI and AI search startups like Perplexity, which is already worth a rumored $3 billion.

If Google is going to compete, it has to move fast. But it also needs to maintain user trust. That could be difficult to regain after AI Overviews told us all to eat Elmer’s glue.