What not to paste into ChatGPT

ChatGPT privacy is mostly a data-handling question: what you paste, which mode you paste it in, and which controls you have afterwards. The model is not the main risk. The text you hand over is.

This guide is deliberately narrow. It does not summarise OpenAI’s retention periods or promise that a setting makes you anonymous. It tells you what not to paste, where the controls live, and what to do when you have already pasted something you should not have.

The one rule

Treat the prompt box like a third party. If you would not email the text to an outside contractor, do not paste it into a chat window.

That rule survives every product change. Feature names move, menus move, policies change; “would I email this?” does not.

The never-paste list

• Credentials. Passwords, API keys, access tokens, connection strings, private keys, session cookies.
• Other people’s personal data. Customer lists, HR records, health information, addresses, phone numbers, anything that identifies a person who did not agree to this.
• Financial details. Bank accounts, card numbers, payroll files, unreleased numbers that would move a share price or a negotiation.
• Legal and investigation material. Privileged advice, witness statements, notes from an active dispute.
• Secrets in code. .env files, internal hostnames, customer identifiers in logs. Paste the shape of the error, not the credentials around it.
• Anything your employer’s policy forbids. That policy overrides this page.

If you need help with one of those, describe the problem without the identifier: “a Postgres connection string with the password replaced by X” is enough to debug the format.

Three modes, three storage paths

OpenAI documents different behaviour for different ways of using the product. The practical version:

The source for the Temporary Chat row is OpenAI’s own FAQ. We are not paraphrasing the retention or training rules beyond that; they live in the privacy policy, and a summary that is six months stale is worse than a link.

Temporary Chat is the right mode for a sensitive one-off. It is not a guarantee of anonymity, and it does not stop you from pasting something you should not paste.

Memory and custom instructions

Memory, when it is on, lets ChatGPT remember facts you shared — your name, preferences, goals — and use them in later conversations. You can view, edit, or delete memories at Settings → Personalization → Memory. Memory is not available in custom GPTs.

That has two privacy consequences:

1. A fact you save once can surface later. If you would not want a detail repeated in another conversation, do not ask it to remember that detail.
2. Memory is not a document store. Do not use it as a place to park client details. Use a project with the files you actually need, or paste the document in the conversation where it is needed.

Custom instructions are standing rules applied to all chats. Put your role, audience, and formatting preferences there. Do not put confidential facts there just because the field is convenient.

Projects and files

Projects keep related chats, files, and instructions together, and ChatGPT can use that context for ongoing work. Plan availability and workspace settings vary.

A project is a container, not a vault. The same question applies: would you email these files to the person whose job is to help with this task? If yes, a project is a reasonable place. If no, keep the file out.

Export and delete

• Export: Settings → Data Controls → Export. OpenAI sends a download of your data.
• Delete a chat: remove it from history. This does not remove a copy you pasted into another tool, and it is not a statement about backups.
• Delete the account: Settings, or the privacy portal. OpenAI documents both paths.

Before you delete an account, export what you need. Deletion is not a reversible archive.

If you already pasted a secret

Do this in order:

1. Rotate the credential. Change the password, revoke the token, reissue the key. Assume it is compromised.
2. Check access. Look for sessions, logins, or API calls you do not recognise.
3. Delete the conversation. It reduces exposure; it does not undo the copy that may exist elsewhere.
4. Tell the right person. If it is a work credential, your security team needs to know before the incident report, not after.
5. Write down what happened. A short note now saves an argument later.

Deleting the chat is step three, not step one. A deleted message is not a revoked key.

A 60-second pre-send check

• ☐ No passwords, keys, or tokens.
• ☐ No other person’s identifying data.
• ☐ The numbers are ones I am allowed to share.
• ☐ This is the right mode for the sensitivity (normal, Temporary, or not at all).
• ☐ If this leaked, I could describe it to my team without panic.
• ☐ The tool I am pasting into is the official one.

If any box is unchecked, rewrite the prompt until it is.