Google's Gemini Help Page Says Human Reviewers See Drive and Gmail Excerpts

Google's Gemini help page says a subset of data from connected apps is reviewed by human reviewers, including "interactions, summaries, excerpts and inferences from your relevant emails and files." The page is Google Help answer 16836988, and it was the centre of a Hacker News thread on Monday, November 9, 2026, that reached 238 points.

The thread was prompted by a GitHub repository titled "Documentation of Gemini File Leaking Behavior", posted by a user named thisisme6598. The repository had 6 stars, 1 fork and 3 commits when read.

Google has not responded on either page, and neither page shows a statement from the company about the claim.

What Google's page says

The help page says reviewers may include trained service providers and that review happens for debugging, safety and improvement. Before review, Google said it takes steps such as "disconnecting interactions from your Google Account and reducing personal information in this data."

On Gmail specifically, the page says "human reviewers may also review some of your Gmail inbox" in limited cases, such as when a user submits feedback or when abuse or harm must be addressed.

The page warns users not to connect apps to Gemini Apps if they contain confidential information they would not want a reviewer to see. It gives no retention period and links to Google's retention policies instead.

Deleting Gemini Apps activity will not stop use for product improvement if the data has already been reviewed, according to the page.

What the GitHub post claims

The repository's author says Google reversed a 2023 promise. A 2023 Google blog post about Workspace extensions said personal content from Gmail, Docs and Drive "is not seen by human reviewers," the author wrote, and says an October 16 edit to that post and the help page now confirm human review. The repository gives no year for the edit.

The author also claims retrieval happens in background conversation, without an @Drive or @Gmail tag, and that the amount captured is probabilistic. They list four outcomes: no retrieval, a generic summary, a partial extraction of two to four sentences, or full-file ingestion.

Claim	Google's help page	Repository evidence
Human review of connected-app content	Yes, on the page	Quotes the page
Retrieval without @ tags	Not addressed	Author's own testing
Full files reach reviewers	Page says excerpts	No logs shown

The evidence is the author's informal testing. The repository includes no logs, screenshots or independent replication, and does not reproduce the help page text.

What is not established

The help page says reviewers see excerpts, summaries and inferences. It does not say whole files or whole inboxes are routed to annotators. The repository's headline claim about files goes further than the page it cites.

The repository also argues that Google's de-identification, described as regex and named-entity matching, cannot scrub narrative content such as account and confirmation numbers. That is the author's reading of Google's Sensitive Data Protection documentation, and it was not tested here.

The privacy argument around connected assistants is not unique to Google. The Terminal has covered a mod that strips an email address from Claude Code's user agent, and separately reported on OpenAI's Codex work with Pierre's Git tooling.

No date has been announced for a Google response. The next check is whether the help page text changes after Monday's thread.