“Anonymising a filing” describes two jobs that are nothing alike. One covers: it takes the value out of the document, permanently. The other substitutes: it puts a placeholder where the value was and keeps the mapping somewhere. Calling both by the same name is convenient until you have to explain what you did, and at that point the difference is everything.
Redaction is for documents that leave
Redaction is the right route when a document leaves the office to stay out: a judgment being published, a filing uploaded to a platform, an exhibit sent to someone who must not read that value. What you want there is irreversibility. Nothing should remain under the box, and nobody should be able to get back to the value.
The practical rule of good redaction is that the text has to be removed, not covered. A rectangle drawn over a PDF is a rectangle: open the file with the right tool and you can select what is underneath. The same goes for metadata, for tracked changes in a .docx, and for the text layer of a scan.
How to redact properly, in five steps
It is worth being concrete, because the commonest mistake is not a theoretical one: it is a black rectangle drawn with the annotation tool, which covers rather than removes. A properly redacted PDF is made like this.
One: use the program’s redaction tool, the one that says it removes content, not the drawing or highlighting tool. Two: apply the redaction. In most programs that is a second, explicit command, and until you give it the document is still whole. Three: re-export the file, so that what you hand over is the result rather than a version with a history inside it. Four: strip the metadata, meaning author, title, originating software and file path. Five: if the document came from a scan, remember there is a text layer under the image, and that it has to go too.
Then there is the check, the part almost nobody does and that costs ten seconds. Open the file you are about to hand over and try to select what sits under the black. Better still: extract its text with a tool that reads the PDF without displaying it, and search for the name. If it comes back, the rectangle was a drawing. What else survives a cleaned document is in the piece about metadata.
Substitution is for documents that come back
If the text is going to a language model, redaction is the wrong route, and not as a matter of form. A model reading a pleading full of black boxes loses the relationships between the parties: it can no longer tell whether the person on page two is the person on page nine, and the summary it returns mixes up claimant and defendant.
A stable placeholder fixes that, and it fixes a second thing. [FULLNAME_1] is the same person throughout the document, so the model still understands who does what. And when the answer comes back with the same placeholders in it, you can read it onto the real case: the placeholder becomes the name again, and the work gets used instead of retyped.
There is nothing to put back under a black box. That is why a tool that redacts permanently is not a safer version of one that substitutes. It is a tool for a different job.
The right word is pseudonymisation
The GDPRopens in a new tab is precise here. Pseudonymisation is defined in Article 4(5): data processed so that it can no longer be attributed to a person without additional information, with that information kept separately. Pseudonymised data is still personal data, and the regulation still applies to it.
Anonymisation is something else. Recital 26 describes it as a transformation after which the person is no longer identifiable by anyone, by any reasonable means. A text with placeholders plus a dictionary that resolves them is not anonymous, by definition, because the dictionary exists.
Saying so is not pedantry. It is the sentence you need the day someone asks what happened to the data. “I pseudonymised the text, the mapping stayed encrypted on my machine and it disappears with the case file” is an answer someone can check. “I anonymised it with a program” is not.
And the Court of Justice gave the distinction some weight
So far pseudonymisation looks like anonymisation’s poor relation: it does less, so it is worth less. The Court of Justice’s judgment of 4 September 2025 in Case C-413/23 Popens in a new tab showed the relationship between the two to be something else, and more interesting.
The case concerned comments gathered by an authority, stripped of identifying data, replaced with a code and passed in that form to an outside auditor. The Court held that pseudonymised data does not automatically stop being personal (whoever holds the mapping remains a controller), but that where the recipient has no means reasonably likely to be used to get back to the individuals, that material may not be personal data for them. The same text, in other words, can be personal in one pair of hands and not in another.
That is exactly the shape of this job. A filing with placeholders is, for the firm that holds the dictionary, pseudonymised personal data: the record, the legal basis and the measures apply in full. For the model provider, receiving a text with no key and no reasonable way of reconstructing one, it may not be personal data at all. Pseudonymisation is therefore not imperfect anonymisation: it is the one operation that separates the two positions, and that separation is where all its usefulness sits. The full picture, with the EDPB opinion and the 2026 guidelines, is in its own piece.
Where the mapping lives is the real question
If the safety of a substitution rests on the mapping staying separate, then the only thing that really matters is where that mapping lives and who can read it. A cloud service that anonymises on your behalf saw the document in the clear before anonymising it, and it holds the mapping.
Omissia keeps all of it on the computer: the analysis, the model that recognises the values, and the mapping between placeholders and real names, encrypted with the system keychain. Delete the case file and the mapping goes with it, after which the model’s answer can no longer be reassembled. That is deliberate: a dictionary that outlives the case is an archive of personal data nobody asked for.
The rest of the argument, including what has to come out of a filing before you hand it to a model, is in the other piece. What the program does, step by step, is in How it works.