By Olivia Hart

Google has reversed a three-year-old promise that Gmail, Docs and Drive content connected to its AI assistant would not be seen by human reviewers. The updated Gemini privacy documentation now allows human review of some information users share with Gemini, including summaries, excerpts and other material extracted from emails and Drive files.

Google says review can extend to an entire email or document. Two further changes raise the stakes: information may train Google’s AI models when “Keep Activity” is on, and human-reviewed conversations may remain stored for three years even after users delete Gemini activity.

This is more than a consent-screen revision. It changes the terms for a service that draws on personal conversations, financial information and work documents. The practical comparison is a locked records room becoming a supervised copy desk: usefulness remains, but access follows Google’s review rules.

What did Google promise three years ago?

When Google linked Bard, then its AI assistant, to Gmail, Docs and Drive, privacy formed part of the proposition. Google said: “If you choose to use 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.”

The pledge covered three specific uses: human review, selection of ads and model training. It told Workspace users that connecting those services would not transfer their content into any of those uses.

That assurance gave practitioners a clear basis for connecting an assistant to files they already manage. Email can supply the context that makes an AI response more relevant, so excluding that material from review and training carried practical weight.

“I don’t imagine anyone really wants Google—or any company—reading their email,” said Jason Aten (@JASONATEN), the tech columnist who provided expert opinion for this report. He said email can contain personal conversations, financial information and work documents.

What changes under the updated Gemini documentation?

Google’s current Gemini privacy documentation says human reviewers may inspect some information shared with Gemini. That material can come from Gmail and Drive, and can include summaries, excerpts or other information extracted from emails and files.

Google also states that reviewers may see an entire email or document. The wording therefore allows access beyond selected snippets in some cases.

Human review serves a routine quality-control function. Aten noted that companies use people to check whether chatbots understand requests and deliver accurate responses.

If an assistant uses email in those checks, the email becomes part of the review unless the provider deliberately excludes it. Google’s original pledge made that exclusion explicit; the updated documentation does not.

Three years after the Bard announcement, Aten described the change as Google quietly rewriting the rules. For users, the former assurance now depends on the scope of Gemini review rather than the launch promise.

Can Gemini use your information for training?

The documentation also expands model-training exposure. If users leave the “Keep Activity” setting on, Google may use the same information for training its AI models.

That changes the training side of the Bard pledge, which excluded Workspace content from training. At launch, choosing the Workspace extensions meant content would not train Bard or influence ads; now, an enabled activity setting can permit training.

For practitioners deciding what Gmail and Drive access to grant, “Keep Activity” is now an explicit control rather than a housekeeping preference. It determines whether the same material can contribute to model development as well as answer requests.

Google’s updated documentation makes the training condition plain. Users reviewing permissions should start with that setting and then assess the separate risk of human review.

Why does three-year retention matter?

Google may retain conversations selected for human review for up to three years. That retention applies even when users delete their Gemini activity.

Deletion therefore no longer guarantees full removal for every exchange. For material drawn from email and files, the stakes include personal conversations, financial information and work documents.

“Up to three years” sets the stated ceiling, but the trigger is participation in human review rather than an explicit retention choice. Users who assume deletion controls every copy of a reviewed exchange are reading the wrong rule.

Human-reviewed chats are different from ordinary activity records. Google’s documentation gives them a longer life span that can outlive the user’s deletion command.

The strategic change is straightforward. Google has moved from excluding connected content from human review and Bard training to permitting some review, and model training when Keep Activity is enabled.

For practitioners, the relevant questions are concrete:

• What Gemini extracts from Gmail and Drive, including whether a full email or file can enter review.
• Whether “Keep Activity” remains enabled and allows the information to support model training.
• Whether a conversation falls within the three-year human-review retention rule.

Connected email still offers the context Aten called central to an AI assistant’s usefulness. The difference is that Google now treats some of that context as material it may inspect, use for training and retain.

Companies should therefore evaluate Gemini privacy settings against the old promise. They should not rely on Bard-era wording as a current safeguard.