Google Caught Manually Taking Down Bizarre AI Answers

Control-Alt-Delete

Google manually removed AI-powered search results from pages where it was caught giving bizarre and incorrect information to search users, The Verge reports.

Last week, on the heels of Google's all-in-on-AI I/O conference, netizens and journalists alike took to social media in droves to share screenshots of Google's new "AI Overview" — in short, a new AI function tacked to the top of Google search pages that gobbles up search results and paraphrases what it finds — spitting out head-scratching, false, and in some cases downright conspiratorial answers to search queries.

Case in point, the AI feature was caught telling users that they should try putting glue on pizza, a recommendation the AI sourced from an 11-year-old Reddit comment posted by a user who goes by "Fucksmith." The AI was also caught telling people they should eat rocks, and even rehashing false "Birther" conspiracy claims that former President Barack Obama is secretly Muslim.

But now, these particular mistakes, among many others, are gone. Per The Verge, in what can only be described as a Whac-A-Mole-inspired strategy, Google quickly moved to manually remove AI Overview from these specific, problem-stirring queries.

In short, it's a short-sighted, band-aid fix that arguably fails to address far more persistent flaws of the tech — and Google has its work cut out to ensure the reputation of its flagship product doesn't take on any more damage.

Keep Scrolling

To be clear, it's a good thing that Google is taking AI Overview off of search pages where it's known to give bad information. But in each of the instances referenced above, the feature came to incorrect conclusions by either hallucinating false information or sourcing its answer from woefully unreliable sources.

These mistakes speak to reliability and accuracy problems that plague generative AI products across the tech industry. Manual take-downs don't address this core issue, and to that end, the method isn't exactly scalable either as Google runs billions of searches every day.

For its part, Google told The Verge that it's "taking swift action" to remove AI Overview "where appropriate under our content policies, and using these examples to develop broader improvements to our systems, some of which have already started to roll out."

But whether Google will indeed be able to meaningfully address these core problems remains to be seen. Google CEO Sundar Pichai, however, is optimistic about the future of AI-enhanced search.

"Are we making progress? Yes, we are," he told The Verge last week. "We have definitely made progress when we look at metrics on factuality year on year. We are all making it better, but it's not solved."

More on Google and AI: CEO of Google Says It Has No Solution for Its AI Providing Wildly Incorrect Information