A week after their algorithms advised people eat rocks and put glue on pizza, Google admitted Thursday that it needed to make adjustments to its bold new generative AI search feature. The episode highlights the risks of Google’s aggressive push to commercialize generative AI, as well as the treacherous and fundamental limitations of that technology.
Google’s AI Overviews feature relies on Gemini, a large language model like the one behind OpenAI’s ChatGPT, to generate written responses to some search queries by summarizing information found online. The current rise of AI relies on LLMs’ impressive fluency with text, but software can also use that ease to put a convincing gloss on falsehoods or errors. Using technology to summarize the promises of online information can make search results easier to digest, but it is dangerous when online sources are conflicting or when people can use the information to make important decisions.
“Now you can get a fast, agile prototype pretty quickly with an LLM, but getting it to not tell you to eat rocks takes a lot of work,” says Richard Socher, who made key contributions to AI for language. as a researcher and in late 2021, he launched an AI-focused search engine called you.com.
Socher says that arguing with LLMs requires considerable effort because the underlying technology has no real understanding of the world and because the web is littered with unreliable information. “In some cases, it’s better to not just give you one answer or show you multiple different points of view,” he says.
Google search chief Liz Reid said in the company’s report. Thursday night blog post which conducted extensive testing before releasing AI Overviews. But he added that mistakes like the examples of eating rock and pizza with glue, in which Google’s algorithms extracted information from a satirical article and a humorous Reddit comment, respectively, had led to additional changes. They include better detection of “meaningless queries,” Google says, and making the system less reliant on user-generated content.
You.com routinely avoids the types of errors that Google’s AI overviews show, Socher says, because his company developed about a dozen tricks to prevent LLMs from misbehaving when used for searches.
“We’re more accurate because we put a lot of resources into being more accurate,” Socher says. Among other things, You.com uses a custom web index designed to help LLMs avoid incorrect information. It also selects from several different LLMs to answer specific queries and uses a citation mechanism that can explain when sources are contradictory. Still, getting AI search right is tricky. WIRED discovered on Friday that You.com failed to properly respond to a query that is known to trip up other AI systems, stating that “according to available information, there are no African nations whose names begin with the letter ‘K.'” In previous tests, he had passed the query.
Google’s generative AI update to its most widely used and lucrative product is part of a tech industry-wide reboot inspired by OpenAI’s launch of the ChatGPT chatbot in November 2022. A couple of months after ChatGPT’s debut , Microsoft, a key OpenAI partner, used its technology to update its Bing search engine. The updated Bing was beset by AI-generated errors and strange behavior, but the company’s CEO Satya Nadella said the move was designed to challenge Google. saying “I want people to know that we made them dance.”
Some experts believe Google was too quick to update AI. “I’m surprised they released it as is for so many queries (medical, financial). I thought they would be more careful,” he says Barry Schwartz, news editor of Search Engine Land, a publication that tracks the search industry. The company should have better foreseen that some people would intentionally try to alter the general descriptions of AI, he adds. “Google has to be smart about this,” Schwartz says, especially when they default to showing results on their most valuable product.
lightning lily, a search engine optimization consultant, was a beta tester for a year on the prototype that preceded AI Overviews, which Google called Search Generative Experience. She says she wasn’t surprised to see the bugs that appeared last week, given that the previous version tended to go wrong. “I think it’s almost impossible for everything to always turn out right,” Ray says. “That’s the nature of AI.”