Consumer Gemini now routes personal files and emails to human reviewers
The documentation Google publishes now acknowledges that human reviewers can see connected Drive files and Gmail messages on consumer Gemini accounts. That is the opposite of the 2023 promise, made when the company launched the Gemini Workspace extensions, that personal content from Gmail, Docs and Drive "is not seen by human reviewers". A write-up published on GitHub on November 9 documents the change, shows how to trigger it, and shows that the content ingestion happens without an explicit request.
What Google's docs say now
When Google shipped the Gemini Workspace extensions in 2023, it attached an explicit promise: personal content from Gmail, Docs and Drive would not be shown to human reviewers. According to the write-up, that commitment was quietly reversed. An updated Google Help answer (16836988) and an October 16 edit to the original 2023 blog post now confirm that connected Drive and email content is routed into human review pipelines. The documentation carries a plain warning: if the apps you connect hold confidential information you would not want a reviewer to see, do not connect them to Gemini Apps.
The write-up scopes this to consumer accounts. Personal Gmail accounts run under the standard Gemini Apps privacy notice, while Workspace Business, Enterprise and Education accounts sit under a separate data processing addendum that excludes human review by default.
How the ingestion works in the author's tests
Two findings from the testing section matter most. First, no explicit request is needed. With Connected Apps enabled, a casual mention of a personal topic in conversation (a trip, a bill, a reservation) can trigger a background Workspace lookup, with no @Drive or @Gmail tag. Second, what gets pulled is probabilistic. The write-up's repro asks the model to print the exact text its retrieval tool injected into its prompt. Repeated trials against a single inbox email, an internet service bill, produced all four observed outcomes:
- Zero retrieval: the model answers from its own weights or from conversation history.
- Metadata or generic summary: "you have 3 emails regarding your flight".
- Partial, unredacted extraction: 2 to 4 sentences, often including names, account numbers and confirmation codes.
- Full file ingestion: the complete unredacted text of the email, headers included.
The author's caveat is that seeing outcome one or two in a single test does not disprove the pipeline. That is one draw from a nondeterministic system, and the write-up says to delete the conversation and test again. It also argues that Google's regex and NER scrubbers cannot sufficiently deidentify narrative data, including account numbers, transaction dates, and combinations of names, cities and events that identify a person in combination.
What it means for your accounts
If you run consumer Gemini with Connected Apps on, the practical read is that connected content can reach human contractors during casual conversation, and the isolation that sold the integration is no longer part of the terms for personal accounts. Keep confidential material out of a connected consumer account, or disconnect the apps. For Workspace deployments the write-up says the contractual position is different: human review is excluded by default. That is what the documentation states today, and the language of the DPA that applies to your deployment is what to check before you rely on any guarantee.
Source: github.com
More news
Have something in mind?
Start a conversation