MorseSign inStart freeOpen Morse

Is it safe to connect Gmail to ChatGPT? What five AI connectors can read, send, and keep

The Morse teamUpdated 10 min read

Connecting Gmail to ChatGPT is reasonably safe for an everyday inbox as of November 13, 2026, but the permission you approve covers sending mail as well as reading it. OpenAI's help page lists the Gmail scope ChatGPT requests as gmail.modify (OpenAI, archived November 3, 2026), which Google describes as access to "Read, compose, and send emails from your Gmail account" (Google). OpenAI doesn't train on what ChatGPT reads from Gmail, but whatever it pulls into a reply is stored with your chats, can become a memory, and is open to prompt injection from any email a stranger sends you.

We read the connector help pages of OpenAI, Anthropic, Google, Microsoft, and Perplexity, plus Google's Gmail scope list and its developer data policy, on November 13, 2026. OpenAI's help center and some of Perplexity's pages block automated reads, so for those we read the latest Internet Archive copies and link to them with their dates. OpenAI deserves credit as the only one of the five whose connector pages list the exact Google OAuth scopes it requests.

What each AI connector can do with Gmail and Drive, as of November 13, 2026

Each cell comes from the company's own page, read on November 13, 2026, or the archived copy linked.

ConnectorAccess it asks forTrains on it?How long it's kept
ChatGPTGmail: gmail.modify. Calendar: calendar.events. Drive: drive.readonly, plus full drive for actions that "create, update, share, move, upload, copy, or delete" files (OpenAI, archived)Not on data "directly from connected Google apps," except feedback you send, content you paste, and what appears in a reply (OpenAI, archived)With the chat. Deleted chats are "scheduled for permanent deletion within 30 days," and Memory can draw on what it reads (OpenAI, archived)
ClaudeGmail: search, read, and draft, plus send, reply, and forward with your approval by default. Drive: read, upload, share, move, and trash (Anthropic)Not on connector data. With training on, pasted content and replies that quote it "may be used to improve our models" (Anthropic)Kept with the chat it was used in. Deleted chats leave back-end storage within 30 days; with training on, data may be kept de-identified for up to 5 years (Anthropic)
Gemini appGmail, Calendar, Drive, Docs, Sheets, Slides, Keep, Tasks, Chat, and Meet, including data added "in the future" (Google)Not directly on your inbox, but on "summaries, excerpts, generated media, and inferences" drawn from it if your Chat content setting is on, and reviewers see a subset (Google)Gemini History auto-deletes after 18 months by default; reviewed chats are kept up to three years (Google)
Microsoft CopilotGmail, Google Drive, Calendar, and Contacts through a Google sign-in, plus OneDrive and Outlook.com; you choose "what data Microsoft Copilot can access" when you authorize (Microsoft)"Microsoft do not use data from connected services to train Copilot" (Microsoft)Conversation activity is stored for 18 months by default (Microsoft)
PerplexityGmail: "Read, compose, and send emails." Calendar: "View and edit events on all your calendars." Contacts, including "your organization's GSuite directory" (Perplexity, archived July 22, 2026). Drive: "See and download all your Google Drive files" (Perplexity, archived October 26, 2026)Gmail and Calendar: the connector page promises "never used for AI training" for Enterprise data only, and the privacy notice says Email Assistant doesn't use "the content of emails" to train models (Perplexity). Pro file connectors: Drive and Dropbox files "will never be used to train models" (Perplexity)Synced Drive files are removed when you delete them in Drive. (Perplexity, archived)

Where each app keeps its disconnect button, and what disconnecting leaves behind, is in our guide to disconnecting Gmail and Google Drive from ChatGPT, Claude, Gemini, Copilot, and Perplexity.

Can ChatGPT send emails from your Gmail?

The scope you grant allows it. Google offers a narrower gmail.readonly scope that only lets an app "View your email messages and settings," but OpenAI lists gmail.modify for ChatGPT's Gmail app. Google classes both as restricted scopes, and says an app that stores or transmits restricted scope data on its servers "must go through a security assessment" (Google).

