As personal AI assistants become more deeply integrated into everyday office ecosystems, data privacy and security boundaries have once again become a focus for the industry and users. Recently, a developer released a test report indicating that after the personal version of Gemini is linked to Google Workspace, it may automatically search files stored in Gmail or Google Drive based on everyday conversations. Google also clearly states in its official help documentation that some linked app data processed by Gemini may be sent for human review.
Automatic Retrieval and Accidental Disclosure: Calls Can Be Triggered Even When Users Do Not Explicitly Ask
According to the developer's hands-on test report, after the personal version of Gemini enables app linking, when users ask questions such as "Are my internet service fees too high?", the AI will automatically search in the background for emails from the internet service provider in their mailbox, even though the user did not explicitly ask it to do so.
In different test scenarios, the content Gemini presents varies widely, including bill summaries, names and account information, email snippets, and in some cases even fairly complete email text. This deep mining means that when users handle everyday privacy or personal matters, the emails and files the AI can access and expose may exceed their initial expectations.
Purpose and Process of Human Review: Outsourced Teams Step In, with De-identification Mechanisms
As for why human review is conducted, Google explains in its help documentation that this data is checked by trained service provider personnel (that is, outsourced content review specialists), with the core aims of troubleshooting system issues, ensuring platform security, and continuously improving the service.
The scope of review includes not only summaries and excerpts of emails and files, but also related inferences and records of user interactions with Gemini. Google specifically notes that if the email content itself is short, or is highly relevant to the user's question, then under certain specific circumstances the summary may even directly include the email or file itself.
To protect user privacy, Google emphasizes that before content enters the review process, it takes technical measures to minimize personal information, such as disassociating the related interaction records from the user's Google account. However, Google also candidly acknowledges that existing mechanisms cannot remove all personal information 100 percent, so it officially recommends that users exercise caution when linking apps that contain confidential or sensitive information.
Model Training and User Control: Users Can Turn Off Activity Saving
Regarding the scope of data use, Google says that when users enable activity saving, Gemini's activity records and the linked app data involved may be used for subsequent product improvements and AI model training.
To give users more control, Google allows users to turn off activity saving in Gemini's app activity settings according to their own security needs, thereby preventing their personal data from being used in the model training process.
