Gmail privacy depends on which Gemini route you use
Smart features, connected apps, activity retention, and managed Workspace protections follow different rules. Current documentation needs a more precise reading than a blanket reassurance.

"Gmail uses AI" is too broad to answer a privacy question. Sorting a message, finding an email for a Gemini response, retaining that response, and using an interaction to improve a model are different operations. The account and product determine which rules apply.
An old reassurance can become misleading when it skips those distinctions. As of 14 November 2026, Google's consumer Gemini documentation contains conditions that a blanket statement about an untouched inbox would leave out.
The useful check follows the data through the feature you actually use. Start with the account, identify the product, then read the controls for that specific route.
Gmail has more than one smart-feature control
Google's current smart-feature documentation separates controls for Gmail, Chat and Meet, for features across Workspace, and for features in other Google products. A setting that allows a flight email to inform another product is different from a setting for a suggestion inside Gmail.
The page also describes additional controls for some experiences. Turning off one smart-feature option should not be interpreted as a complete inventory of every permission or activity setting attached to the account.
Before changing anything, write down the outcome you want. You may want to stop a particular cross-product lookup, disable an unwanted suggestion, or prevent a work document from entering a consumer chat. Those intentions lead to different checks.
Avoid changing several controls at once if you need to understand their effects. Note which feature stopped working and which remained available. An interface change provides evidence about that feature, not a complete audit of backend data handling.
Consumer Gemini has its own activity rules
Google's Connected Apps personalisation page says Gemini does not train directly on an entire Gmail inbox or the other listed source libraries. It also says Gemini Apps activity, including data from connected apps, can be used for model training when Keep Activity is on.
The page explains that responses can use summaries, excerpts, generated media, and inferences from relevant source material. Those can enter training under the described conditions. For a brief or particularly relevant message, a summary may contain the message itself. Disconnecting a source app does not delete information already in Gemini Apps activity.
That distinction rules out two easy assurances: that connecting an inbox necessarily trains on every message, and that email-derived content can never enter a training interaction. The documented route depends on what is retrieved, the interaction, and the relevant activity control.
Read the current deletion and retention explanations before assuming that a changed switch erases earlier activity. Stopping future access and deleting retained records are different requests. A privacy decision needs both questions answered when past interactions contain sensitive information.
Work accounts require a separate reading
The Google Workspace generative-AI privacy hub describes protections for Workspace services, including limits on model training outside the organisation's domain without permission. It also describes administrator controls and retention. Those terms should be read for the service and account in use.
A company's Workspace commitment cannot automatically be transferred to a personal Gemini session. Conversely, a consumer activity explanation does not establish how a managed Workspace interaction is handled. The same email address on a screen is insufficient evidence about the applicable service terms.
For organisational data, ask the administrator which service is approved and what controls apply. Include whether past conversations can be deleted by the user and whether retention is centrally managed. This is a concrete configuration question, not a reason to accept a reassuring screenshot from another product.
Check the complete sharing route
Even a disconnected app does not prevent someone from manually pasting a message into a chat or uploading a screenshot. Review those intentional sharing actions alongside background connections. The content has crossed a boundary because of the action, regardless of which inbox toggle was selected.
Record the product, account type, connection, activity setting, and date of the documentation used for the decision. Recheck when a new personalisation feature is enabled. That record gives a later reviewer something to compare with a changed policy. A broad claim that "Gmail is safe" cannot do the same work.