Skip to content

How-to

How to redact a PDF or Word document before ChatGPT

Prepare a separate copy on your Mac: replace names, addresses and identifying numbers with consistent placeholders, check the complete output, and share only that reviewed copy with ChatGPT. Keep the original and placeholder key private for restoring details later. This reversible process is pseudonymization, even when people call it anonymization. Remaining facts can still identify someone, so removing names alone is not enough.

Last reviewed · 6 sources

Steps

  1. 01

    Copy the document

    Keep the original private and work on a copy.

  2. 02

    Replace personal details

    Use consistent placeholders such as [PERSON_001].

  3. 03

    Review everything

    Read the full copy, including headers, tables and signatures.

  4. 04

    Share and restore

    Share only the reviewed copy, then restore the real details in the answer.

Step 1: work from a copy, not the original

Never edit the only copy of a client document. Make a working copy, and keep the original somewhere ChatGPT never sees. If the document is a scan, you need its text first, which means OCR, and OCR mistakes can hide names from simple search.

Step 2: replace details consistently

Replace people, organisations, addresses, emails, phone numbers, dates of birth and ID numbers with typed placeholders. Consistency matters more than style: the same person must get the same placeholder every time, or the assistant may confuse the parties in its answer.

Keep a private key, such as [PERSON_001] = the client, stored with the original and never pasted into the chat.

Step 3: review the whole copy

Read the entire redacted copy, not just the highlighted changes. Look at headers, footers, signatures, tables and file names, where details often survive. The Law Society of Ontario’s white paper warns that even anonymized facts can sometimes be pieced back together, so remove unusual identifying facts too.

Check the actual PDF or Word copy you will share

A rectangle drawn over text is different from applied redaction. In Acrobat, mark the content, apply the redactions and save the result; Adobe also offers a separate hidden-information cleanup option. Preview has its own Redaction Selection tool: do not confuse that with drawing a shape over a name.

Reopen the output locally, search for known names and numbers, and try copying the text into a local editor. Inspect the pages too. These checks can expose an incomplete edit, but a failed search is not proof that the file contains no personal information. Check its filename, comments, properties and any attachments before uploading it.

For Word, review revisions, comments, headers and properties as well as the body. For a scan, replacing the OCR text does not remove identifying content from the underlying page image. If your AI task only needs text, share the reviewed text copy instead of attaching the original PDF or .docx.

Step 4: share the approved copy, then review the return

Share the reviewed text or file, or use a connection limited to approved copies. Check what you are attaching: the reviewed copy and the original can have similar names. Ask the assistant to keep each placeholder exactly as written and never include the private mapping in the prompt.

Restore the details locally when the draft returns, then verify the parties, facts and amounts against the source. The assistant may preserve the tokens yet change the meaning around them. Its provider still processes the content you share; local preparation does not make a cloud conversation local.

Anonymization or pseudonymization?

Pseudonymization replaces identifiers while keeping additional information that can connect the copy back to a person. That is useful for a draft you need to restore. It does not establish that the remaining facts are anonymous: a distinctive event, role or combination of details may still identify the client.

Decide whether the task needs those facts at all, and follow your organisation’s rules for confidential material. The private key must stay separate, and your obligations do not disappear when you replace names.

Where PiBye fits

How PiBye handles this

PiBye prepares a separate approved copy from PDFs, scans, Word .docx, Markdown and text files on your Mac. It keeps the original in the private Vault, proposes stand-ins, and lets you review the complete text before sharing. This is a preparation workflow; it is not a promise that every identifier has been found or that the original file has been sanitized.

With a connected assistant, finished drafts arrive in the Vault automatically. Open a draft and choose Put details back to restore from the private mapping on your Mac, then review the result. The original, mapping and restored copy stay local; the approved content you share is still processed by your chosen assistant.

PiBye showing an approved, pseudonymized copy ready to copy into an AI chat
PiBye showing an approved, pseudonymized copy ready to copy into an AI chat

The real PiBye window, with invented sample details. Open full-resolution image ↗

Version 1.0.16 · Apple Silicon (M1 or later) · macOS 14.8.5 or later · 1.1 GB

Frequently asked questions

Can I just ask ChatGPT to anonymize the document for me?

That sends the original, names and all, to ChatGPT first, which is the exposure you were trying to avoid. Strip identifiers before the document reaches the AI.

Does anonymizing make using ChatGPT compliant?

No tool makes a use compliant. Replacing identifiers with tokens reduces the personal information you disclose; it does not replace your own judgment on confidentiality and client consent.

Is a copy with placeholders anonymous?

No. As long as a key exists, it is pseudonymized and still personal information for you. Keep the key out of the chat, and review remaining context that could still identify someone.

Sources

  1. Redact sensitive content in PDFs in Acrobat Pro, Adobe. Checked 13 November 2026.
  2. Annotate a PDF in Preview on Mac, Apple Support. Checked 13 November 2026.
  3. Inspect document: Word for Mac properties, Microsoft Support. Checked 13 November 2026.
  4. Remove hidden data and personal information by inspecting documents, presentations, or workbooks, Microsoft Support. Checked 25 September 2026.
  5. NISTIR 8053: De-Identification of Personal Information, National Institute of Standards and Technology, October 2015. Checked 25 September 2026.
  6. Licensee use of generative artificial intelligence (white paper), Law Society of Ontario, April 2024. Checked 17 September 2026.