AI features in an inbox can feel a little like inviting a highly efficient assistant into the room, then realizing the room contains every overdue bill, launch plan, password-reset request, travel itinerary, and message you meant to answer three weeks ago. Gmail’s Gemini-powered tools make that question more immediate: when the service summarizes an email thread, drafts a response, creates a task list, or responds to a voice request, what happens to the contents of the inbox?

The short version is not a simple yes or no. Gemini needs access to relevant email content to perform an action a user requests. Google says, however, that it does not train its foundational AI models, including Gemini, on personal emails. The distinction matters. Processing data to answer a prompt is different from using that data as training material to improve an underlying model. Yet the settings and connections around Workspace and Search mean users should not treat every Google AI feature as operating under the exact same rule.

For anyone who uses Gmail for game-code receipts, Discord notifications, work correspondence, tournament logistics, or the usual avalanche of newsletters, the practical move is to understand which controls govern which behavior. Turning off Gmail smart features can limit Gemini-style assistance, but it can also remove useful non-generative tools such as spelling and grammar checks. Separately, Search-related activity and recommendation settings can affect the ways information is used once Workspace content is connected to Search services.

The central distinction: processing a request versus training a model

Gemini is a large language model, or LLM: a type of AI system designed to interpret and produce language. In Gmail, that can translate into a summary of a long thread, a suggested reply, an extracted list of follow-up items, or help organizing information. The model cannot generate a useful email summary without examining the emails relevant to the request. That access is the operational part of the feature.

Google’s position is that this access is limited to what it calls isolated tasks. In its April 2026 explanation of Gemini’s data practices, the company said Gemini processes information to fulfill a user’s particular request and does not retain that data afterward. It also said personal emails are not used to train its foundational models.

Training, in this context, means using material to help improve or build the model itself, rather than using it only to produce an answer in the moment. A person asking for a summary wants the system to inspect that conversation and return a shorter version. That does not, by itself, mean the full conversation becomes part of a broad training dataset. Google says it does not use personal Gmail messages that way for its foundational AI models.

This is a meaningful assurance, but it should not be confused with a promise that Gemini never sees email text. If a user asks Gemini to summarize an email, draft a reply based on it, or organize tasks from it, the service necessarily processes relevant content. The more precise privacy question is therefore twofold: what information is needed to perform the immediate task, and what account connections or permissions could allow related outputs to be used more broadly?

Workspace protections cover more than Gmail

Google extends its stated Workspace protection beyond email. The company says it does not use Workspace data to train or improve the generative AI and LLM systems that power Gemini, Search, and other systems outside Workspace without permission. The listed Workspace products include Chat, Calendar, Docs, Drive, Sheets, and Slides.

Related coverage includes How to Control Gemini Smart Features and Data Settings in Gmail.

That scope is important because an inbox rarely exists alone. A request about a game event might involve a Gmail confirmation, a Calendar entry, a Drive document, and a Chat message. A user assessing their comfort level with AI assistance should consider the account’s Workspace data as a connected set, not simply a pile of emails.

The key phrase is “without permission.” Google’s own guidance identifies an exception path when users connect content apps from Google Workspace to Search services. Those connections can enable Search features including AI Mode, and Google says it may use that data to improve services for everyone, including through generative-AI training.

That does not mean Search directly trains generative AI models on a Gmail inbox. Google explicitly says Search services do not do that. Its explanation instead says that, during Search interactions, summaries, excerpts, generated media, and inferences derived from relevant media, emails, and files can help answer a prompt. Google says it trains models on those summaries, excerpts, generated media, and inferences to make responses more relevant, helpful, and high-quality.

Why that wording deserves a careful read

There is a real difference between training directly on raw inbox messages and training on derived material such as summaries or inferences. But from a user’s point of view, both concern information that originated in personal content. The safest interpretation is not that one policy statement cancels the other. Instead, they describe different contexts: Gmail and Workspace assistance on one side, and Search services that users have connected to Workspace content on the other.

In plain language, asking Gmail to summarize a thread falls under Google’s stated isolated-task approach. Linking Workspace content to Search services may introduce another layer where information derived from that content can be used to improve Google services, including AI. If minimizing that possibility matters, it is worth reviewing the Search and personalization settings rather than stopping at Gmail’s smart-feature switch.

This is also why privacy controls can feel less like one big red button and more like a settings menu designed by a quest designer who loves optional objectives. Gmail controls, Workspace smart-feature controls, Search Services History, and Personalized Recommendations address related but distinct parts of the account experience.