ChatGPT's app permissions set what it does with that access. OpenAI's options run from Always ask to Allow all actions, and it warns that the last one "carries elevated risk because supported actions may run without another confirmation." Those settings change only whether ChatGPT asks first. They "do not grant an app new access," and removing access means disconnecting the app (OpenAI, archived November 8, 2026). A Gmail-triggered task can also act on new mail with that same access, which we cover in are ChatGPT scheduled tasks private.

Can the Claude Gmail connector send email?

Anthropic's own pages disagree. Its help center article, updated November 13, 2026, says Claude can "Send, reply to, and forward emails from Gmail," asks for your approval before each one by default, and that "Google's OAuth screen mentions email sending permissions" (Anthropic). Its Claude Docs page for the same connector says "Claude reads and searches email only; it can't create, send, or modify messages" (Anthropic).

Go by the Google consent screen, since that's the grant Claude holds. If it mentions sending, treat the connector as able to send. On Team and Enterprise plans, Anthropic says owners decide whether members can let send actions run "without asking each time."

Do ChatGPT, Claude, and Copilot train on your Gmail?

They say no, and Google's own rules for outside apps forbid it. Apps that use Google Workspace APIs may not use the data "to create, train, or improve a machine learning or artificial intelligence model beyond that specific user's personalized model" (Google). OpenAI, Anthropic, and Microsoft each repeat a version of that promise on their connector pages.

The promise covers only data read through the connector. OpenAI carves out conversations you rate, content you paste, and anything "included in ChatGPT's response." Anthropic's carve-out, on consumer plans with training on, is pasted content and Claude's responses "which include specific information from these connectors." So with training on, the reply that quotes your email falls on the trainable side of both lines.

Google's own Gemini app isn't an outside app and works differently. Google says Gemini "does not train generative AI models directly on your Gmail inbox" (Google), but can train on the summaries, excerpts, and inferences it draws from your mail when your Chat content setting is on, and that a summary "may just be the email" when it's short. On gemini.google.com and iPhone, Connected Apps aren't available at all while History is off (Google). Google's Connected Apps page still says that activity is used to improve Google services "including by training generative AI models, if your Keep Activity setting is on" (Google), while its newer history page says the settings that let it train generative AI on your chats "are off by default" (Google). Google's page doesn't say whether accounts that had Keep Activity on before were switched over, so check Settings, then History. Google's advice is plain: "Please don't connect your apps to Gemini Apps if they contain confidential information that you wouldn't want a reviewer to see." Our guide to whether Gemini is private covers its reviewers in more depth.

Does a connected AI keep a copy of your email or files?

Mostly not of the whole mailbox, but Drive is where copies start. Perplexity's standard Drive search queries Google "in real time" and "does not index your entire drive," while its high-precision search indexes the files you pick and keeps them searchable in Perplexity (Perplexity, archived). A Drive file added to a private Claude project becomes project knowledge, and its content "refreshes from Drive periodically" (Anthropic).

What every app does keep is the chat that quotes your mail, plus any memories drawn from it and, on Gemini, your Gemini History. Our guide to what each AI keeps after you disconnect it covers those chats, memories, and Activity records app by app.

Can a connected AI be tricked into leaking your email?

Yes, and the companies' own pages say so. Google's Gemini help explains that attackers can share "malicious content with a user, who then unintentionally references this malicious content in prompts," and gives the example of a summary request where one email is malicious (Google). Google's developer policy requires apps holding restricted scopes to protect "against prompt injection techniques" (Google). Anthropic tells you that connecting means "granting Claude permission to access and potentially modify data," and to "Only connect services you trust and need for your workflows" (Anthropic).

Microsoft's connectors page doesn't mention the risk. Varonis reported a flaw it called CoSnitch that could pull Gmail, Calendar, and Drive data out of Copilot's connectors, which Microsoft fixed on August 18, 2026 under CVE-2026-24301 (Varonis). We cover that flaw and others, including ShadowLeak against ChatGPT's Gmail access, in what is prompt injection.

Gemini's Connected Apps also reach beyond Google. Google says it "does not monitor or secure data from custom third-party Connected Apps," and that connecting them "may expose your data, passwords, devices, and accounts to unauthorized access" (Google). On November 10, 2026, Adobe and Google announced that Adobe will expand its connectors in Gemini with Acrobat and more creative tools (Adobe).

