Manual redaction before prompting: placeholders, a quick checklist, and the settings that help

What to take out

Swap these for placeholders every time:

• Names of customers, patients, coworkers, and kids.
• Street addresses, and anything like gate codes or alarm codes.
• Phone numbers and email addresses.
• Account, policy, invoice, and order numbers.
• Card numbers, bank details, passwords, and login codes.
• Health details, legal matters, and anything about money that's specific to a person.

Watch for combinations, too. "The only dentist on Elm Street with a red door" names someone without a name in it.

Step 1: Paste into a plain text app first

Don't paste straight into the chat. Paste into a plain app on your computer: TextEdit on a Mac or Notepad on Windows. Now you can see everything, and nothing has been sent anywhere.

Step 2: Find and Replace

Use the app's Find and Replace. Replace the customer's name with [CUSTOMER], the street with [ADDRESS], and so on. Use the same bracket every time the same thing appears, so the reply still makes sense. Then read the whole thing once more for anything you missed, like a phone number in an email signature.

Step 3: Tell the AI to keep the brackets

Paste the cleaned version and add:

Some details are replaced with placeholders in square
brackets. Keep every placeholder exactly as written.
Don't guess what they stand for.

When the reply comes back, swap the brackets for the real details in your email app, not in the chat.

Settings that help, but don't replace this

ChatGPT has a setting to stop your new chats being used to improve its models: Settings → Data controls → Improve the model for everyone, turned off. It also has Temporary Chat, which OpenAI says isn't saved to your history and isn't used to train its models, though a copy may be kept for up to 30 days for safety. Both help. Neither means the text was never sent. Redacting first does. For the full picture, see is your business data safe in ChatGPT?

If you use AI at work, your company's rules come first. Some workplaces allow only certain tools or plans.

Under the hood

Why manual redaction beats asking the model to redact: if you paste raw text and ask the AI to anonymize it, the raw text has already been sent. Redaction only protects data when it happens before the request.

Practical rules for the tokens:

• Stable mapping. The same entity gets the same token throughout, so references stay coherent.
• Typed tokens. [CUSTOMER] and [ADDRESS] carry enough meaning for the model to write naturally around them.
• Preserve instruction. Models sometimes "helpfully" fill in a placeholder. Say explicitly not to.
• Quasi-identifiers. Rare combinations (small town plus trade plus date) can re-identify a person. Generalize them.

OpenAI documents that turning off "Improve the model for everyone" stops new conversations being used for training, and that Temporary Chats may be retained up to 30 days for safety. Business plans have their own terms; check yours.

Short close

Paste into a plain text app, swap private details for brackets, tell the AI to keep them, and put the real details back yourself. If you wouldn't write it on a postcard, don't paste it.