2026-08-26 · 5 min read

Document redaction done right — and where everyone fails

Why is a black box over PDF text not redaction?

Because a PDF separates its text layer from its graphics layer. Drawing a black rectangle adds a shape on top of the text — but the text itself remains in the file, intact and searchable. Selecting the area, copying it and pasting into any text editor recovers the 'redacted' content in full.

This failure has happened again and again in filings by courts, government bodies and companies around the world: a document that looks fully redacted, opened by any curious reader in seconds. The same applies to black text on a black highlight, or shrinking the font to zero — the data is still there.

What other redaction failures are common?

Even people who delete the visible text usually stumble on the file's hidden layers. Office documents and PDFs carry far more than what appears on the page.

  • File metadata — the author field, organization name, file paths and edit history survive content deletion. We cover this in depth in our guide to document metadata risks.
  • Tracked changes and comments — unaccepted revisions and margin comments stay in the file and surface with one click.
  • The scan behind the redaction — in a scanned PDF, deleting only the OCR text layer leaves the original image underneath.
  • The one missed occurrence — a name appearing 30 times; manual redaction almost always misses at least one, and one is enough.
  • Indirect identifiers — "the CFO of the northern subsidiary" identifies a person without any name.

How do you redact a document properly?

Proper redaction operates on the data, not the display: find every occurrence of the sensitive information, delete it from the text layer itself, clean the metadata — and only then export a fresh, clean file.

  1. Complete detection — scan the whole document automatically: names, IDs, phone numbers, addresses, amounts. Never rely on the eye alone.
  2. Human review — go over the findings and decide what goes and what stays.
  3. Removal from the data layer — delete or replace the text itself, never draw over it.
  4. Metadata cleaning — author, comments, tracked changes, document properties and previous versions.
  5. Verify the output — open the final file and try to select, copy and search for the removed information.

Redaction or token replacement — which is better for AI?

Redaction leaves black holes that destroy context: you can't tell who signed, who received, who committed. Anonymization with consistent tokens — the same person becomes PERSON_001 throughout — keeps the document fully readable, without the identities. When the goal is to analyze the document with an AI tool rather than archive it, that's the right approach: the AI gets the full context, and none of the content you should never paste.

Frequently asked questions

Does a black rectangle in a PDF really hide the text?

No. In most cases the rectangle is a graphic shape over a live text layer — select, copy and paste recovers the content in full. Real redaction deletes the text from the file itself.

I deleted the sensitive text — is that enough?

Not necessarily. File metadata — author, comments, tracked changes, document properties — may still contain the information or expose identities. Metadata cleaning is an inseparable part of proper redaction.

What's the difference between redaction and anonymization?

Redaction removes the information and leaves a hole — the document loses context. Anonymization replaces each detail with a consistent token (PERSON_001), so the document stays readable and useful without exposing identities.

Is manual redaction good enough for a short document?

It's risky to rely on. A name repeated dozens of times, spelling variants of the same name, and indirect identifiers are all easily missed by hand. Automated scanning followed by human review is the safe combination.

Can a properly redacted document be fed to an AI tool?

If all direct and indirect identifiers were removed and the metadata cleaned, the document no longer contains personal data, so feeding it in is not processing personal data. Consistent tokens are preferable, so the analysis stays meaningful.