A Legal Team's Guide to Defensible PDF Redaction

Redaction failures in litigation follow a pattern. A filing goes out with black bars over the sensitive passages. Someone — often a reporter who reads dockets for a living — opens it, selects the page, and pastes it into a text editor. The "redacted" material is right there, and now it's a story about the failure as much as the content.

The uncomfortable part is that these mistakes are rarely made by careless people. They're made by people who did what looked like the right thing.

Why legal work is unusually exposed

Filings are public and permanent. Once a document is on a docket it has been downloaded, mirrored and archived. You cannot recall it, and a correction announces the error.

Adversaries are motivated and technical. The person most interested in what's under the box is the party you're redacting it from.

Volume invites shortcuts. Discovery runs to thousands of pages. Any process that's slow gets streamlined, and the step people drop is verification.

Documents are mixed. Exhibits are scans. Attachments are photographs. A production set is rarely uniformly digital, and tools behave differently on each.

The three failure modes

1. Covering instead of removing. A rectangle drawn in a PDF editor is an annotation; the text underneath is intact. This is the classic and still the most common failure.

2. Marking without applying. Acrobat's redaction is genuine, but it's a two-step process: Mark for Redaction, then Apply. A file saved after marking looks redacted and isn't. This one catches experienced users precisely because they were using the "right" tool.

3. Redacting the text layer of a page whose exhibit is an image. A page can carry a typed body and a scanned insert. Tools that classify each page as "text" or "scan" and pick one path will handle such a page as text — correctly removing text-layer content — and never examine the embedded image. The privileged name inside the scan survives, in a document that passes every visual check.

A workflow that holds up

1. Keep the original untouched. Redaction is irreversible. Work on a copy, under a naming convention that makes it obvious which is which.

2. Build a term list before you start. Party names, witnesses, minors, account numbers, matter references. Names should be entered explicitly rather than left to pattern-matching — no pattern distinguishes a person's name from a company's, and name detection that guesses will over-redact your own client's name into unreadability. Supply the name and the whole cluster goes, middle names included.

3. Run structured detection alongside it. Social Security numbers, financial account numbers, dates of birth, contact details — the high-volume identifiers a human skims past on page 300.

4. Verify every output. Not a sample. See the checklist below.

5. Record what you did. Which terms, which categories, who reviewed it. If the redaction is ever challenged, "we ran a tool" is a weaker answer than a documented process.

The verification checklist

Before anything is filed or produced:

• Copy-paste test. Select all, copy, paste into a plain text editor, search for each redacted term. Nothing should appear.
• Text extraction. pdftotext filing.pdf - prints every text object in the file regardless of what's drawn on top.
• Scanned pages magnified. No characters showing through the black.
• Embedded images checked on pages that also contain text.
• Metadata stripped. Document properties routinely carry an author name and the original filename.
• A human read it. Someone who knows the matter, reading the actual pages.

The first two take under a minute for a whole document and would have prevented essentially every redaction failure that has made the news.

On "defensible"

There's no certification that makes a redaction defensible. What makes it defensible is that it is actually irreversible and that you can describe the process that made it so.

That means being able to say: the content was removed from the file rather than covered; scanned material was OCR-read and the pixels destroyed; embedded images were checked, not just the text layer; metadata was stripped; and a named person reviewed the output before it went out.

It also means being candid about limits. Automatic detection finds well-formed patterns and the terms you supply. It does not read handwriting — so signatures and margin notes need human attention — and no OCR is perfect on a poor-quality exhibit. A process that assumes otherwise isn't defensible; it's optimistic.

Where a tool helps and where it doesn't

A good tool removes the mechanical burden: finding every instance of a name across 600 pages, catching the account number on page 412, handling the scanned exhibits without a separate workflow, and doing it without your case material being retained on someone's server.

It does not decide what is privileged. That judgement is the part that can't be automated, and it's why the review step stays.

More on the mechanics in how to redact a PDF, and on exhibits specifically in how to redact a scanned document. If you want to see how the tool handles a production document, try it on one — or read what we do with your files first.