Redact and Sanitize a PDF Before Uploading It to ChatGPT

Redaction and sanitisation solve different problems

Visible redaction targets content on the page: text, images, or selected areas. In Adobe Acrobat, marking content identifies it for redaction, but the change becomes permanent only after the redaction is applied and the file is saved.

Sanitisation targets hidden information. Adobe lists items such as metadata, embedded content and attached files, scripts, hidden layers, overlapping objects, and information stored from prior saves. A document can need visible redaction, hidden-content sanitisation, or both.

Password protection is an access control, not redaction. Cropping changes the visible page area but can preserve cropped content. Converting or printing a PDF can change structure but should not be assumed to remove every hidden item.

Step 1: create the smallest useful PDF

Duplicate the original and identify which pages ChatGPT actually needs. Extract only those pages when permitted. Remove entire attachments or sections that are outside the task before marking individual redactions.

Create an inventory of names, email addresses, signatures, account numbers, legal references, confidential clauses, credentials, and distinctive project details. Add indirect identifiers such as exact dates, locations, roles, and rare facts.

If the PDF came from a word processor or spreadsheet, consider preparing the source document first. Redaction may be clearer in the source format, but the final PDF still needs its own inspection.

Step 2: apply real visible redaction

1. Use a redaction function designed to remove content rather than a drawing or annotation tool.
2. Search for known terms and inspect every result manually.
3. Mark text, images, and regions that must be removed.
4. Apply the redactions and save to a new filename.
5. Close and reopen the saved output before verification.

Search is an aid, not complete coverage. Adobe notes that text inside images or line art may not be found as normal text. Variants, misspellings, split words, and OCR errors can also escape a search.

Step 3: handle scanned PDFs and OCR

A scanned PDF can contain page images rather than searchable text. OCR creates a text layer that can help locate terms, but recognition is imperfect. Review low-quality pages, handwriting, rotated text, tables, stamps, and uncommon names visually.

After OCR-assisted redaction, inspect the page image and the searchable text layer. Search again for each identifier and try selecting or copying text around the redacted area.

Do not assume that a page is safe because a search returns zero results. It may mean the text was never recognised.

Step 4: sanitize hidden content

Use the PDF application's hidden-content or sanitisation tools to inspect metadata, comments, attachments, hidden layers, embedded files, links, scripts, form data, and other supported objects. Read the results and retain anything that the document genuinely needs.

Review document properties manually after sanitisation. Check the author, title, subject, keywords, application, and custom properties. Also check the filename itself.

Some workflows require interactive forms, signatures, links, or accessibility structures. Removing them can change function or evidentiary value, so work on a copy and consult the responsible reviewer.

Step 5: verify before upload

• Reopen the exact output in a new process.
• Search every sensitive term and common variant again.
• Attempt to select and copy around redacted regions.
• Review all pages at readable zoom, including margins and background images.
• Inspect properties, comments, attachments, layers, forms, and scripts again.
• Confirm the PDF opens normally and contains only pages needed for the task.
• Check the ChatGPT workspace and organisational approval separately.