Google Gemini routes personal Drive files and emails to human annotators
In 2023, Google introduced Gemini Workspace extensions with a promise that personal content from Gmail, Docs, and Drive would remain inaccessible to human reviewers. However, recent changes have led to the reversal of this commitment. Google has confirmed that connected Drive and email content is now routed into human review systems, as stated in updated Google Help Answer 16836988 and a revised blog post from October 16.
The company explicitly warns users that what reviewers see may include only the email or file they are interacting with, but cautions that confidential information should not be shared in apps connected to Gemini. Testing reveals that this ingestion happens automatically in the background, without the need for specific tags like @Drive or @Gmail. Personal topics mentioned in casual conversations with consumer Gemini can trigger background Workspace lookups.
When using Gemini with Connected Apps enabled, the system generates one of four possible outcomes based on its probabilistic tool-calling and context injection: no retrieval, a high-level summary, partial unredacted extraction, or full-file ingestion. To test this behavior, users should ensure Connected Apps (Google Workspace) are active in their Gemini settings, set the system to Flash with High Thinking or better, and use casual queries about Drive or Gmail content without using @ tags.
Upon successful retrieval, users should print the exact text and metadata block returned by the retrieval tool. Testing with personal emails, such as an invoice from an internet provider, shows that Gemini can provide generic responses, metadata summaries, or even full personal data including account numbers and payment details.
Despite Google's attempts to deidentify sensitive information, the company's own documentation shows that automated scrubbers are not always effective, and the complete text of emails may be exposed.
Written by urgent.news from Hacker News's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.