Anonymization or pseudonymization: what the GDPR says, and what it changes for AI

The two words are used as if they were interchangeable. In the GDPR they are not: one takes data out of the regulation, the other keeps it in. For a lawyer, a notary or a doctor who wants to use AI on a file, the difference decides what you can do with the answer.

1. Pseudonymisation: the GDPR definition

Article 4(5) of the GDPR defines pseudonymisation as processing personal data "in such a manner that the personal data can no longer be attributed to a specific data subject without the use of additional information, provided that such additional information is kept separately and is subject to technical and organisational measures". In practice: "Jean Martin" becomes "Person_1", and the table linking "Person_1" to "Jean Martin" is kept elsewhere, under lock.

Recital 26 draws the consequence: pseudonymised data which could be attributed to a natural person by the use of additional information should be considered information on an identifiable natural person. It remains personal data, and the GDPR still applies.

2. Anonymisation: no way back

For the CNIL, the French data protection authority, anonymising means making any identification of the person impossible in practice, by any means and irreversibly (our translation). The same recital 26 excludes from the GDPR data rendered anonymous in such a manner that the data subject is not or no longer identifiable. The bar is high: removing the name is not enough if the job title, the town and a date are enough to find the person again.

3. The difference in one table

4. What the GDPR expects of pseudonymisation

The GDPR names pseudonymisation twice as a safeguard (chapter IV): article 25 (data protection by design) refers to appropriate technical and organisational measures such as pseudonymisation, and article 32 (security of processing) lists the pseudonymisation and encryption of personal data. Pseudonymising does not exempt you from the GDPR, but it is one of the measures the GDPR expects of you.

The European Data Protection Board published Guidelines 01/2025 on pseudonymisation in January 2025, as a version for public consultation. They confirm that pseudonymised data remain personal data, even when the data and the mapping table are not held by the same person, and introduce the idea of a "pseudonymisation domain". In July 2026 the EDPB opened consultation on Guidelines 02/2026 on anonymisation, also still a draft.

5. The SRB judgment: a nuance, not a loophole

On 4 September 2025, in case C-413/23 P, EDPS v SRB, the Court of Justice of the European Union held that pseudonymised data must not be regarded as constituting personal data in every case and for every person: pseudonymisation may, depending on the circumstances, effectively prevent persons other than the controller from identifying the data subject. But the Court also held that, for the controller, identifiability is assessed from the controller's own point of view.

Three caveats before drawing conclusions. The judgment interprets Regulation 2018/1725, which governs EU institutions, not the GDPR itself, even though the definitions are parallel. The case was sent back to the General Court. And for a firm that pseudonymises its own files and keeps the mapping, nothing changes: it holds the key, so for it the data are personal data, and the duty of secrecy applies in full.

6. Why AI work usually needs pseudonymisation

A lawyer who asks an AI to draft a letter, summarise exhibits or compare clauses needs the answer to come back about the real client. With anonymised text, the draft is addressed to nobody and the summary names nobody. With pseudonymised text, the AI reasons on "Person_1" and "Company_2", and the mapping kept by the firm puts the real names back into the answer.

It is also the route taken by the French Bar's guide "Ethics and artificial intelligence" (2026), which recommends to "systematically pseudonymise or anonymise data before any use of a generative AI tool" (our translation). We go through its recommendations in a dedicated article.

7. The trap: indirect identification

Replacing names is not enough when the rest of the text still points to one person: the only paediatric surgeon in a town of 3,000, a date of birth next to a postcode, a very specific amount. That is the gap between pseudonymised and anonymous, and the reason why a pseudonymised copy should be reviewed before it goes out.

We wrote an article on re-identification risk, and we measure it on RAT-Bench, Imperial College London's benchmark, which tries to find the person among 3 million real records from the anonymised text.

8. How Hexagone AI handles both

• Masking. Each identifier is replaced by a neutral, stable marker (PNAME_1, CNAME_2). The same person keeps the same marker across every document of a folder.
• Synthetic replacement. Identifiers are replaced by realistic but fictional values, so the document still reads naturally.
• A mapping that stays with you. Detection, replacement and the mapping all run on your own computer. That is what makes the process reversible for you, and only for you: the AI receives the protected copy, and the answer is re-identified on your machine.