Google is expanding its “AI Overviews” feature to include medical advice, despite a history of providing inaccurate information. Chief health officer Karen DeSalvo announced that advancements in Gemini models will allow the feature to cover “thousands more health topics.”
The update introduces “What People Suggest,” a feature that aggregates health advice from users on the internet. While DeSalvo stated that this feature is available on mobile in the US, users have reported difficulties accessing it on both the Google app and the web.
Concerns surrounding the accuracy of Google’s AI Overviews have been well-documented. Reports highlight instances where the AI provided erroneous information, such as claiming baby elephants could fit in a human hand and suggesting glue as a pizza topping. A study by Columbia’s Tow Center for Digital Journalism found that Google’s Gemini chatbot returned incorrect answers to basic questions 60 percent of the time.
A Google spokesperson indicated that the “What People Suggest” feature aims to help users find relatable health information. The company claims this feature has undergone rigorous testing and clinical evaluation by licensed medical professionals and appears alongside authoritative health content on Search.
In 2021, Google closed its health division, leading to layoffs and reorganizations of staff focused on health-related roles. Experts have expressed skepticism about the ability of Google’s AI to deliver reliable healthcare information, citing past failures in health tech initiatives. Emarketer senior analyst Rajiv Leventhal told Bloomberg that “nobody in the big tech world has succeeded” in disrupting the healthcare industry, which he described as a “unique beast.”
Google has been questioned about its methods for ensuring the accuracy of the health information presented by its AI. Previously, representatives stated that insufficient high-quality web content may contribute to errors in the AI’s responses. The company claimed to have “guardrails and policies” to protect against low-quality outputs and asserted that it uses problematic cases to guide improvements.
Featured image credit: Kerem Gülen/Imagen 3
