Google's Own Help Page Now Says Don't Connect Confidential Files to Gemini
Google's Own Help Page Now Says Don't Connect Confidential Files to Gemini
A Hacker News thread on 9 November 2026 rose near the top of the site with a claim that reads like a leak: Google is routing personal Drive files and emails from Gemini to human annotators. The thread pointed at a public GitHub writeup by one user, not at a document dump. The most useful evidence turned out to be somewhere else entirely. It was in two pages Google publishes itself.
Neither page is hidden. Both are indexed and linked from the product. One is the help article about connected apps. The other is a 2023 blog post on responsible AI guardrails that now carries an update note dated 16 October 2026. Read together, they say something a lot of teams have not absorbed: the promise that your work files would stay away from human eyes has been rewritten in place, with conditions attached.
The promise, and the condition added to it
The 2023 post, Our responsible approach to building guardrails for generative AI, made a clear statement about Workspace extensions in Bard:
If you choose to use the Workspace extensions in Bard, 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. Immediately after it, in the same paragraph, the page now carries this:
(Update, October 16, 2026: You can view the latest about how some new features work here. As we've continued to make Google products more personalized and helpful, we've added features that enable users to choose to connect Workspace apps to other Google services. If users choose to enable these new features, data from Workspace Apps will be processed to personalize other Google experiences, and depending on a user's settings, that same subset of data can be used to train generative AI models. It can also be subject to human review in limited circumstances.)
No new announcement, no separate post. An update note inside the original claim, which is how a lot of vendor terms change.
What the help page says today
The help article on personalization with Connected Apps is more specific, and it is the page that describes the feature.
When connected to Gemini, a subset of Connected Apps data processed by Gemini is reviewed by human reviewers (including trained service providers) to debug our services, help keep them safe, and improve them.
It then gives an instruction that is unusual to find in first party product documentation:
Important: Please don't connect your apps to Gemini Apps if they contain confidential information that you wouldn't want a reviewer to see.
The reviewed material includes interactions, summaries, excerpts and inferences from your relevant emails and files, plus prompts, responses and connected Wallet and payment information. For files, there is this clause:
In some cases, for example, where the email or file is brief or particularly relevant to respond to the prompt, the summary may just be the email or file.
So the boundary between "a summary" and "the document" is decided by the model's judgement about relevance, not by a redaction rule. It is worth being precise about what this does and does not mean:
- It is documented policy. Google names the reviewers (including trained service providers), the purposes (debugging, safety, service improvement) and the data classes.
- It is not a claim that a person reads every chat. The page says a subset of connected app data is reviewed.
- It is not limited to abuse reports. Normal file content pulled in to answer a prompt is in scope.
- Photos are the exception. The page says humans do not view media from your Photos library, except for abuse and harm cases or when you submit feedback.
Deleting the file does not delete the reviewed copy
The practical trap sits in the same article, under disconnecting apps.
If you disconnect a connected app or delete any data in that app, it doesn't delete any data from Gemini Apps Activity. That data can still be used for product improvement, including training generative AI models. You can delete your Gemini Apps activity to prevent it from being used for future product improvement, unless it has already been reviewed by trained reviewers.
Read the last clause twice. Once a reviewer has seen it, deleting your activity does not reach that copy. Search Engine Journal's 2024 writeup of Google's Gemini privacy pages collected the retention detail: reviewed conversations are kept separately from the account, are not removed when you delete your activity, and are retained for up to three years. The same reporting notes that turning activity saving off stops future data from going to reviewers, which is a control that has to be on before the file is shared, not after.
Anyone who has cleaned up a spreadsheet the day after a board meeting will recognise the shape of this. The cleanup happens in the system you control. The copy already left.
One tester's finding, and what it does not show
The GitHub writeup that started the thread, gemini_files, goes further than the policy pages. The author reports that with Connected Apps enabled, Gemini triggers background Workspace lookups from ordinary conversational mentions, with no @Drive or @Gmail tag, and that what comes back varies between attempts: nothing, a metadata summary, a short verbatim quotation, or the full text of the file.
That is one person's testing, and it should be read that way. The author argues the variability is expected because tool calling is probabilistic and a single test proves little either way. Commenters in the Hacker News thread pushed back from the other side, noting that behaviour like this is hard to separate from model variance and that no evaluation with published methodology is attached.
What the writeup does establish is narrower and still useful. The policy already says brief or relevant files may be reproduced rather than summarised. Whether a given prompt pulls in a file is a decision made at request time, so "would this happen to us" cannot be settled by reading the settings screen.
What to put in place
This is a governance item that arrives disguised as a feature toggle. A few things worth doing this week:
- Inventory the connectors, per account type. Consumer Gemini Apps and a managed Workspace account do not run on the same terms. List what is connected, in which account, and for whom. Most organisations find personal accounts doing work the company account was meant to do.
- Keep a dated copy of the review section. Vendor help pages are edited in place, as the 2023 post shows. A saved copy with a date turns "we were told" into evidence.
- Define confidential the way the reviewer test defines it. If "would I be comfortable with a contractor reading this document" is false, the file does not go through a consumer assistant, whatever the NDA says.
- Scope connectors to a folder, not a drive. Where the connector allows it, point it at material that is safe to expose and keep client files and contracts outside that scope.
- Write the rule as one sentence. For example: no client documents through a consumer AI assistant. Long policies lose to a person trying to finish a task on a Friday afternoon.
- Record it in the vendor register. Not just "Gemini approved", but which connectors, which data classes, and which retention exceptions apply.
None of this is exotic. It is the review a finance team would run before sending payroll data to a new SaaS tool. The difference is that this tool arrived through a settings toggle a single employee can flip, and the terms changed after the rollout.
The contract is the help page
The uncomfortable lesson is about where the promise lives. It was never in the sales conversation. It sat in a blog post that has since grown an update note, and in a help article a user reads once, at setup, before anyone in legal has seen it.
That is a fair description of how most AI features ship now. The capability lands first, the data handling lands in a support article, and the people affected find out when a thread reaches the front page. A business that reads the support article before the rollout, instead of after the thread, is doing the one thing that actually scales.
If you are working out which AI assistants should touch which documents, talk to us about how we map it.
Sources
- Our responsible approach to building guardrails for generative AI (Google, with an update note dated 16 October 2026)
- About personalization with Connected Apps (Gemini Apps Help, Google, accessed 10 November 2026)
- gemini_files (writeup by a Gemini user; single source, not independently replicated)
- Hacker News discussion (9 November 2026)
- Google Gemini Warning: Don't Share Confidential Information (Search Engine Journal, Roger Montti, 12 February 2024)
