Redaction that actually works

Replace the role, not the letters

The technique that does most of the work costs nothing and takes no time: write the role instead of the identity.

• “Mrs Adeyemi has missed three payments” becomes “the tenant has missed three payments”.
• “Acme Ltd’s contract with Northern Supplies” becomes “our client’s contract with their supplier”.
• “Patient DOB 12/03/1978, hypertension” becomes “a patient in their late forties with hypertension”.

The answer you get back is the same. The model was never using the name for anything; it needs the relationship, the sequence and the numbers. Once this becomes a habit rather than a step, most of the confidentiality problem in most people’s work disappears, because the sensitive material is not being sent at all.

When you need consistency

For longer documents where several people appear, replace consistently. Person A, Person B, Company X. Keep a small mapping in a file you do not send, and substitute back afterwards. This preserves the structure the model needs while removing the identities.

A find-and-replace in any editor does this in a minute. Free tools that automate it exist: Microsoft Presidio is an open-source library that detects and replaces names, numbers and identifiers in text, and spaCy’s named-entity recognition does a rougher version in three lines of Python. For occasional work, doing it by hand is faster than installing anything.

The parts of a file you did not mean to send

This is where careful people still get caught, because the sensitive material is not in the text.

Document metadata. Word and PDF files carry author names, organisation, editing history and sometimes tracked changes and comments that are not visible in the document. Use the built-in inspector (File, Info, Check for Issues, Inspect Document in Word) before sending anything outward.

Spreadsheet hidden content. Hidden rows, hidden sheets, filtered-out data, and the source data behind a chart all travel with the file. Deleting the visible rows is not enough.

PDF layers. A black rectangle drawn over text hides it from the eye and not from a text extractor. This has produced published court filings where the redacted names could be copied out with a mouse. Proper redaction removes the text; the free qpdf and most PDF editors can flatten, and checking with pdftotext afterwards takes ten seconds.

Image metadata. Photographs carry EXIF data including GPS coordinates and the device. Free tools such as ExifTool strip it; most phones can be set not to record location in the first place.

Screenshots. The most common leak by far. A screenshot of one message includes the browser tabs, the notification that arrived, the file names in the sidebar and the account you are logged into. Crop before sending, always.

Why “anonymised” is harder than it sounds

Removing names does not always make data anonymous. A well-known line of research has shown that a surprisingly small number of ordinary attributes (a postcode, a date of birth, a sex, or a handful of location points) is often enough to identify a specific person within a large population. Combining a de-identified dataset with a public one is a documented re-identification technique.

For a single document with one person in it, role substitution is effective. For a dataset of thousands of records, removing the name column is not anonymisation, and treating it as such is a mistake with regulatory consequences. If you are handling data at that scale, the question needs somebody who does this properly.

What redaction costs you

Be honest that it is not free. Removing detail sometimes removes information the model would have used: a name can carry a gender, a nationality or a professional register that changes how a letter should be written, and a company name lets a model apply what it knows about that industry.

Where that matters, put the useful attribute back without the identity (“a supplier in the garment trade”, “a female patient in her sixties”) which supplies the context and not the person. And where the identity is the question, such as research about a named public figure, redaction is not applicable and the decision is whether to send it at all.

The rule that survives everything

Before pasting, ask one question: could I email this to a stranger?

If the honest answer is no, either redact it or use a local model. That single question, asked reliably, prevents more problems than any technical measure in this course, because it is applied at the only moment when prevention is still possible.