Every image sent through Lens, every spoken query, and every file fed to Translate can feed Google's generative models — unless you manually flip two settings buried in Search preferences that most users have never opened.
Google is now training its AI on Lens photos, voice queries, and Translate uploads by default. Every image sent through Lens, every spoken query, and every file fed to Translate can feed Google's generative models unless a user manually flips two settings that most people will never find.
The shift, first reported by TechCrunch and summarized by Engadget, is the company's response to a tightening training pipeline. Generative AI work is running out of public text and images, and the consumer Search funnel now counts as a renewable source. Every Lens snapshot, every spoken query, every uploaded file that crosses a Search product is fair game by default.
Google's Search Services History page lists the products in play: Lens image searches, voice and voice search, and Translate file uploads all count. Personal Google Photos is not in this set, and the company has not changed the Photos data terms in parallel. A photo backed up to Photos stays on the older privacy footing; a photo sent through Lens does not.
Every account, including signed-out sessions where Google can still stitch the upload to an identity, is enrolled automatically. The opt-out lives across two pages few users have ever visited. On the Search Services History page, the relevant control is a "Save Media" checkbox tied to each Search surface. On the Search Services Personalization page, users must confirm that nothing tied to their account is being retained for AI training. Both toggles default to on, and both require a Google account. Signed-out Search users have no settings to flip.
Google did not surface a one-time prompt when the policy expanded, did not email existing users, and did not announce the change on a Search or Help Center blog, according to Engadget and the underlying TechCrunch report. The data started flowing without any such notice. An opt-out toggle in a setting most users will never open is the kind of checkbox that satisfies a compliance reviewer, not the user who never agreed to the change.
The workaround most users will hear about is a query-trick suffix that solves a different problem. Prefixing a query with "-AI" suppresses AI Overview results in Search, which is about what the model returns, not what it was trained on. The two settings pages are the actual lever, and they only matter to a user who already knows the data is in use.
The next real test is whether the same default-on posture reaches Workspace or Education Search, where consent rules are tighter and a forced-enrollment policy would face a faster legal response.