Google expands AI training on user media unless you opt out

Google announced that all signed‑in users are now automatically enrolled in a policy that saves uploaded images, audio, video and other media to its Search Services History for training its large language models, unless the user manually disables the setting [1].

Expanded data collection for AI models

The updated privacy settings, communicated via email last month, add “saved media” to the list of data Google may use to improve its AI products and services [1]. The language now says that any media saved to Search Services History can be employed for “future searches and training AI,” covering everything from photos uploaded to Google Photos to voice queries made through Search Live or Google Translate. This broadens the data pool beyond text, mirroring Google’s push to enhance its Gemini‑based search experience.

How users can retain privacy

Google provides two settings to control the new data flow. In the Google app, users can open Search history, ensure the “Save media” toggle is off, and then turn off the overall Search Services History entry. A second page, Search personalization, also offers a toggle to stop personalized results [1]. For devices where these options are missing, the web portal myactivity.google.com lets users uncheck “Save media” and choose retention periods of 3, 18 or 36 months [1]. Disabling personalized ads via myadcenter.google.com completes the privacy stack.

What to watch

• Policy enforcement timeline – Google may roll out the opt‑out controls to additional platforms or regions in the coming weeks.
• AI model performance – Track whether Google cites improvements in Gemini or other AI services that it attributes to the expanded training data.
• Regulatory response – Monitor any statements from data‑privacy regulators that could pressure Google to modify or clarify the policy.

The move signals Google’s willingness to leverage ever more user‑generated content to sharpen its AI, while still offering a manual opt‑out path. Whether the trade‑off between richer models and reduced privacy will shift user behavior remains to be seen.