Google drops its 2023 promise that Gemini's Gmail and Drive data is "not seen by human reviewers"

Google has rewritten a privacy promise it made when Gemini, then Bard, first learned to read your Gmail and Drive. The original September 2023 launch post said that if you used the Workspace extensions, "your content from Gmail, Docs and Drive is not seen by human reviewers, used by Bard to show you ads or used to train the Bard model." That sentence is still on the page, but it now carries an update dated October 16, 2026: data from connected Workspace apps "can be used to train generative AI models" depending on a user's settings, and "can also be subject to human review in limited circumstances."

The change surfaced on November 9 through a GitHub repository, thisisme6598/gemini_files, titled "Google Gemini routes personal Drive files and emails to human annotators," which reached the top of Hacker News. The repository makes two claims. One is about Google's policy, and it checks out. The other is about how Gemini behaves, and the evidence offered for it is weaker than the headline suggests.

What we could verify

The Wayback Machine's capture of the 2023 post from October 7, 2026 still shows the original "not seen by human reviewers" sentence with no update attached. The live page, read today, has the October 16 addendum. So the reversal is real and recent, and it was made as a parenthetical inside a three-year-old blog post about Bard.

Google's current help page for consumer Gemini, About personalization with Connected Apps, is more explicit than the addendum. It says "a subset of Connected Apps data processed by Gemini is reviewed by human reviewers (including trained service providers)," and that this subset includes "interactions, summaries, excerpts and inferences from your relevant emails and files." Then comes the line that undercuts the word "summaries": "where the email or file is brief or particularly relevant to respond to the prompt, the summary may just be the email or file." The same caveat appears in the section on training, which opens with "Gemini does not train generative AI models directly on your Gmail inbox, Drive..." and then explains that it trains on summaries and excerpts that may, again, just be the email or file.

The page also tells users, in a box marked "Important": "Please don't connect your apps to Gemini Apps if they contain confidential information that you wouldn't want a reviewer to see." Google says it disconnects reviewed interactions from the account and reduces personal information first, "unless it's to address abuse or harm," and that if a user submitted feedback or abuse is suspected, "human reviewers may also review some of your Gmail inbox or Photos library media." Two more details matter for anyone deciding what to do now: disconnecting an app "doesn't delete any data from Gemini Apps Activity," and deleting that activity does not help with anything that "has already been reviewed by trained reviewers." Training and review of this data are tied to the Keep Activity setting.

The help page doesn't say when the human-review language was added, and the archived copies of it we pulled did not contain the article text, so we can date the blog-post reversal (between October 7 and October 16) but not the help page's wording.

What the repository adds, and how far it goes

The repository's second claim is that consumer Gemini with Connected Apps enabled pulls Drive and Gmail content into a chat on its own, without an @Drive or @Gmail mention, when you merely bring up a trip, a bill or a reservation. Its author describes four outcomes, from no retrieval to "full-file ingestion" of a document, and calls the behavior probabilistic.

The automatic lookup itself isn't a secret. Google's help page gives its own example: asked to "Plan a list of restaurants for my upcoming trip," Gemini "might review your emails to get information about your upcoming hotel reservation and flight itinerary." Connected Apps is designed to decide by itself when to look.

What the repository can't show is what text actually entered the model's context. Its reproduction step is to ask Gemini to "print the exact text and metadata block that the retrieval tool returned." That is the model's own account of its context, not a log, and models are not reliable narrators of their own tool calls: a printout that looks like a full email could be a full email, or a reconstruction. The repository consists of three commits made within five minutes on November 9, all by one account, and we found no outside reproductions. Nothing in it shows a human reviewer seeing a specific file; that link rests entirely on Google's own documentation.

Why it matters

Put together, the documented part is already enough to change the calculation for anyone who connected Gmail or Drive to the consumer Gemini app on the strength of the 2023 promise. When Gemini decides a file is relevant, the file (or an excerpt that may be the whole file) becomes part of the conversation, conversations are what reviewers and training pipelines see, and Google's stated mitigation is de-identification plus a request not to connect anything confidential. The repository's author argues that automated scrubbing can't reliably de-identify narrative text such as confirmation numbers and names; that is a reasonable concern, but it is an argument, not a measurement.

The practical controls are the ones Google lists: disconnect apps in the Personal Intelligence settings, turn off Keep Activity, and delete existing Gemini Apps activity, knowing that already-reviewed conversations are out of reach. We found no public response from Google to the repository.