Redact a PDF Before Uploading to ChatGPT (Free)

Asking ChatGPT to summarise a contract, explain a medical letter or check a bank statement is genuinely useful. But uploading the file sends everything in it — names, account numbers, addresses, signatures — to OpenAI’s servers. If you want to redact a PDF before uploading it to ChatGPT, the method matters: the usual black box drawn in a PDF viewer hides text from your eyes, but not from an AI that reads the text layer.

How do you redact a PDF before uploading it to ChatGPT?

Remove the sensitive data itself, not just its appearance. Use a true redaction tool that rebuilds each marked page so the text underneath is deleted, check that nothing can be copied or searched, then upload only the redacted copy. Slate PDF’s Redact PDF does this in your browser, so the original never leaves your device.

1. 1Open Redact PDF and choose the file.
2. 2Drag a box over every name, number, signature or photo to hide; use Whole pages to black out pages entirely.
3. 3Click Redact and download the new copy.
4. 4Check it: search for a redacted name and try copying the blacked-out area. Nothing should come back.
5. 5Rename the file if the filename itself contains a name, then upload the redacted copy.

Why a black box isn’t enough for AI

Most PDF viewers let you draw a filled rectangle, or highlight text in black. Both add a shape on top of the page; the text underneath is still in the file. A person looking at the page sees black, but ChatGPT does not look at the page the way you do — it extracts the text layer, and the “hidden” name is right there in it. Ask it a question and it may quote the very number you thought you had covered.

Slate PDF’s redaction is the true kind: every page you mark is re-rendered as an image with the black boxes burned in, and the original page content is discarded. Our guide on how to redact a PDF properly covers the difference in more detail.

What to redact before you upload

• Full names and initials — yours and everyone else’s in the document.
• Dates of birth, ID, passport, tax and national insurance or social security numbers.
• Bank account, card, policy and customer reference numbers.
• Home addresses, phone numbers and email addresses.
• Signatures and photos of people or ID cards.
• Medical record numbers, diagnoses tied to a named person, case and docket numbers.
• Anything in headers and footers — reference numbers often repeat there on every page.

Keep what the AI needs to do its job. To ask about a contract’s notice period, it needs the clauses, not the parties’ names; to check a payslip’s tax calculation, it needs the figures, not the employee number. Redacting too little is a privacy problem; redacting the substance makes the answer useless.

The catch: redacted pages become images

Because each marked page is rebuilt as a picture, the rest of the text on that page is no longer selectable — and ChatGPT may then report that it cannot read those pages. Pages you did not mark keep their text. If you need the AI to read the remaining text on redacted pages, add a text layer back afterwards:

1. 1Redact the PDF and download the redacted copy.
2. 2Run that copy through OCR PDF, which recognises the visible text and makes it searchable again.
3. 3The black boxes contain no text, so nothing you redacted comes back — only what is still visible is recognised.
4. 4Search the OCR’d file for a redacted name one more time, then upload it.

Don’t forget the hidden parts of the file

A PDF can carry information you never see on the page. Redact PDF starts a fresh document, so the original title and author fields are not copied across, but comments and annotations on pages you left unmarked do come along. Delete stray comments before you redact, and read how to remove metadata from a PDF if the document passed through other hands. And check the filename: “J-Smith-divorce-settlement.pdf” undoes careful redaction the moment you upload it.

An alternative: paste a placeholder version

For text-heavy documents, you can skip the upload entirely. Extract the text with PDF to Markdown, use find-and-replace to swap each name for a placeholder such as [CLIENT] or [EMPLOYER], and paste the cleaned text into the chat. You see every word the AI receives, and placeholders keep the meaning intact — “[CLIENT] must give [SUPPLIER] 30 days’ notice” is still easy to reason about.

Does ChatGPT keep the files you upload?

That depends on your account type, your settings and OpenAI’s current policy, all of which change. On consumer plans, conversations may be used to improve models unless you turn that off in the data controls, and deleting a chat does not guarantee immediate deletion everywhere. Business and enterprise plans typically have different terms. Read OpenAI’s current privacy documentation for your plan — and assume anything you upload could be retained. Redacting first means it does not matter as much.