Google has also been embarrassed by its AI profiles. After a week of criticism and memes decrying the poor quality and outright misinformation that stemmed from the tech giant's unfinished AI-powered search feature, the company issued an apology of sorts on Thursday. Google, a company whose name is synonymous with web search and a brand focused on “organizing the world's information” and putting it at users' reach, actually wrote in a blog post that it “certainly saw some strange, inaccurate, or unhelpful AI profiles.”
That's an understatement.
This admission of failure, written by Liz Reid, Google's vice president and head of search, seems like evidence that the push to incorporate AI technology into everything has somehow made Google search worse.
In a post titled “That thing from last week” (did this get through PR?), Reed goes into detail about the various ways that AI Overviews makes mistakes: While it doesn't “hallucinate” or make things up like other large-scale language models (LLMs), Reed says it can make mistakes for “other reasons,” including “misinterpreting the query, misinterpreting the nuances of language on the web, or not having a lot of good information available.”
Reid also noted that some of the screenshots shared on social media in the past week were fake, while others were of nonsensical queries that no one has ever searched for before, such as “how many rocks should I eat.” With little factual information on the topic, Google's AI directed users to satirical content (in the case of the rocks, the satirical content was published on the website of a geology software provider).
It's worth pointing out that if you Google “how many rocks should I eat?”, you won't be surprised if you get a bunch of useless links and joke articles. What people are reacting to is the AI confidently answering “geologists recommend eating at least one small rock per day” as if it were a factual answer. It may not technically be a “hallucination,” but the end user doesn't care. It's insane.
Even more disturbing is that Reed claims that Google “thoroughly tested the feature prior to release, including through an extensive red team effort.”
So does no one at Google have a sense of humor? Did no one ever think of a prompt that would produce bad results?
Additionally, Google downplayed the AI feature's reliance on Reddit user data as a source of knowledge and truth. People have been adding “Reddit” to searches so often that Google finally made it a built-in search filter, but Reddit is not a body of fact-based knowledge. And yet the AI shows Reddit forum posts as answers to questions, unable to understand when firsthand Reddit knowledge is helpful and when it isn't, or even worse, when it's a troll.
Reddit is currently making a ton of money by providing data to companies like Google and OpenAI to train their models, but that doesn't mean that users want Google's AI to decide when to look for answers on Reddit or suggest that someone's opinion is fact. Learning when to search Reddit is nuanced, and Google's AI hasn't figured it out yet.
“While forums are often great for providing a reliable, first-hand source of information, in some cases they can lead to less-than-useful advice, like using glue to stick cheese to pizza,” Reid acknowledged, pointing to one of the AI's most notable failings over the past week.
A Google AI summary suggests adding glue to pizza to make cheese stick, but the source of that suggestion turns out to be an 11-year-old comment by Reddit user F*cksmith 😂 pic.twitter.com/uDPAbsAKeO
— Peter Gyang (@petergyang) May 23, 2024
But if last week was a disaster, at least Google is responding quickly in its wake — or at least that's what it says.
The company says it has reviewed examples of AI Overviews and identified patterns where it can make improvements, including improving detection mechanisms for nonsensical queries, restricting users from user-generated content for responses that may provide misleading advice, adding trigger restrictions for queries where AI Overviews would not be helpful, hiding AI Overviews for hard news topics where “freshness and factuality are important,” and adding improved triggers for protections against health searches.
With AI companies building ever-improving chatbots every day, the question isn’t whether chatbots will be able to surpass Google Search in making sense of the world’s information, but whether Google Search can catch up with AI and challenge chatbots.
However silly Google's mistake may be, it's too early to assume the company is out of the running, especially given the sheer size of its beta testing team, which includes virtually everyone who uses search.
“The fact that millions of people are using this feature for so many new searches is unique,” Reed said.