Google is imposing limits on its AI Reviews search feature after producing several bizarre and inaccurate results, the company announced on May 30.
AI Overviews, introduced at Google I/O 2024 on May 14 after nearly a year of experimentation, offers users quick summaries of AI-generated topics with links for deeper exploration, appearing at the top of search results, according to Moneycontrol.
Initially made available to “hundreds of millions” of users in the United States, the tech giant plans to expand AI previews to more than a billion people by the end of year. The launch is part of Google’s efforts to revamp its search product for the generative AI era, amid renewed competition from companies like Microsoft, OpenAI and new entrants like Perplexity.
Google’s AI accused of “hallucination”
However, last week the feature reportedly produced unusual and incorrect advice, such as suggesting users eat rocks or use glue on pizza to help the cheese stick, screenshots of which have been widely circulated online.
In a May 30 blog post, Liz Reid, head of search at Google, said the company has now implemented mechanisms to detect “nonsense queries” and has limited the inclusion of satirical and humorous content in AI insights.
Additionally, Google updated its systems to reduce the use of user-generated content that could offer misleading advice and added restrictions for queries where the AI insights weren’t particularly helpful. The company is also addressing the “small number of AI previews that violate content policies,” which include potentially harmful or obscene information.
Reid noted that policy violations were found in “fewer than 1 in 7 million unique queries” featuring AI insights. She argued that AI insights generally do not “hallucinate” or invent information; errors usually come from misinterpretation of queries, nuances of language or insufficient information.
Google AI Presentation Improvements
Reid compared AI previews to “featured snippets,” a long-standing search feature that uses AI to highlight key information with links to web content, saying their accuracy rates are comparable. She noted that the feature underwent extensive testing before launch, including a rigorous red team, evaluations with typical user queries, and testing on some search traffic.
Additionally, Reid said Google won’t show AI previews for hot topics where timeliness and accuracy are crucial, and that for some health topics the company has added improvements additional to improve its quality protection.
These updates come as Google aims to monetize its AI features by testing search and shopping ads in AI previews.
First publication: May 31, 2024 | 7:29 p.m. STI