How to turn off Gmail and Workspace smart features

Users who would rather not have Gemini-style assistance in Gmail can disable smart features. The setting applies across the account, so it should not be necessary to repeat the process in every individual Google service.

In the Gmail mobile app

  1. Open Gmail and tap the three-line menu at the upper left.
  2. Scroll down and select Settings.
  3. Choose the email account to adjust.
  4. Find Smart features and clear the selection.
  5. Find Google Workspace smart features and clear the relevant checkboxes to remove the features from Workspace apps.

In Gmail on the web

  1. Select the gear icon near the upper-right area of Gmail.
  2. Choose See all settings from the quick-settings panel.
  3. Remain on the General tab and scroll until reaching Smart features.
  4. Uncheck the option and confirm the change.
  5. Still under General, locate Google Workspace smart features.
  6. Select Manage Workspace smart feature settings, turn off both available options, then save.

It is sensible to check the page labels carefully while making changes, because the controls distinguish Gmail smart features from Workspace-wide smart features. The first is most directly tied to Gmail’s assistance functions; the latter concerns broader Workspace behavior.

The trade-off: useful tools may leave with the AI features

Disabling smart features is not a clean “remove only the intrusive bits” option. Google notes that turning them off can also disable spelling and grammar checks. That creates a practical trade-off: users may be able to reduce AI-driven inbox assistance, but they may also give up editing help they consider basic functionality.

Whether that trade is worthwhile depends on the sensitivity of the inbox and the value of the tools. A person using Gmail for low-stakes retail notices may reach a different decision than someone whose account is full of work documents, financial matters, health discussions, or confidential community administration. There is no evidence here that Gmail messages are being used to train Gemini’s foundational model; Google says the opposite. But users who prefer a smaller amount of automated analysis in their account can still opt out of features that perform that analysis.

The issue is part of a wider argument over AI features embedded by default into familiar software. Convenience is real: summaries can shorten sprawling threads, drafting can reduce repetitive writing, and task extraction can help with organization. The cost is that the feature must inspect personal material to do its job. Choosing whether that convenience is worth the access is a user decision, not a technical misunderstanding.

Concerns about how creative and personal data feeds AI systems have already spilled into other entertainment-tech debates. For example, the dispute around an alleged AI 4K treatment of a television series shows why authorization and the use of source material remain central questions even when a tool is marketed as an upgrade. That separate AI-upscaling controversy is not the same as Gmail’s account controls, but it underscores why clear consent boundaries matter.

How to review Search Services History and recommendations

Users who want to keep some Gmail or Workspace smart features but reduce the opportunity for Search-connected usage should review two additional account controls: Search Services History and Personalized Recommendations. By default, these settings draw on the account’s most recent Web & App Activity and Search Personalization settings.

To adjust Search Services History, sign into the Google account and open the Search Services History settings. From there, users can turn the setting off or delete previously saved activity. Those who leave it enabled can restrict what it saves by adjusting the Save Media setting.

For Personalized Recommendations, navigate through the Google account’s Data & privacy section to Personalization settings. The Personalized Recommendations in Search services option can be switched on or off there.

These are not replacements for Gmail’s smart-feature controls. They serve a different purpose: they govern activity and personalization associated with Search services. A sensible review process is to decide first whether Gemini assistance inside Gmail is wanted at all, then separately decide whether Workspace content should be connected to Search-oriented features and personalization.

A practical checklist before changing anything

  • Decide what you are opting out of. Disabling smart features limits assistance in Gmail and Workspace, while Search history and recommendation settings address Search-related usage.
  • Account for lost conveniences. Smart-feature changes may disable spelling and grammar checks as well as Gemini tools.
  • Check Workspace-to-Search connections. Google’s stated training exception is tied to permission and connected content apps.
  • Separate direct inbox training from derived data use. Google says Search does not train models directly on Gmail, while also describing training on summaries, excerpts, generated media, and inferences used to answer Search prompts.
  • Revisit settings when features change. AI controls and connected-service options are not interchangeable, and the names of individual features can make a difference.

Google’s current statements draw a boundary that many users will find important: personal emails are not used to train Gemini’s foundational AI models, and Gemini’s inbox access is framed as task-specific. At the same time, that boundary has adjacent doors. Smart features require content access to work, and Workspace connections to Search services introduce permissions around derived data that deserve an informed choice. The best privacy setting is not automatically “everything off”; it is the one that matches what a user is comfortable allowing each service to do.