The call almost always comes the same way. Somebody inside a company or a public body has seen something wrong, tried to say it, and nothing happened, and now they are holding a document. They send it to you. From that moment their safety depends on how you handle a file.
The law says exactly what makes them identifiable
Italian Legislative Decree 24/2023opens in a new tab transposes the European whistleblowing directive and puts a confidentiality duty at its centre. The identity of the person reporting may not be revealed, without their express consent, to anyone other than those competent to receive and handle the report.
The interesting part is not the prohibition, it is how the object is defined. Protection covers the identity and all information from which the identity can be inferred, even indirectly. The legislator wrote the protection at the right level, which is not the level of the name: they understood that inside an organisation a person is identified by elimination, and that context does the eliminating.
Formally that duty binds whoever handles the report inside the organisation, not the journalist receiving a document from an employee. But the definition describes with clinical precision the risk you are holding, and that is why it is worth reading from outside too: if the legislator had to write “even indirectly”, it is because indirect inference is the normal way this goes wrong.
The field narrows on its own
Try looking at a company document through the eyes of someone inside who wants to know who let it out. They do not read the content: they read who could have held it.
The circulation list narrows it to a department. The revision date narrows it to whoever worked on it that week. A margin note narrows it to whoever annotates that way. The filing number says which system it was pulled from, and systems keep access logs. If the document is a scanned printout, the fold and the shadow say which printer. In a few moves the list goes from “the company” to five people, and from five people to one.
It is the same logic the code of criminal procedure uses when it defines a journalist’s privilege: Article 200opens in a new tab does not protect the source’s name, it protects anything that could lead to identifying them. Two different laws, written in different eras for different purposes, arrive at the same technical definition because they describe the same phenomenon.
The moment the document stops being yours
A confidential document leaves your protection in three ways, and only the first is visible.
The first is publication, and a newsroom thinks about that: what to show, what to redact, what to hold back. It is the only moment when anyone stops to consider it.
The second is internal circulation. The full file ends up in the desk’s chat, the section editor’s mail, the newsroom’s shared drive, the laptop of the colleague helping with the numbers, the lawyer’s folder. Each copy sits on a system with its own retention policy, its own backups and its own logs, and none of those policies was written with this document in mind.
The third is the tool. Four hundred pages of board minutes cannot be read by hand overnight, and a language model summarises them in three minutes. That is exactly what those tools are good at, and it is also the moment the full document is transmitted to a third-party company, which receives it, keeps it for as long as it decides and, under some contracts, uses it for training. What a model actually remembers is a technical question with a long answer; the short question is whether this is the file to find out with.
Why “remove the name” is the wrong measure
The instinctive reaction is to black out the whistleblower’s name. In most cases it achieves nothing, for a simple reason: the whistleblower’s name is not in the document. An employee who lets a set of minutes out does not appear in the minutes. What is there is the set of details that leads to them, and that set stays intact.
The right measure is the one both rules describe: take out the elements that narrow the field. Names of the people mentioned, roles and departments, staff numbers, addresses, contact details, filing numbers, bank details, dates that identify a shift. Replace them with stable placeholders, so the text stays readable (“[ORG_2] resolved on [DATE_3]” summarises and translates like the original) and keep the map between placeholder and real value on your own computer, and only there.
On what remains in the file after the text is clean (the author in the metadata, the text layer under a scan, the revisions in a .docx) there is a whole piece, and here it is not a detail: the author of a .docx is often the corporate username of whoever created it.
The thing worth deciding first
The GDPRopens in a new tab requires anyone processing personal data to apply minimisation and measures appropriate to the risk, and the journalistic regime loosens duties towards the data subject without removing that. But the reason to anonymise first is not compliance: it is that the map between placeholders and names is an object you can delete.
When the investigation is over and the case is deleted, what remains is a text that leads back to nobody. To nobody for you, for the colleague who still has it in the drive, and for anyone who might ask the newsroom to hand over what it holds. It is a property you only get by deciding on it at the start, because downstream the full document is already in six places.