Skip to content
News

GUIDES · September 24, 2026 · 6 min read

How to Redact a Document Before Sending It to AI

Two things go wrong before anybody starts. Black boxes in a PDF are not redaction, and deleting the names breaks the answer you were trying to get.

Why a black box is not redaction

Drawing a filled rectangle over a name in a PDF adds a shape to the page. The text is still underneath it — selectable, copyable, and read in full by anything that extracts the document. It has been the cause of a long line of public embarrassments, most of them by organisations with lawyers.

Proper redaction removes the underlying content rather than covering it. If your tool does not say it does that, assume it does not.

Why deleting is the wrong fix

The instinct is to strip the names out. It works, in the sense that nothing identifying is sent — and it ruins what comes back.

A model given [NAME] writes about [NAME]. You get a letter addressed to a placeholder, with pronouns that do not agree and sentences built around a gap, and you spend the time you saved repairing it.

Substitute. Do not delete.

Replace each value with a plausible one of the same shape: a different name, a different number, a different street. The document stays coherent, the model writes an ordinary reply about a person who does not exist, and you swap the real values back in before you read it.

Doing it by hand

This works and is worth knowing, even if you end up automating it.

  • •1. Find every identifier. Not only names. Emails, phone numbers, addresses, postcodes, dates of birth, account and card numbers, case or patient references, employer, and any unusual combination of ordinary facts.
  • •2. Replace each one with a stand-in of the same shape. Keep it plausible; keep it consistent, so the same person is the same stand-in throughout.
  • •3. Write the mapping down — which stand-in stood for which real value — and keep it on your own machine.
  • •4. Read what you are about to send. Not the original. The substituted version, looking for anything you missed.
  • •5. Reverse the mapping on the reply.

The weakness is step one. People are reliable at names and unreliable at the fifth account number on page nine, and the cost of missing one is the whole point of the exercise.

Doing it automatically

This is what Secure AI does. The identifying detail is found and replaced on your own device before anything is sent, the mapping stays with you, and the real values go back into the reply. You see what was found before it goes, and you can cover anything else yourself.

The file itself is never uploaded — PDFs, Word documents, spreadsheets and photographs are read where they are.

Try Secure AI free

Frontier AI chat — encrypted by default, with your identity stripped before any request reaches the provider.

Common questions

What about scans and photographs?

A photographed contract has no text to find until it is read. That recognition has to happen somewhere — and if it happens on a server, the document was uploaded to be read, which is the thing being avoided. It should happen on your own machine.

Does this make the document anonymous?

It makes it non-identifying for the purpose of asking a question about it — the model never receives a real person. Formal anonymisation is a separate legal standard about the data you keep, rather than about the copy you transmit, and the two are answering different questions.

Can I not just use a local model?

You can, and nothing leaves at all. The trade is capability: the models that run on a laptop are behind the ones that do not, and that gap is usually why you wanted a model to begin with.

More from News
Secure AI

Use any AI. Stay anonymous.

© 2026 Secure Artificial Intelligence, Inc.Private details never reach the AI.
How to Redact a Document Before Sending It to AI · Secure AI