Should you connect a work Gmail or Google Drive to AI?

Only if your admin has approved it. OpenAI says Google Workspace admins decide which OAuth scopes the ChatGPT app may request, and that Enterprise owners can allow or block entire shared drives (OpenAI, archived). In ChatGPT Business, "many apps are enabled by default" (OpenAI, archived). On Claude's Team and Enterprise plans an Owner must turn the connectors on first, and a Google Workspace admin may need to mark Claude as trusted under Security, Access and data control, API controls (Anthropic).

Shared spaces widen who can read it. Perplexity says anyone with access to a shared Project "will be able to search those files" synced to it from Drive (Perplexity, archived), and anyone in a shared session sees the calendar answers (Perplexity, archived).

How do you revoke an AI's access to your Google account?

Remove the app on your Google Account's linked apps page, knowing that this doesn't delete the chats that already quote your mail, and follow our guide to disconnecting AI apps from Gmail and Drive for each app's own steps.

How to ask AI about one email instead of your whole inbox

The safest grant is none. Every risk above comes from a standing grant, a token that lets an AI company's servers read, and often send, from your whole mailbox whenever a prompt, a task, or a stranger's email asks, with what it reads saved in chat history on those servers.

On Morse, paste in only the part of the email you need help with. The chat's history is stored in your browser, encrypted, not on our servers. Our servers don't keep your messages or log what you write. Morse doesn't train AI on your content, sell your data, or show ads, so there's no carve-out for pasted text to read about.

Pick a mode for the chat:

  • Private, for drafting a reply you'd rather no host keep. New chats start here. Open-weight models like DeepSeek, Kimi, and GLM, plus xAI's Grok, answer with your account, email, and IP address stripped. Morse has zero data retention agreements with its inference, search, and other providers, and where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
  • E2EE, for pasted passages of a contract, a medical letter, or a thread no one else should read. Your browser checks the secure enclave's hardware attestation before anything leaves your device, then encrypts the message so only that enclave can read it, and our servers relay ciphertext they can't read. It rests on the enclave hardware and attestation working as designed.
  • Anonymous, to ask GPT, Claude, or Gemini about an email without your account behind it. Our proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider sees the text without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy, so take names and account numbers out first.

The history locks when you sign out, and because we don't store chats, we usually have very little to give authorities. Clearing your browser deletes those chats for good and we can't recover them, so paste the finished draft back into Gmail. Our privacy policy lists what the account itself holds.

Which connector settings should you change?

  • Sending: on ChatGPT, keep the Gmail app on Always ask. On Claude, keep send, reply, and forward on approval.
  • Training: turn off "Improve the model for everyone" in ChatGPT and model training in Claude, since quoted replies fall outside their connector promises. On Morse there's no switch to find.
  • Copies: delete chats that quote sensitive mail, clear related memories, and remove Perplexity's synced Drive files when you disconnect.
  • Prompt injection: connect only the accounts you need, and turn a connector off in chats that don't need it.
  • Revoking: remove the app on Google's linked apps page, then delete what it left in your chats.

Try a chat from the Morse home page without an account, or start free with 15 messages a day and no card. E2EE chats and unfiltered models come with Pro at $19 a month, and Max is $49 a month.

Common questions

Will ChatGPT watch my inbox in the background?

Only if you set up an event-triggered task. OpenAI says eligible users can create a task in Work that runs on new messages in a connected Gmail account, and that disconnecting Gmail removes the connection and stops "related mailbox monitoring."

Can my employer see what I do with Gmail connected to ChatGPT?

On a work account, possibly. OpenAI says eligible Enterprise and Edu workspaces can review supported app and conversation events through its Compliance Platform, and Google Workspace admins decide which OAuth scopes the ChatGPT app may request.

What happens to a Drive file in an AI chat if I lose access to it?

It depends on the app. Anthropic says Claude removes the document preview but keeps the conversation history, while Perplexity says a file you lose access to in Google Drive is "immediately removed from Perplexity."

Ask what you actually want to ask.

80+ models, characters, and images, with every chat encrypted and kept on your device. Free to start.

Start a private chat

Read next