The other pieces in this series are about what must not leave. This one is about what is left: the work still has to be done, the pages are four hundred, and the deadline is tomorrow. A language model on an anonymised file genuinely helps, and it is worth being precise about where.
What works, and why placeholders do not get in the way
A stable placeholder is an identifier. If “[FULLNAME_3]” always means the same person across the whole file, the model can count how often they appear, in which documents, next to whom and in what chronological order, exactly as it would with the real surname, because to a model the real surname was already just a recurring string.
From there, the expected things work. The orienting summary: four hundred pages reduced to a map of where to look, which is the most honest and most useful use. The timeline: pulling out every date and putting them in order is mechanical work that costs an afternoon by hand. Comparing versions: where the March minutes say something different from the November ones. Translating an annex in a language nobody in the newsroom reads. Finding recurrences: which documents have [ORG_2] and [FULLNAME_5] in them together.
None of these tasks gets worse because the names are placeholders, and each returns a result you map back onto the real values on your own computer. The full file never left.
What does not work, and is better known in advance
Anonymisation has a real technical cost, and it deserves stating. If two documents write the same name differently (“Mario Rossi”, “M. Rossi”, “Rossi Mario”) and anonymisation treats them as different people, the model will see three people where there is one. The link you cared about most is exactly the one you broke.
That is why confirming the detections is not a formality: it is the moment you decide those three spellings are one person and take one placeholder. Once done, the link holds across the whole file.
What stays out, and stays out for good, is anything that requires knowing who that person is in the world: that [FULLNAME_3] is the councillor’s brother, that [ORG_2] is controlled by whoever won the contract. That work happens on the real document, on your computer, in your head. It is also the work models get wrong most often and most confidently, so it is not much of a loss.
The fact nobody checked
There is a specific way these tools fail on an investigation, and it is not the flagrant invention. It is silent omission. You ask “find the contradictions between these two documents” and you get a list of three. The list is correct: those three contradictions really are there. But there were five.
A result like that is dangerous precisely because it is checkable and checks out. You verify the three, you find them, you trust the list. None of the three is false, so no check discards them, and the two that are missing leave no trace anywhere. A model that fails to find something is not telling you it is not there: it is telling you it did not find it, and the two sentences look alike only until you build a story on top of them.
The operational consequence is simple and unexciting: a summary is a map for deciding where to look, not a substitute for reading the pages that matter. It is the same distinction that holds up the argument about human review when it is not a checkbox, and what happens when somebody skips it is on the record in a courtroom.
Who signs
The AI Actopens in a new tab puts the heavy obligations on those who provide systems, not on those who use them to work. What touches a newsroom is transparency about generated or manipulated content, and the good sense of knowing which tool touched what.
The rest is already covered at home. The rules of conductopens in a new tab and the responsibility of whoever signs a piece were not written with language models in mind and did not need to be: checking the facts is a duty resting on a person, and no tool has ever counted as a good reason for not having done it. A badly summarised file produces a wrong story exactly as a hastily read archive did, with the difference that now there is something to blame.
The GDPRopens in a new tab, for its part, asks for measures appropriate to the risk. On a file holding sixty people who will never appear in the story, the appropriate measure has an obvious shape: do not send them.
The flow, in five lines
The document arrives and stays on your computer. Omissia reads it locally and proposes what it found (names, contact details, staff numbers, bank details, amounts, dates) and you confirm, which is also where you merge the different spellings of the same person. Out comes text with placeholders, and that is what you give the model. The answer comes back with the placeholders in it and you map it onto the real values, on your computer. When the investigation closes you delete the case: from then on the worked text leads back to nobody, not even to you.
That last line is the one worth deciding at the start, because it is the only one that cannot be added afterwards.