D1: what Perplexity returned
Unedited, before any fact check. Kept for the record.
This is D1, the raw draft from Perplexity, exactly as returned. It has errors that were fixed in the final article. Read the final article.
MODEL: Perplexity by Perplexity AI. Exact model name and version, UNSURE. Search mode on. SLUG: ai-email-privacy-risks TITLE: The Privacy Risks of Letting AI Read Your Email META_DESC: Letting AI read your email can expose private messages and create new records. Learn what Google and Microsoft promise and how to limit the risks.
Your inbox becomes another input
Giving an artificial intelligence assistant access to your email means letting software process private messages to answer questions and generate responses. [support.google](https://support.google.com/mail/answer/14615114?hl=en) That convenience deserves scrutiny. The issue is not just whether your messages teach a model. It is also what gets processed, what gets saved, and who can obtain the resulting records. [support.google](https://support.google.com/mail/answer/14615114?hl=en)
An assistant cannot necessarily read everything or take every action. Its reach depends on the product and its permissions. [learn.microsoft](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy) But connecting an outside app adds another service whose access and data policies you need to understand. [support.google](https://support.google.com/gemini/answer/16836988?hl=en) Treat inbox access as a serious permission, not a routine setup screen.
The privacy promise depends on the product
Google says Gemini inside Gmail uses your content to answer requests but does not use it to train its generative models. Its consumer guidance also says prompts and generated output are not stored without permission. [support.google](https://support.google.com/mail/answer/14615114?hl=en)
Do not extend that promise to every Gemini experience. Google says personal Gemini Apps with connected services follow different rules. With Keep Activity enabled, interactions can improve its models. Although Google says it does not train directly on your inbox, summaries, excerpts, and inferences drawn from connected emails can be used for training. [support.google](https://support.google.com/gemini/answer/16836988?hl=en)
Microsoft says its organizational Microsoft 365 Copilot, including email assistance in Outlook, does not use prompts, responses, or retrieved organizational data to train its underlying models. It does store interaction history, including prompts and responses. Administrators can manage that material through compliance tools. [learn.microsoft](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy) No training does not mean no records.
An email can carry instructions
Prompt injection means placing instructions inside material an AI reads, hoping it mistakes that material for commands. An email can hide those instructions in its body, attachments, or invisible formatting. [learn.microsoft](https://learn.microsoft.com/en-us/defender-office-365/step-by-step-guides/prompt-injection-protection-defender-for-office-365)
For example, a malicious message could tell an assistant to reveal other mailbox content instead of summarizing the message. Microsoft warns that successful attacks can leak sensitive information or make dangerous messages appear safe. Providers build defenses, but the attack surface remains. [learn.microsoft](https://learn.microsoft.com/en-us/defender-office-365/step-by-step-guides/prompt-injection-protection-defender-for-office-365) Reading an email should not grant its sender authority over your assistant.
Limit access and the paper trail
Your questions can become records too. Microsoft supports electronic discovery, finding and producing digital information for legal cases, for stored Copilot prompts and responses. Organizations can preserve, search, and export them. [learn.microsoft](https://learn.microsoft.com/en-us/purview/ai-m365-copilot) That does not make every conversation public or automatically admissible. It means an AI chat is not a private escape from ordinary records.
Before connecting an assistant, check its permissions, training settings, storage rules, and deletion controls. [support.google](https://support.google.com/mail/answer/14615114?hl=en) Prefer limited access. Avoid granting sending or deletion powers unless needed. Review recipients, attachments, and drafts before approving actions. [learn.microsoft](https://learn.microsoft.com/en-us/defender-office-365/step-by-step-guides/prompt-injection-protection-defender-for-office-365) Disable smart features you do not want, and disconnect unused services. [support.google](https://support.google.com/gemini/answer/16836988?hl=en)
Keep sensitive requests narrow. Do not assume deleting visible chat history removes preserved organizational records. [learn.microsoft](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy) The useful question is not simply whether you trust AI. It is whether this task justifies giving this system this much access.
Links and sources as returned
LINKS_CHECK: artificial intelligence, [https://en.wikipedia.org/wiki/Artificial_intelligence](https://en.wikipedia.org/wiki/Artificial_intelligence) Google, [https://en.wikipedia.org/wiki/Google](https://en.wikipedia.org/wiki/Google) Gemini, [https://en.wikipedia.org/wiki/Gemini_](https://en.wikipedia.org/wiki/Gemini_)(chatbot) Gmail, [https://en.wikipedia.org/wiki/Gmail](https://en.wikipedia.org/wiki/Gmail) Microsoft, [https://en.wikipedia.org/wiki/Microsoft](https://en.wikipedia.org/wiki/Microsoft) Microsoft 365 Copilot, [https://en.wikipedia.org/wiki/Microsoft_365_Copilot](https://en.wikipedia.org/wiki/Microsoft_365_Copilot) Outlook, [https://en.wikipedia.org/wiki/Microsoft_Outlook](https://en.wikipedia.org/wiki/Microsoft_Outlook) Prompt injection, [https://en.wikipedia.org/wiki/Prompt_injection](https://en.wikipedia.org/wiki/Prompt_injection) electronic discovery, [https://en.wikipedia.org/wiki/Electronic_discovery](https://en.wikipedia.org/wiki/Electronic_discovery) SOURCES: 1. [Google, How Gemini in Workspace protects your data](https://support.google.com/mail/answer/14615114?hl=en). [support.google](https://support.google.com/mail/answer/14615114?hl=en) 2. [Google, About personalization with Connected Apps](https://support.google.com/gemini/answer/16836988?hl=en). [support.google](https://support.google.com/gemini/answer/16836988?hl=en) 3. [Microsoft, Data, Privacy, and Security for Microsoft Copilot](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy). [learn.microsoft](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy) 4. [Microsoft, Manage data security and compliance for Copilot](https://learn.microsoft.com/en-us/purview/ai-m365-copilot). [learn.microsoft](https://learn.microsoft.com/en-us/purview/ai-m365-copilot) 5. [Microsoft, Prompt injection protection for email](https://learn.microsoft.com/en-us/defender-office-365/step-by-step-guides/prompt-injection-protection-defender-for-office-365). [learn.microsoft](https://learn.microsoft.com/en-us/defender-office-365/step-by-step-guides/prompt-injection-protection-defender-for-office-365) 6. [Google, Smart features and controls](https://support.google.com/mail/answer/15604322?hl=en&co=GENIE.Platform=Android). [support.google](https://support.google.com/mail/answer/15604322?hl=en&co=GENIE.Platform=Android)