Google struggles to manually remove strange AI answers in search

Social media is full of examples of Google’s new AI Overview product saying strange things, from telling users to put glue on their pizza to suggesting they eat rocks. The confusing rollout means Google is rushing to manually disable AI overviews for specific searches as various memes are posted, which is why users are seeing so many of them disappear shortly after being posted on social media.

It’s an odd situation, as Google has been testing AI overviews for a year now – the feature launched in beta in May 2023 as Search Generative Experience – and CEO Sundar Pichai said the company has served more than a billion queries in that time. period.

But Pichai also said that Google has reduced the cost of delivering AI responses by 80% over the same period, “driven by technical, hardware and engineering advances.” It appears that this type of optimization may have happened very early, before the technology was ready.

“A company once known for being on the cutting edge and shipping high-quality products is now known for low-quality results that are being memeable,” said an AI founder, who wished to remain anonymous. On the edge.

Google continues to claim that its AI Overview product largely provides “high-quality information” to users. “Many of the examples we saw were unusual queries, and we also saw examples that were tampered with or that we were unable to reproduce,” Google spokeswoman Meghann Farnsworth said in an email to On the edge. Farnsworth also confirmed that the company is “taking swift action” to remove AI overviews in certain queries “when appropriate under our content policies, and using these examples to develop broader improvements to our systems, some of which have already began to be implemented.”

Gary Marcus, AI expert and professor emeritus of neural sciences at New York University, said On the edge that many AI companies are “selling dreams” that this technology will go from 80% correct to 100%. Achieving the initial 80% is relatively simple as it involves approximating a large amount of human data, Marcus said, but the final 20% is extremely challenging. In fact, Marcus thinks the last 20% might be the hardest thing of all.

“In fact, you need to reason to decide: is this plausible? Is this source legitimate? You have to do things like a human fact-checker would do, which may actually require artificial general intelligence,” Marcus said. And Marcus and Meta’s head of AI, Yann LeCun they both agree that the large language models that power current AI systems, like Google’s Gemini and OpenAI’s GPT-4, will not be what creates AGI.

Look, it’s a difficult situation for Google. Bing excelled in AI before Google with the famous “we made them dance” quote, OpenAI is supposedly working on its own search engine, a new AI search startup is already worth $1 billionand a younger generation of users who just want the best experience are switching to TikTok. The company is clearly feeling the pressure to compete, and the pressure is what makes AI launches confusing. Marcus points out that in 2022 Meta launched an AI system called Galactica that had to be withdrawn shortly after its launch because, among other things, told people to eat glass. Sounds familiar.

Google has big plans for AI overviews — the feature that exists today is just a small slice of what the company announced last week. Multi-step reasoning for complex queries, the ability to generate an AI-curated results page, Google Lens video search – there’s a lot of ambition here. But right now, the company’s reputation just depends on getting the basics right, and that doesn’t look great.

“[These models] they are constitutionally incapable of verifying the sanity of their own work, and that is what affects this industry,” Marcus said.