People use "anonymize", "pseudonymize" and "redact" as if they meant the same thing. They do not, and the difference decides what you can do with the result and what rules still apply. Here is each one in plain terms, and which fits when you want AI help with client documents.
Redaction
Redaction means removing or blacking out parts of a document so that a reader cannot see them. The parts are gone, and nothing links the gap back to the original.
It is the right tool when you hand a document to another party, a court or the public. Done properly, it is permanent. Done badly, it is not: a black box drawn over text in a PDF can leave the text underneath. Redaction for filing or disclosure calls for a tool and a process built for that purpose, and a second person checking the result.
For AI work, plain redaction has a drawback. If every name becomes a black bar or "[REDACTED]", the AI cannot tell who is who. A contract with two parties and a payment between them turns into a contract with two blanks. You can often still get a summary, but the answer loses the relationships that made the question worth asking.
Pseudonymization
Pseudonymization replaces identifying details with consistent stand-ins, such as Person_1 and Company_1, and keeps the key that links them back to the real values somewhere separate.
The European Data Protection Board defines it, following the GDPR, as processing personal data so that it can no longer be attributed to a specific person without additional information, provided that this additional information is kept separately and protected by technical and organisational measures. In its guidelines, the Board describes such additional information as, for example, a table matching pseudonyms with the details they replace.
The point to remember is what it does not do. According to the same guidelines, pseudonymised data is still personal data, because it could be attributed to a person with the additional information. The UK ICO says the same: pseudonymised data is personal data in the hands of someone who holds the additional information. Pseudonymization reduces risk. It does not take the data out of data protection law.
For AI work, it has a real advantage: the placeholders are consistent, so the AI sees the structure of the document. And because you hold the key, you can put the real names back into its answer.
Anonymization
Anonymization goes further. The data is changed so that people can no longer be identified at all, by you or by anyone else, and there is no key to return to. Whether data is truly anonymous depends on whether identification is still possible using means reasonably likely to be used, not only on whether names are missing.
This is a high bar for documents about individual clients. A letter that describes a unique situation may identify its subject even with every name removed, because the story itself is the identifier. Full anonymization usually means removing or generalizing details until the document is no longer useful for the case, which is rarely what you want when you need help with a specific client matter.
Many tools and articles say "anonymize" when they mean "replace names". If the tool keeps a key so the names can come back, what it does is pseudonymization, whatever the button says.
The same sentence, three ways
Start with: "Maria Lopez, born 4 May 1980, owes Northwind Ltd 12,000 under the lease."
- Redacted: "[REDACTED], born [REDACTED], owes [REDACTED] 12,000 under the lease." Nothing links the gaps to anyone, and the AI cannot tell that the first and third gaps are different parties.
- Pseudonymised: "Person_1, born Date_1, owes Company_1 12,000 under the lease." The structure is intact, and a key held by you says Person_1 is Maria Lopez.
- Anonymised: "A tenant owes a landlord a sum under a lease." Nobody can be identified, and most of the detail is gone.
The sentence in the middle is the one you can ask an AI useful questions about and still map back to the right people afterward.
Which one fits AI work
| Goal | What fits |
|---|---|
| Send a document to another party, a court or the public | Redaction, using a tool and process made for it |
| Get AI help with a specific client matter and read the answer in terms of the real people | Pseudonymization, with the key kept on your side |
| Publish or share data so that nobody can be identified | Anonymization, which needs careful assessment and often loses detail |
For most professional AI use, pseudonymization is the practical choice. Consistent placeholders keep the document readable to the AI, the key lets you restore the answer, and your obligations stay in view. Because pseudonymised data is still personal data, treat the safe copy as sensitive, keep the key out of the chat, and check what your profession's rules ask of you when a third-party service receives the safe copy.
What none of them promise
No method can promise that nothing identifying is left. Automatic detection misses things: a nickname, a misspelled surname, a detail that only identifies in combination. Whatever you use, a person has to review the result. The sensible question is not "is this perfectly safe?" but "have I reduced what leaves my computer to what the task needs, and checked it?"
ClientSafe, a Windows app from GagarinSoft, makes pseudonymised safe copies of client documents on your PC, keeps the key on your PC, and restores real names in the AI's answer. It is not a redaction tool for documents you send to another party or a court. Read more about ClientSafe.
General information, not legal advice. Check the rules that apply to you.