If you Googled “miserable failure” in 2005, President George W. Bush’s official biography on the White House website would show up as the top result.
This was not a political act on the tech company’s part, but rather, a prank. Early jokesters of the internet may remember the 2000s fun of “Google bombing,” of which the Bush insult is the most famous.
Google bombing occurred when trollers linked a webpage – like Bush’s biography – to specific text on their own sites – like, in Bush’s case, “miserable failure.” With enough instances, the search engine’s algorithm misinterpreted the terms as popularly linked. This created all kinds of entertaining results: Googling “Liar” and “poodle” produced the webpage for UK’s prime minister at the time, Tony Blair, while “Dangerous cult” returned the Church of Scientology’s website as the first result.
Google adjusted its algorithm to prevent further Googlebombs in 2007. However, the ghost of Googlebombs has come back to haunt the tech giant, and this time, it’s through their own technological goof.
After Google rolled out an AI-powered search overview in May, users quickly noticed the false and wacky results that these overviews sometimes provided. If you asked Google about the health benefits of running with scissors, it told you the activity was a good cardio exercise to “improve your pores and give you strength.” Another query led Google to recommend the health benefits of rocks, apparently referencing a satirical article from the Onion.
“Eating the right rocks can be good for you because they contain minerals that are important for your body’s health,” Google’s AI overview replied to a reporter’s query.
‘Glue in pizza’ remains strong
However, nothing made the internet riot as much as a suggestion from the AI Overview to “mix about 1/8 cup of non-toxic glue into the sauce” to keep cheese from sliding off your pizza slice.
A spokesperson from Google brushed off the false results, writing at the time that “the examples we’ve seen are generally very uncommon queries, and aren’t representative of most people’s experiences.” They added that the “vast majority” of AI Overviews provide high quality information, with links to allow the inquiring person to dig deeper into a search.
However, even as Google publicly expressed confidence in its new AI tool, the company quietly began reducing its visibility. Google gradually reduced the amount that AI Overviews popped up in search results from 84% to 15%, according to research from content marketing platform BrightEdge.
A spokesperson from Google disputed the data, noting that the numbers were different than what the company had seen. They added that this was likely because BrightEdge looked at a narrowed query set that is not a representative sample of Google Search traffic, including those who had opted out of AI overviews.
One example where AI overviews sometimes still shows up is the same glue-in-pizza result that had the internet rolling. The Verge recently reported that if you ask Google how much glue to add to your pizza, it will give the same result, only this time citing an article from Business Insider about the fiasco.
Which means that,it seems like the more journalists write about the ridiculous AI Overviews, the more it feeds the algorithm to produce the same, wrong results. It’s a kind of self-fulfilling feedback loop that recalls the absurdity of the Googlebombing days, although the only troll in play is Google itself.
When Fortune reporters attempted various searches for pizza, cheese, and glue, no AI overview would pop up, possibly meaning that Google may have caught wind of their continuous blunders and promptly adjusted their platform.
The Google spokesperson said the queries continue to show for a large number of searches, but the technology is undergoing edits.
“We’re continuing to refine when and how we show AI Overviews so they’re as useful as possible, including technical updates to improve response quality,” they told Fortune.