ChatGPT, Claude and Gemini Can Read Your Inbox: What You Grant and How to Revoke It

Linking your mailbox or cloud drive to a chatbot is one of the most practical things you can do with generative AI, and also one of the riskiest. The moment you press “Connect,” you hand an assistant permission to search and read your messages and documents, and in many cases to draft, send emails or move files on your behalf. ChatGPT and Claude do this through connectors authorized with your Google or Microsoft account, while Gemini relies on its Google Workspace app. Disconnecting inside the chatbot is not always enough: to cut access completely, you also need to review permissions in your Google or Microsoft account.

  • Broader than you think: Authorization screens typically request full mailbox access even when you only want a thread summarized.
  • Write access is common: Several connectors can send, reply, forward, label or move content, not just read it.
  • Two-step revocation: You must disconnect the app in the chatbot and then remove its access in your Google or Microsoft account.

What you actually authorize when you press Connect

The permission you grant through a Google or Microsoft screen is usually wider than the task you have in mind. You may want a summary of one thread, but you concede access to the entire inbox. OpenAI’s help documentation for its Outlook app lays this out plainly: the Microsoft permission list for mail includes reading and modifying messages (Mail.Read, Mail.ReadWrite), sending them (Mail.Send), reading and editing contacts, and maintaining access while you are offline (offline_access). For calendar, it requests the ability to create and modify events. OpenAI notes that granting these permissions does not by itself activate the actions; in organizations, an administrator enables each one separately.

Anthropic issues a similar warning on its Google Workspace connectors page: Google’s authorization screen mentions permissions to send email. It adds a useful guarantee: Claude only accesses the connected account and respects the permissions you already have in Google, so it cannot see what you could not see yourself.

What each assistant can see and do in your name

The differences between services are not minor. ChatGPT offers the finest-grained control. According to its app permissions page, you can choose among four levels: always ask, allow reads, allow low-risk actions, or allow everything. OpenAI warns that the last option carries elevated risk because actions execute without further confirmation.

Claude, for Gmail, Calendar and Drive, can search and read emails and documents, though for Gmail attachments it only sees metadata. It can draft, label, send, reply and forward; create events; and upload files. By default, every action taken in your name requires approval, except in Auto mode. With a personal Microsoft 365 account, Claude covers Outlook and OneDrive, and its writing tools come enabled; in chat it asks before sending, creating, changing, moving or deleting, except in automatic approval modes, and each tool can be adjusted in its permissions.

Gemini, across Gmail, Docs, Drive, Calendar, Keep and Tasks, summarizes and searches your emails and documents but does not see comments or images in Docs or Gmail. It can work with tasks, notes and events, but does not manage Drive, so it cannot create folders or move files. Google’s help pages do not detail its confirmation behavior. Gemini also requires history to be enabled and, for Gmail, the smart features setting across other Google products.

The risk is not theoretical. A chatbot that reads your email also reads messages from strangers, and a single message can carry hidden instructions for it. This is known as prompt injection. With permission to send without confirmation, the door stays open.

Do your emails and documents train the AI?

The three companies give different answers, and OpenAI’s has two layers. Its general connected-apps page states that on Free, Go, Plus and Pro plans it may use information obtained from apps to train its models if the “Improve the model for everyone” setting is on; on Business, Enterprise and Edu it does not do so by default. But its Google-specific FAQ says it does not train its general models on data taken directly from Gmail, Drive or Calendar. The exceptions are what you submit as feedback, what you copy or paste into the chat, and what appears in responses. If you turn the setting off, none of that is used.

Anthropic says it does not train on data from its Gmail, Drive and Calendar connectors, although if you have accepted training, content you copy or responses that include it may be used. Retrieved data is stored with the chat and deleted when the chat is deleted.

Google is the broadest. Its Gemini privacy notice, updated on November 16, 2026, says Gemini stores and uses information from connected apps to provide and improve the service, that human reviewers examine some of the data, and that content from connected apps accompanying feedback is reviewed by specialized teams. From history settings you can prevent future chats from being used for training. One common detail: disconnecting does not erase what is already inside. OpenAI clarifies that disconnecting Google does not delete saved chats or memories.

How to revoke access: first in the chatbot, then in your account

These are two separate places, and you should go through both:

  1. ChatGPT: Settings > Apps (or Plugins, depending on the version) > choose the app > disconnect the account. OpenAI warns that changing a permission does not revoke access already granted and that the provider may have its own controls.
  2. Claude: Customize > Connectors > select the connector > Disconnect. With a personal Microsoft account, Anthropic warns that the permission remains in your Microsoft account until you remove it there.
  3. Gemini: on the connected apps page at gemini.google.com/apps.
  4. Your Google account: in Linked apps (myaccount.google.com/linkedapps), filter by access to your Google account and remove ChatGPT or Claude. According to Google, from that point the app can no longer access your account.
  5. Microsoft work account: in the My Apps portal you can revoke permissions you granted yourself; those approved by an administrator can only be removed by that administrator.

It is the same habit as reviewing browser extension permissions and cleaning out the ones you no longer need.

My reading is that connecting email to a chatbot is among the most useful things you can do with AI and, at the same time, the one that opens the most risk surface. The sensible approach is to start in read-only mode, leave confirmations enabled, and connect only the account you need rather than your primary one out of inertia. Companies should request send permissions only when the user activates that action, not in advance on the Google or Microsoft screen. Until that changes, periodic review of linked apps is on you.