Google announced that it will introduce a new "Personal Intelligence" feature to its Gemini app. This feature allows Gemini to securely connect and read information from users' Gmail, Google Photos, YouTube, and other applications, providing assistance that is more closely aligned with the user's individual needs through cross-source reasoning capabilities.
This feature is currently in beta and is being offered first to Google AI Pro and AI Ultra subscribers in the United States.
Cross-application reasoning: No more rummaging through drawers to find license plates or plan your trip.
Google emphasizes that the core advantages of "personalized intelligent services" lie in "cross-source reasoning" and "detailed extraction." After users actively enable the connection in the settings, Gemini can combine text, photos, and video information to answer questions.
Google gave a relatable example: when a user needs to change the tires on their 2019 Honda SUV, Gemini not only provides tire specifications like a typical robot, but also analyzes vehicle photos in Google Photos to suggest tire models suitable for local road conditions. Even when the user needs the license plate number, Gemini can directly identify the plate from the photo album and confirm the specific vehicle model by searching Gmail, saving the user the hassle of switching between different apps.
In addition, when it comes to travel planning, Gemini can analyze users' past records in Gmail and photo albums to understand the interests of family members, thereby avoiding crowded tourist attractions and recommending itineraries and board game activities that are more in line with family preferences.
Privacy design: Off by default, not used directly for model training.
Regarding the privacy concerns that users care about most, Google stated that the feature is "off by default," giving users complete control over which applications to connect to.
Google emphasizes that Gemini accesses data to respond to requests while operating, but "does not directly use users' Gmail inboxes or Google Photos libraries to train the model." Google only uses "specific prompts" and "model responses" after filtering and obfuscating personal data for subsequent optimization training. The system is trained to "understand how to find data," not to "remember your data."
This is still in beta: Beware of AI "over-interpreting" things.
Despite its powerful features, Google admits that it is still in the testing phase and may be subject to misjudgments due to "over-personalization".
For example, if a user's photo album contains hundreds of photos taken at golf courses, Gemini might mistakenly assume the user "loves golf," overlooking the fact that the user was simply accompanying their son to play and may not actually enjoy the sport themselves. Google encourages users to correct such misjudgments directly in the conversation (e.g., "I don't like golf") or provide negative feedback to help the model adjust its understanding of interpersonal relationships and interests.
Online Information
Personalized AI services will be rolled out to eligible Google AI Pro and AI Ultra individual subscribers in the United States starting today, over the next week (excluding Workspace Business/Education versions). The feature is available on web, Android, and iOS, and will be expanded to Google Search's AI mode and more countries and regions in the future.
Analysis of viewpoints
Compared to Apple's emphasis on the privacy advantages of "device-side computing," Google's strength lies in its vast "cloud ecosystem." For many, Gmail and Google Photos are practically the entire digital memory bank of their lives. And Gemini's ability to connect this data means it understands your life's context better than anyone else.
However, the challenges remain with "trust" and "accuracy." Although Google has repeatedly assured users that it will not use raw data to train models, will users dare to entrust important decisions (such as medical or financial advice) to AI in cases of hallucination or "over-interpretation" as mentioned in the article? Furthermore, how to enable AI to understand complex human social relationships (e.g., people in a photo album are not necessarily friends, but could be unpleasant exes) will be a technological hurdle for the true realization of "personalized intelligence" in the future.
