Gemini AI Now Scans Your Google Photos (If You Opt In)

Google has started rolling out an update to Google Photos that lets its Gemini AI model scan a user’s photo library to support more personalized AI image generation, according to the company. The rollout began on April 18, 2026 , requires users to opt in, and is launching first in the U.S. before expanding globally.

The feature is positioned under Google’s “Personal Intelligence” initiative. With the connection enabled, Gemini can draw on personal images of users and their contacts to produce AI-generated content that better reflects an individual’s context and preferences, the company said. Google framed the change as a way to make image generation more direct by reducing the need to manually upload reference photos or write long, detailed descriptions.

Google said linking Google Photos to Personal Intelligence helps Gemini interpret user preferences and apply personal context inside its outputs. The company also described this Photos connection as part of a broader approach in which Gemini can access information from other Google services to provide more tailored assistance. In that same framework, Google said Gemini can use data from products such as Calendar and Gmail when users choose to connect them.

On data use, Google said the Gemini app does not directly train its models on private Google Photos libraries. At the same time, the company said it does train on limited information—such as specific prompts and model responses—to improve how the system works. Google emphasized that users remain in control of the opt-in experience and can change settings at any time.

The company advised users to weigh privacy considerations before connecting personal data to the AI platform. The update’s U.S.-first rollout, followed by a wider international release, means the feature could become relevant to users and regulators across multiple jurisdictions as it expands. Google’s approach combines personalization across services with user-controlled settings, while also acknowledging that some interaction data may be used to refine performance.