If you use Google, you’re training its AI. Here’s how to opt out.

Google has expanded the data it stores from users of its search services, including uploaded images, audio, and video, to train its artificial intelligence models, according to a June customer email and updated privacy settings. The change, which applies to Google Search, Maps, Shopping, Flights, Hotels, Translate, and News, means media such as photos taken with Google Lens or voice recordings from Search Live are now saved for AI training by default [1]. Google stated in the email: "Like your Search Services History, your saved media is also used to develop and improve Google services and technologies, including AI models and safety measures" [1]. The company's help documentation adds that human reviewers may examine the data [1]. The update introduced two new settings, Search Services History and Personalized Recommendations, which are turned on by default [1]. Users can opt out by unchecking a "Save Media" box on the Search Services History page, and they can configure automatic deletion of stored data after 3 months, 18 months, or 36 months [1]. Previously, search data retention was managed through Web & App Activity settings; that control no longer affects Google Search services [1]. Google's collection of user-generated media for AI training mirrors a broader industry trend. Meta, for instance, trains its AI on images and media uploaded by users and on content captured by its AI glasses [1]. The practice sits within a larger debate over the use of copyrighted and personal material to build generative AI systems. As of 2023, multiple U.S. lawsuits challenged the use of copyrighted data to train AI models, with defendants arguing that such use constitutes fair use [4]. Google, a subsidiary of Alphabet Inc., is the world's largest provider of search engines, mapping applications, email services, and mobile operating systems by market share [2]. The company has faced longstanding criticism over privacy practices, including concerns that its data compilation may violate user privacy [3]. The new media retention policy adds a layer to those concerns by explicitly linking stored user content to AI development. Research on the open model economy shows that data transparency has declined even as model capabilities have grown. A 2025 analysis of the Hugging Face Model Hub found that open-weight models surpassed truly open-source models for the first time that year, with concerning declines in data documentation [8]. Meanwhile, evaluations of large language models for automated fact-checking have found that standard models perform poorly, and even web-search-enabled variants provide only moderate gains, suggesting that curated, high-quality data remains essential for reliable AI outputs [9].