Google warns human reviewers may see your Gemini Connected Apps emails

Google has confirmed through its official Gemini help pages that human reviewers may see portions of your emails and files when you connect apps to the Gemini AI assistant. The disclosure sits in Google's Gemini Connected Apps help documentation, alongside a plain warning: do not link apps that hold confidential information you would not want a reviewer to read.

The feature at issue is called Connected Apps, part of a Google capability known as Personal Intelligence. Eligible users can link certain Google services to Gemini so the assistant can draw on personal data for more tailored responses. Once linked, data from those apps, along with other information from the Google Account, is applied to personalizing Gemini, carrying out tasks on the user's behalf, and improving Google services, a category that includes training AI models.

What Connected Apps gives Gemini access to

Linking an app gives Gemini access to a wide range of personal data. That includes emails, files, calendar events, contact details, photos, videos, payment information from Google Wallet, location data, and insights derived from all of it. Google notes the data can involve sensitive topics such as race, religion, and health, as well as confidential material.

Users can manage which apps are linked through the Personal Intelligence section of their Google Account settings. Gemini remains usable without any connected apps; the personalization features simply will not apply. Connecting or disconnecting apps does not automatically change other Google personalization settings that are already active on the account.

When human reviewers may see emails and files from connected apps

Once apps are linked, Google states that a portion of the data Gemini processes from those apps is reviewed by human reviewers, among them trained staff from outside service providers. The goal is to identify problems, keep the service safe, and improve it. Reviewers judge whether Gemini responses were poor, inaccurate, or harmful, and suggest better answers to feed back into the models.

The reviewed material can include interactions, summaries, excerpts, and inferences drawn from a user's relevant emails and files. Where an email or file is brief or closely tied to a prompt, the summary passed to reviewers may amount to the whole email or file. If a user submits feedback, or if Google needs to respond to abuse or harm, reviewers may also look at content from a user's Gmail inbox or Photos library directly.

"Human reviewers improve Google services, including generative AI," Google says in its Connected Apps help page. The company frames the practice as routine quality control. Even so, the same help pages acknowledge that this process can reach content users consider private.

How AI training differs from human review

Google draws a line between human review and direct AI training. The company says its training pipeline does not draw directly from a user's raw Gmail inbox, Drive, Contacts, Calendar, other Workspace apps, Google Wallet data, or media and audio stored in a Google Photos library.

What goes into the training pipeline is different. When Keep Activity is turned on, Google may train models on summaries, excerpts, generated media, and inferences from relevant emails, files, media, and transactions, as well as on prompts, responses, and information about how those responses were built. So while the original inbox stays out of direct model training, a summary derived from it may not.

Retention, privacy steps, and what the Privacy Hub says

Before outside service providers review interactions that drew on Connected Apps data, Google says it takes steps such as separating the interaction from a user's Google Account and cutting down the personal information it contains. Those steps are set aside when a review is needed to address abuse or harm.

The Gemini Apps Privacy Hub adds an important detail on how long reviewed data sticks around. Chats that have been reviewed by human reviewers, along with related data such as the user's language, device type, location information, or submitted feedback, are not erased when a user deletes their Gemini Apps activity. Instead, that data is kept for up to three years.

Steps users can take to limit exposure

Several controls exist to reduce how much data reaches reviewers. Users can review and change which apps are linked through Personal Intelligence settings, delete their Gemini Apps activity at any point, and switch off Keep Activity to prevent future chats from being used to train Google's AI. Disabling Keep Activity does not undo reviews that have already taken place.

Removing a connected app, or deleting data inside one, does not wipe data that has already been saved to Gemini Apps Activity. That data can still support product improvement, including AI model training, until a user deletes it directly from Gemini Apps Activity, and only then if trained reviewers have not already processed it.

Google's own guidance is straightforward. The company warns users plainly not to connect apps that contain information they would not want a reviewer to see. For anyone handling sensitive work through Gmail, Drive, or other linked services, that caution is worth taking seriously before opting in to the feature.

What the documentation leaves unanswered

Several questions go unaddressed in Google's published pages. There is no figure for how large the reviewed portion of Connected Apps data actually is. There is no breakdown of how reviewers are selected or trained beyond the note that some are outside contractors. The only retention figure in the documentation is the three-year cap for reviewed chats noted in the Privacy Hub.

It is also unclear whether enterprise or Workspace contracts change these defaults for all account types. Unverified online discussions have raised the possibility of a separate data pipeline sending additional files or emails to annotators beyond what the Connected Apps policy covers, but those claims have not been confirmed by any major news outlet and should be treated as unverified.

What the official pages do establish is clear: connecting Google apps to Gemini means a portion of what Gemini processes, including content drawn from emails and files, may reach human reviewers. For anyone handling confidential material through connected services, that is a fact worth knowing before turning the feature on.