Skip to content

Glossary

Data masking

Stripping or replacing personal details in text before it reaches a model, so that names and account numbers never leave your control at all.

In plain terms

Before text is sent, something finds the personal details in it and takes them out or swaps them for placeholders. The model works on the altered version, and where the answer needs the real values they can be put back afterwards. It genuinely reduces what leaves your organisation, and it does not reduce it to nothing.

01

Why it matters

Because it is frequently the control that makes a tool usable at all where material would otherwise be too sensitive to send, and because it is routinely described as making material safe when it makes it safer. That gap matters most in exactly the situation the control was adopted for, which is why an honest account of its limits is more useful than an enthusiastic one.

02

How it works

Structured details are found reliably and unstructured identity is not. Account numbers, email addresses and card-like patterns have shapes that can be recognised consistently. A person identified by their role, their location and a date has no shape, so the detail that a general reader would use to work out who this is tends to survive the process untouched.

Replacing rather than deleting preserves usefulness and carries its own consideration. Swapping a name for a consistent placeholder lets a model follow who did what, which deleting does not, and consistency across a document is exactly what makes the surrounding context still point at somebody. The technique that keeps the text useful is the one that keeps more of the identifying structure.

Restoring afterwards is where the arrangement gets its value and its complexity. Putting real values back into an answer means holding the mapping somewhere for the duration, which is a small piece of sensitive material in its own right and one more thing to think about. It is manageable and it is not free.

It reduces exposure rather than removing the question, which is the framing worth adopting. Something identifying can survive, so the honest description is that less leaves rather than that nothing does. A reviewer told the second will treat the tool as unconstrained, and a reviewer told the first can make a proportionate decision, which is the outcome you actually want.

What comes out, and what stays

What comes out, and what staysThe reason this limit is structural rather than a matter of tuning is that the right-hand column has no form to recognise. A name looks like a name and an account number looks like an account number, so a tool can find them consistently, in any document, without knowing anything about your organisation. A role, an office and a date are ordinary words that identify somebody only in combination and only to a reader who knows the context, which is exactly the reader you were worried about. That is why the useful framing is a reduction with a residual rather than a clean removal, and why it is worth saying so when the control is described to anybody making a decision on the strength of it. A reviewer who understands the residual can decide proportionately; one who has been told the material is anonymous has been given no way to.Removed reliablyThe name.The account number.The email address.Frequently left behindThe role, and which office.The date it happened.Enough for a colleague to knowwho this is.The left column is what thecontrol was built to catch andit catches it well. The rightcolumn is what identifiessomebody to anybody who alreadyknows the organisation, and nopattern matches it.
The reason this limit is structural rather than a matter of tuning is that the right-hand column has no form to recognise. A name looks like a name and an account number looks like an account number, so a tool can find them consistently, in any document, without knowing anything about your organisation. A role, an office and a date are ordinary words that identify somebody only in combination and only to a reader who knows the context, which is exactly the reader you were worried about. That is why the useful framing is a reduction with a residual rather than a clean removal, and why it is worth saying so when the control is described to anybody making a decision on the strength of it. A reviewer who understands the residual can decide proportionately; one who has been told the material is anonymous has been given no way to.
03

Seen in the wild

  • Removing customer names and account numbers from a support thread before pasting it into an assistant for drafting help.

    ChatGPT
  • Stripping identifying fields from records inside an automation before the step that sends them to a model.

    n8n
  • Running a model on your own hardware, where masking is unnecessary because nothing is sent anywhere.

    Ollama
04

Common misconceptions

People assume

Masked material is anonymous.

In fact

It has had recognisable details removed, which is a real reduction. What frequently survives is the surrounding context: a role, a location, a date and an organisation size can still leave one candidate. Describing it as reducing exposure keeps a reviewer able to judge; describing it as anonymising invites them to stop asking.

People assume

It solves the sensitive material problem.

In fact

It makes some material sendable that otherwise would not be, which is genuinely valuable and is narrower than solving anything. For material where the rules decide the tool rather than your judgement, masking usually does not change what is permitted, and treating it as though it does is the error worth avoiding.

05

Telling them apart

Data masking vs PII

Data masking

A control that removes recognisable details before sending.

PII

The category of material the control is trying to catch.

The category is defined by what can be worked out; the control catches what has a recognisable shape. That gap is the whole story.

06

Questions

What does it reliably catch?
Details with a recognisable shape: account numbers, email addresses, card-like patterns. What it struggles with is identity carried by context rather than by format, such as a role plus a location plus a date, which is precisely the combination a human reader would use to work out who is being discussed.
Is replacing better than deleting?
For usefulness, generally yes, because consistent placeholders let a model follow who did what across a document. That same consistency preserves the structure that makes surrounding context identifying, so the version that keeps the text useful is also the version that keeps more of the shape.
How should we describe it to a reviewer?
As reducing what leaves rather than as anonymising. The first invites a proportionate judgement about residual risk; the second reads as a solved problem and tends to end the enquiry. The honest description is also the one that holds up when somebody looks closely later.
07

Key takeaways

  • Recognisable shapes are caught reliably; identity carried by context is not.
  • Replacing keeps text useful and keeps more identifying structure with it.
  • Restoring real values means holding a mapping, which is sensitive in itself.
  • Describe it as reducing exposure, never as anonymising.
09

Tools that use this

  • ChatGPT

    Removing names and account numbers from a support thread before pasting it.

  • n8n

    Stripping identifying fields before the step that sends a record to a model.

  • Ollama

    Unnecessary, because nothing is sent anywhere.

Last checked July 2026

All glossary terms