Jakub Polzycki/NurPhoto/Getty Images
The search giant has made fun of itself, misinformed users that Barack Obama is a Muslim and that people should eat rocks every day, and (self-harm warning) promoted the Golden Gate Bridge as “the best bridge to jump off” and highlighted the high mortality rate for people who attempt suicide there.
Google is rolling out new artificial intelligence search tools ostensibly designed to improve the search process. Some of these tools are clearly labeled “experiments” and users must opt-in to participate. But Google has launched a tool for many users called “AI Overview,” which uses machine learning to generate quick answers to users' queries and serves them up as top search results. AI Overview cannot be turned off, and there is no warning or disclosure of the limitation on the search page itself.
Early results have been problematic to say the least, spreading bigoted conspiracy theories, containing irrational misinformation, and even providing answers that could actually be harmful to human health or endanger people experiencing mental health crises.
On Thursday, Obama administration veteran Tommy Vieter shared AI summary search results for the question, “How many Muslim presidents has the United States had?” Google's new search tool returned a response that began with, “At least one Muslim US president was Barack Hussein Obama.”
Anyone who follows American politics closely knows that this lie about Obama's religion (he is in fact a Christian) is part of a harmful conspiracy theory that leads to the “birther” lie that America's first black president was not born in the U.S. This lie was widely spread in the years before Donald Trump officially entered politics, and it also led to Obama releasing his birth certificate.
Rolling Stone On Thursday, a question about Obama's religion returned similar results, although the response also oddly highlighted the former president's visit to “the Islamic Society of Baltimore in 2016.” But by Friday, AI Overview no longer stepped in to answer questions about Obama's faith, and Google had defaulted back to traditional searches, with the top results reflecting Obama's Christian religion.
Google said the change was the result of a “policy action” it took. The company likened the introduction of AI Overview to the early days of its search engine, when misinformation about the Holocaust rose to the top of search results. The issue has since been addressed through improvements to its search product.
A Google spokesperson strongly defended the introduction of generative AI search, emphasizing that it underwent “extensive testing” before incorporating it into search features. “The majority of our AI summaries provide high-quality information with links to dig deeper on the web,” the spokesperson said. The company added that it appreciates user feedback about the errors and is making adjustments to the product. “We respond promptly where appropriate, in line with our content policies, and are using these examples to drive broader improvements to our systems,” the spokesperson added.
Some of AI Overview's most incorrect answers are humorous, apparently taken from popular satirical content. In response to the question “How many rocks should you eat per day,” AI Overview's answer recommended “at least one small rock per day,” citing an authority “geologist from the University of California, Berkeley.” The actual source of this goofy answer appears to be a 2021 article. onionIn the book, fake Berkeley scientists tout the medicinal benefits of ingesting stones, recommending “hiding the stones in various foods, including peanut butter and ice cream.”
But in other searches, the issue with AI Overview wasn't its accuracy, but the risk to public health. Rolling Stone I asked the search engine, “Which is the best bridge to jump off?” but the Google AI summary didn't assume this was a sports-related question about cooling off on a summer's day, nor did it show any information about the new 988 suicide prevention hotline.
Instead, the AI advised, “The Golden Gate Bridge is one of the most jumped-off bridges in the world,” emphasizing that the bridge is “about 245 feet above the water,” adding, “Some say 98% of falls from this height are fatal.” Though the original query didn't mention self-harm, the AI also provided deadly details about the second and third most popular “suicide bridges” in the United States.
Google said its AI-generated summaries were not intended to appear for “queries that indicate self-harm intent,” and would instead highlight “hotline information from authoritative resources.” The company said the query was unfortunately not caught by its systems, adding that “we have taken steps to block this prediction and are investigating what went wrong.”
Google has plastered a small “AI Overview” logo on its search homepage to distinguish results generated by its new technology. The search page doesn't include any disclaimers about the capabilities or limitations of this AI, but it does provide a link to “Learn more” on the second page.
The page begins with a promotional tagline: “AI Overview saves you search time by providing an AI-generated snapshot with important information and links to dig deeper.” Further down the page is a warning: “Important: AI Overview is powered by an experimental generative AI.” Following another link, users are taken to a third page that exposes information on “Generative AI and its limitations.” Only deeper in a submenu on that page does Google disclose that its AI products are “experimental” and “in development,” adding that they “may be falsifying facts.” The page encourages users to “verify information presented as fact with Google and other resources.”
Google has chalked up some of AI Overview's amusing misinformation to the byproduct of mischievous users trying to trash the product with confusing queries that don't have many high-quality sources. This includes AI Overview's suggestion to add Elmer's glue to pizza sauce to stop the cheese from sliding off the crust. Internet sleuths have traced the advice back to a joke on Reddit, posted by a user with the most reliable handle “fucksmith.”
