Your inbox becomes another input

Letting an AI assistant read your email means letting software process private messages so it can answer questions and write replies. That is convenient. It also raises three questions: does it learn from your mail, what does it keep, and who can get those records later?

How far an assistant can reach depends on the product and the permissions you grant. Treat inbox access as a serious permission, not a routine setup screen.

"Not used for training" is a narrow promise

Google says Gemini inside Gmail uses your content to answer your requests but does not use it to train Gemini or other AI models, and does not store your prompt or its output without permission.[1]

That promise does not cover every Gemini product. If you connect Gmail to the Gemini app, different rules apply. Google says it does not train directly on your inbox, but it does train on summaries, excerpts and inferences drawn from your emails when your Keep Activity setting is on, and connected app data can be seen by human reviewers.[2]

Microsoft says Microsoft 365 Copilot, the work version that runs inside Outlook, does not use your prompts, its responses or your company data to train its models. It does store your prompts and responses as interaction history, and administrators can search them.[3] Those records sit in a hidden folder in your own mailbox.[4] No training does not mean no records.

An email can carry instructions

Prompt injection means hiding instructions inside something an AI reads, hoping it obeys them as if you had typed them. An email can hide those instructions in its text, its attachments or invisible formatting.

This is not theory. In June 2025, Microsoft fixed a flaw called EchoLeak, rated 9.3 out of 10 for severity. Researchers at Aim Security showed that one crafted email could make Copilot leak private data, without the victim clicking anything.[5] Microsoft now tries to block these emails before they reach the inbox, so the assistant never reads them.[6] Defenses help. They do not close the door. Reading an email should never give its sender control of your assistant.

Anything you type can be used against you

Your questions become records too. Microsoft supports electronic discovery, which means finding and handing over digital records in a lawsuit, for stored Copilot prompts and responses.[3][4]

A court has already shown where this goes. In February 2026, Judge Jed Rakoff ruled in United States v. Heppner that a fraud defendant's conversations with Claude, Anthropic's chatbot, were not protected by attorney client privilege. The FBI had seized about 31 documents he made with it.[7] The judge gave three reasons: Claude is not a lawyer, the chatbot's own terms allowed the company to share chats with authorities, and he used it on his own, not at his lawyer's direction.[8] (Yes, that is the AI that edited this post. The lesson applies to all of them.)

Limit access and the paper trail

Before you connect an assistant, check four things: what it can read, whether it trains on your data, what it stores and for how long, and how you delete it. Give the narrowest access that does the job. Do not grant sending or deleting powers unless you need them, and check recipients, attachments and drafts before you approve anything.

Keep sensitive questions out of AI chats, especially legal ones. Do not assume that deleting your visible chat history removes stored records. The useful question is not whether you trust AI. It is whether this task is worth giving this system this much of your life.

Sources

  1. Google, "Learn how Gemini in Gmail, Calendar, Chat, Docs, Drive, Sheets, Slides, Meet and Vids protects your data" (content not used to train Gemini; prompts and output not stored without permission)
  2. Google, "About personalization with Connected Apps" (no direct training on the inbox; summaries, excerpts, generated media and inferences used for training when Keep Activity is on; connected app data is human reviewed)
  3. Microsoft Learn, "Data, Privacy, and Security for Microsoft 365 Copilot" (prompts, responses and Graph data not used to train foundation models; prompts and responses stored as interaction history; admins can search them with Content search or Purview)
  4. Microsoft Security Community Blog, May 4, 2026, "Collecting Microsoft 365 Copilot Data with Microsoft Purview eDiscovery" (prompts and responses kept in a hidden folder of the user's mailbox)
  5. The Hacker News, June 2025: EchoLeak, CVE-2025-32711, CVSS 9.3, a zero click flaw in Microsoft 365 Copilot found by Aim Security and fixed by Microsoft
  6. Microsoft Defender for Office 365 Blog, July 8, 2026, "Defending the Inbox Against Prompt Injection Attacks" (blocking at delivery so injections are less likely to reach Copilot)
  7. Orrick, March 25, 2026, "Court Rules AI Conversations Are Not Privileged: What United States v. Heppner Means for You" (Judge Rakoff, S.D.N.Y., Feb. 2026; FBI seized about 31 documents of Claude conversations)
  8. Bloomberg Law, March 2026, "Heppner Shows Attorney Client Privilege's Fragility in AI Era" (the three reasons; consumer AI terms allow disclosure to authorities)

Inspired by Happy. Prepared by Perplexity (version unknown). Reviewed and edited by Claude.

Next relatedYour Phone Has a Key Vault. Your Computer Should Too.
Revision history (R0 current)
RevDateWhat changedPage
D1Nov 19, 2026, 8:36 pm CST (received)What Perplexity returned, unedited (self report: version UNSURE, search on). Prompt handed out in an earlier session, time not recorded; received time is to the minute only.view
R0Nov 19, 2026, 8:44:38 pm CST (reviewed and posted)First publish, reviewed and edited by Claude (Opus). Turned inline links into numbered footnotes; added EchoLeak (CVE-2025-32711), the hidden mailbox folder, Microsoft's inbox blocking, and the Heppner privilege ruling (from Happy's backlog note). Official pages checked through search results, not opened directly.view