Google details AI Overviews changes designed to tackle inaccurate content – SiliconANGLE

Google LLC has updated its recently launched AI Overviews feature to reduce the likelihood that it will display inaccurate information in search results.

Liz Reid, the head of Google Search, detailed the changes in a Thursday blog post. The update follows a series of highly publicized user reports about nonsensical results in AI Overviews.

Introduced at Google I/O earlier this month, AI Overviews is a new panel above the standard search results. It displays a detailed answer to the user’s query that can contain up to several sentences, as well as other types of data such as restaurant reviews. The feature is powered by Google’s Gemini series of large language models.

AI Overviews began rolling out to users in the U.S. on May 14. Over the next few days, reports started emerging about inaccurate and nonsensical responses. Typing “how many rocks should I eat?” into Google, for example, promoted AI Overviews to suggest a daily intake of one rock per day. In response to a query about pizza, the feature generated a recipe that listed glue as one of the ingredients.

Last week, Google issued a statement saying that it was taking “swift action” to address the errors. The company detailed that it was already in the process of rolling out technical modifications designed to tackle the issue. In the blog post, Reid provided more detailed information about what steps the search giant is taking to improve AI overviews.

The executive wrote that the company has so far made more than a dozen technical changes as part of the effort. Some of those updates, Reid detailed, will help Google’s search engine more accurately detect when it receives a nonsensical query that should not be answered with by AI Overviews. For added measure, AI Overviews will reduce the amount of satire and humor it includes responses.

Nonsensical content wasn’t the only factor behind the feature’s inaccurate answers. Some of the unhelpful content that AI Overviews generated was produced because the underlying LLM misinterpreted text on webpages, Reid detailed. 

“We worked quickly to address these issues, either through improvements to our algorithms or through established processes to remove responses that don’t comply with our policies,” the executive wrote.

Reid also shared an overview of how AI Overview works. According to the executive, the LLM that powers the feature is connected to the “core web ranking systems” Google uses to generate search results. AI Overviews is “built to only show information that is backed up by top web results,” she added.

Google’s push to improve the quality of the information in AI Overviews comes as it gears up to significantly expand the feature’s availability. Currently, the capability is only available in U.S. When Google detailed AI Overviews earlier this month, it detailed that the plan is to make the feature accessible to more than a billion users by the end of the year.

The search giant also intends to explore whether AI Overviews could be monetized, which adds more urgency to its push for better response accuracy. On May 21, Google detailed an upcoming test that will evaluate whether ads can be embedded in the feature’s AI-generated answers. The ads are set to appear in a dedicated section that will display a “Sponsored” label. 

Image: Google