Google is rolling out a new Search Services History setting that pulls images, audio, and video from a user's interactions with Search into its AI model training pipeline — and the setting is on by default. The change reaches accounts globally over the next few months, and notification emails with the subject line "New privacy settings for Search services" began landing on June 23. Training data extracted under the new policy will be retained for up to 4 years, even after a user deletes the underlying activity.
The scope of the collection is wider than typed queries. Per Google's own description, saved media includes Google Lens images, recordings from Search Live, Translate speaking-practice audio, uploaded files, and voice searches. Users who had previously disabled Web & App Activity and Search Personalization will find the new toggle off; everyone else gets it enabled, with the "Save media" box already checked.
To opt out, users can visit the My Activity page, open the Search Services History tab, uncheck the Save media box, and optionally delete prior activity. Google said in a statement that the controls can be flipped at any time.
“These new settings help users get more relevant results and revisit their searches—including visual and voice searches—and they can be turned on or off at any time.”— Davis Thompson, Google spokesperson
Key facts
- 01Google's new Search Services History setting is enabled by default and feeds Lens images, voice queries, and Translate audio into AI model training.
- 02Training data is retained for up to 4 years even after a user deletes the original activity from their Google account.
- 03The rollout begins globally over the next few months, with Google notifying users by email starting June 23.
- 04Users can opt out via Google's My Activity page by unchecking the 'Save media' box under Search Services History.
Google did not address why the feature ships enabled by default. The 4-year retention window is disclosed inside a pop-up that appears only when a user moves to turn the setting off — a placement that puts the longest-lived consequence behind the action most users will never take.
The mechanic matters because multimodal training data is now the scarce resource at frontier labs. Text scraped from the open web is largely exhausted as a marginal source of capability gains, and the next round of model improvements depends on diverse inputs: images, spoken audio, video, and the messy real-world artifacts users hand to a search engine. Google Search, Lens, and Translate sit on one of the largest such pipes in existence.
That scale is also the leverage point. Google ships Search, Gmail, Maps, Android, Chrome, Photos, YouTube, and Workspace to billions of users who have spent a decade or more inside those products. Switching costs are high, and a default-on setting buried in account preferences will convert at rates a competitor running an opt-in flow cannot match.
Opt-out as the industry default has hardened across the AI sector over the past 18 months, from Meta's training disclosures in Europe to Anthropic and OpenAI's evolving data-use terms. Thorin Klosowski of the Electronic Frontier Foundation argues the bar should be higher: "I think 'opt in' is really asking the bare minimum of these companies. Asking their users to consciously choose to enable these features is the least they can do." Without that, he says, Google would need to make a real case to users for why the feature helps them.
Ben Winters, director of AI and privacy at the Consumer Federation of America, frames the user-side cost in time and attention. "It creates this extra layer of math that a consumer has to do about whether they feel comfortable using the tool they've been using for a long time," he says.
“There's an increasing feeling of powerlessness and hopelessness about even trying to protect your data, because every little thing is going to be squeezed out of you.”— Ben Winters, Director of AI and privacy, Consumer Federation of America
There are limits to what the disclosure actually promises. Google says training data is "disconnected from your Google Account" once ingested, which means deletion requests against the account no longer reach the model's training corpus. The 4-year clock runs from ingestion, not from the user's last interaction, and Google has not published a mechanism for users to confirm whether any specific image or recording was in fact used.
Regulators in the EU and UK have spent the last two years pressuring large platforms on exactly this pattern — default-on data flows into AI training, with retention windows that outlast the source record. Google's rollout is global, which means the same setting will land in jurisdictions with very different enforcement postures, and the company has not said whether EU users will see a different default.
For Google, the calculus is straightforward: more multimodal data tightens the feedback loop on Gemini and on the AI features now embedded across Search, Lens, and Translate, and a default-on toggle is worth more model-quality progress than any single licensing deal. For users, the practical move is a two-minute visit to My Activity to flip the Save media box before the setting reaches their account. Expect the opt-out-by-default pattern to remain the industry norm until a regulator forces a flip, and expect retention windows measured in years, not months, to become standard disclosure across every major AI provider.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
