Where your prompts go

The postcard and the envelope

Imagine posting a letter to a helpful expert. The envelope keeps the postman from reading it on the way. But the expert has to open the envelope to answer you, so they read every word. A cloud AI works the same way: the trip is protected, the destination is not.

The journey of one prompt

Say you paste a draft email into a chat at chat.example and ask "make this friendlier". Here is what typically happens, step by step:

1. Your browser encrypts it with HTTPS and sends it. People on your Wi-Fi or at your internet provider can see that you talked to chat.example, but not what you wrote. (The encryption module explains what HTTPS does and does not hide.)
2. The provider's servers decrypt it. They must, or the model could not read it. From here on, your text is readable by the provider's systems.
3. The model reads the whole conversation. Most chats send the earlier messages again with every new one, so something you pasted an hour ago is still being sent.
4. The request is usually logged. Logs often include your account, the time, your IP address and, depending on the service, the text itself.
5. The text may be kept. Chat history, safety checks for abuse, and, on some services and plans, improving future models. Some providers let you switch training off; the settings differ by provider and plan.
6. Tools and plugins can pass it on. If the AI searches the web, reads your email or calls another service for you, parts of your prompt can travel to those services too.

What the provider can learn

The words you type are only part of it. Around each prompt sit metadata: facts about the message rather than its content.

Put together over weeks, prompts and metadata can paint a detailed picture: your job, your health worries, your money troubles, the names of your colleagues. People often tell a chat assistant things they would never type into a search box, because it feels like a private conversation.

A common misconception

"I deleted the chat, so it is gone." Deleting usually removes it from your history view. Copies may still exist for a while in backups, safety logs or review queues, for as long as the provider's policy says. Treat anything you send to a cloud AI as something that might be kept.

A simple rule of thumb

Before you paste, ask: "Would I be fine if a stranger at this company read this?" If the answer is no, do one of three things:

• Remove the personal parts first (names, emails, phone numbers, account numbers). You will write code for this in the lesson after next.
• Use an on-device model, which never sends your text anywhere. That is the next lesson.
• Do not use AI for it. Some things, like passwords and other people's secrets, never belong in a prompt.