Redaction that survives somebody trying

"I removed the names" is not redaction

It is the most common workaround in every office: strip the name, keep the text, paste it in. It feels careful, and for a large class of documents it does almost nothing.

The reason is that names are rarely the strongest identifier in a real case file. The combination of ordinary, boring details is.

The combination is the identifier

Work published from US census data has repeatedly shown how few attributes it takes. A widely cited 1990s analysis estimated that around 87 per cent of the US population was uniquely identified by the combination of five-digit postal code, date of birth and sex. A later reanalysis with different data and assumptions put the figure closer to 63 per cent. The two estimates disagree, and the disagreement does not matter for your purposes: on either number, three unremarkable facts identify most people.

Now look at a typical anonymised case note. A 34-year-old warehouse supervisor in a named town, with a specific rare condition, whose incident happened in March. There is no name anywhere in it, and anybody local could tell you exactly who it is.

Rarity is the thing to watch. A common diagnosis in a big city is anonymous. An unusual job title, an uncommon condition, a distinctive sequence of events, a small organisation — each collapses the pool of people it could be, and they multiply.

What actually reduces the risk

Generalise rather than delete. Age band instead of date of birth. Region instead of town. "A manufacturing employer" instead of the company. This keeps the text usable while widening the pool, which is what anonymity actually is.

Use consistent pseudonyms. Replace the individuals with Person A, Person B, Company C — consistently, so relationships in the text still make sense and the answer you get back is still about the right people. Keep the key in a separate file that never goes near the tool.

Remove the narrative fingerprints. "The fire at the distribution centre in March" identifies far more precisely than a surname. So does a quoted phrase from a well-known local dispute. These are the details people never think to remove because they are not personal data in the everyday sense.

Suppress small numbers. In aggregate data, a cell containing one or two people identifies them. Public bodies routinely suppress or round counts below five, and the same discipline applies to any table you paste into a tool.

The trap in automated redaction

Tools exist that find and mask personal data automatically. Presidio, from Microsoft, is free and open-source and does a respectable job. Several commercial products do the same.

Two cautions, one of them badly under-appreciated.

First, measure before trusting. Automatic detection is trained on typical patterns, and it misses what is unusual in your material — a local reference number format, a place name that is also a common word, an identifier written in a way it has not seen. Run it over fifty of your own documents and count the misses. That count is your actual protection level, and it is not the number in the vendor's brochure.

Second, and decisively: do not use an online model to redact text you are then afraid to send to an online model. If you paste the unredacted case note into a chat window and ask it to remove the identifying details, the unredacted case note has already been sent. Every duty you were protecting was breached at that moment, and the tidy redacted version that comes back changes nothing about it. Redaction has to happen before anything leaves — by hand, or with a tool running on your own machine.

Structured data has its own version

For a spreadsheet, the same logic in numbers. Look at whether any combination of columns produces groups of one or two rows — postcode plus job title plus start year will often do it. Where it does, generalise a column or remove one, and check again.

The formal name for this idea is k-anonymity: every record should be indistinguishable from at least k−1 others on the identifying columns. You do not need the formalism to use it. You need to sort by the combination and look for groups of one.

And do not forget the file itself

A document carries metadata: author, organisation, previous file names, tracked changes, comments, and in spreadsheets, hidden columns and the source data behind a pivot table. Sending "the summary tab" often sends the raw sheet with it.

mat2 strips metadata from many formats, exiftool handles images and PDFs, and every office suite has a "inspect document" or "remove personal information" function. Use one, then reopen the file and look at what is actually in it.

The test, and the case where nothing works

The check that beats every tool: give the redacted version to a colleague who knows the area and ask them to guess who it is. If they can, so can anybody with the same background knowledge, including whoever might later read the tool's logs.

And be honest about the residual case. Sometimes the facts themselves are the identifier — a single incident, a unique role, a matter that was in the local news. There is no redaction that survives it. Then the choice is a model running on your own machine, or not using one, and "I anonymised it as best I could" is not a third option.