There has been a buzz on social media about Google’s new AI Overview product providing bizarre responses like suggesting users to put glue on their pizza or eat rocks. Due to the chaotic launch, Google is now working quickly to manually deactivate AI Overviews for specific searches as memes flood social networks, leading to many disappearing shortly after being shared.
This situation is unusual given that Google has been testing AI Overviews for a year, with the feature initially launching in beta in May 2023 under the name Search Generative Experience. CEO Sundar Pichai mentioned that the company handled over a billion queries during this period.
Despite making significant progress in reducing the cost of delivering AI responses by 80 percent, as per Pichai, through hardware and technical improvements, it seems that this optimization might have been implemented prematurely, before the technology was fully prepared.
An anonymous AI founder expressed disappointment, noting that a company once known for its high-quality output is now associated with low-quality results that become memes, according to The Verge.
Google maintains that its AI Overview product largely provides “high quality information” to users. Google spokesperson Meghann Farnsworth mentioned that many peculiar instances were uncommon queries or edited content that could not be replicated. The company is swiftly addressing these issues by removing inaccurate AI Overviews for specific queries, using them to enhance their systems further, with some improvements already underway.
Gary Marcus, an AI expert and professor, mentioned that while many AI companies promise that their tech will achieve 100 percent accuracy from the current 80 percent level, the last 20 percent is the most challenging. Identifying the plausibility of responses and verifying sources akin to human fact-checkers may require artificial general intelligence, a level of reasoning beyond current capabilities.
According to Marcus and Meta’s Yann LeCun, future advancements in AI are unlikely to be solely based on large language models like Google’s Gemini and OpenAI’s GPT-4, indicating that current AI systems may not lead to artificial general intelligence (AGI).
The competitive landscape and the pressure to innovate are driving companies like Google to push out AI products quickly, leading to occasional messy releases. Marcus referenced Meta’s experience with the Galactica AI system in 2022, which had to be removed due to issues such as suggesting unsafe actions like eating glass.
Google envisions expanding AI Overviews beyond its current capabilities with features like multistep reasoning, an AI-organized results page, and video search in Google Lens. However, the company’s current focus remains on ensuring the accuracy and reliability of its basic functions for now, as maintaining a positive reputation is crucial.
Marcus highlighted that these AI models lack the ability to fact-check their own output, which has become a significant challenge for the industry.
