Pseudonymisation vs anonymisation under the GDPR: differences, examples and the CJEU ruling

What is the difference between pseudonymisation and anonymisation?

Pseudonymisation separates the data from the information that links it to a person, but the link can be restored. Anonymisation means the person can no longer be identified. Only anonymous information falls outside the GDPR.^[2] People search for both spellings, so for the record: "pseudonymization vs anonymization" is the same topic in US spelling, and the legal tests below are the same.

The GDPR defines pseudonymisation as processing personal data in such a way that it can no longer be attributed to a specific data subject without additional information, provided that this additional information is kept separately and is subject to technical and organisational measures (Art. 4(5)).^[1] Recital 26 adds that pseudonymised data which could be attributed to a person by the use of additional information is to be regarded as information on an identifiable person. Whether someone is identifiable depends on all the means reasonably likely to be used, by the controller or by another person, taking cost, time and available technology into account.^[2]

The Baden-Württemberg data protection commissioner puts it briefly in a discussion paper on AI: only lasting anonymisation removes the link to a person, and anonymising is itself a processing operation that needs a legal basis.^[6]

What do the two look like in three examples?

Example 1: a CSV export

A customer table holds name, e-mail, postcode, date of birth and revenue. If you replace name and e-mail with running numbers (CUSTOMER_001) and keep the number list somewhere else, that is pseudonymisation. Whoever has the list can match every row. Even without the list, the combination of postcode and date of birth stays unique in many rows. Moving towards anonymisation would mean birth years instead of birth dates, regions instead of postcodes and merging small groups, and even then you would have to test whether anyone can still be singled out.

Example 2: a letter

In a payment reminder to "Hanna Lenz" you replace the name with "[PERSON_1]". The text still contains the invoice date, the amount and "head of group audit". Anyone who knows the company recognises the person. That is pseudonymisation with gaps, far from anonymity. Making a single letter truly anonymous usually works only if you cut it so far that it no longer serves its purpose.

Example 3: a chat prompt

A prompt with "[PERSON_1]", "[COMPANY_1]" and "[IBAN_1]" is pseudonymised. You keep the table, the AI never sees it, and you can swap the answer back. For you as key holder the prompt remains a text containing personal data. How to do it step by step is covered in how to anonymise data before ChatGPT, which explains why the realistic result is pseudonymisation.

How do I decide which case I have?

Assume "pseudonymised" as long as you or anyone else can find the person again with reasonable effort. Three questions derived from Recital 26 help:^[2]

For prompts you will almost always end up in the first row, because you want to swap the answer back. That is not a flaw but the purpose: you use pseudonymisation as a safeguard, not as a way out of data protection law. The GDPR itself names pseudonymisation as a suitable technical measure, for example for data protection by design (Art. 25) and for security of processing (Art. 32).^[8][9]

Why is replacing names often not enough?

Because context identifies. The German Data Protection Conference (DSK) writes in its guidance on AI that removing names and addresses regularly does not suffice, because the context can still make the person identifiable.^[4] Dates, amounts, rare job titles and places work like parts of an identification number. The more unique the combination, the less swapping the name protects.

What did the CJEU decide in EDPS v SRB (C-413/23 P)?

On 4 September 2025 the Court of Justice ruled in the case of the European Data Protection Supervisor against the Single Resolution Board (SRB). It set aside the General Court's judgment and referred the case back. According to the Court's press release it held three things:^[3]

• A person's opinions and views relate to that person.
• Pseudonymised data must not be regarded "in all cases and for every person" as personal data. Pseudonymisation can effectively prevent persons other than the controller from identifying the data subject.
• For the controller's duty to inform, identifiability is assessed at the time of collection and from the controller's point of view. The duty applied before the data were passed to the consultancy, whether or not the data were personal data for the consultancy.

The political picture is open. On 19 November 2025 the Commission proposed making the concept of personal data relative. According to law-firm reports, the Council rejected that change in its position of 21 October 2026 and proposes a special rule for pseudonymised data held by third parties who cannot identify the person. Parliament has no position yet, and a conclusion is not expected before 2027.^[7] For today: the definition is unchanged.

When is real anonymisation worth the effort?

When you never need to trace the result back: published figures, analyses across many cases, or training material. There you merge individual details into groups, coarsen time and place, and remove identifiers. For single texts such as a letter or a complaint it is rarely sensible, because the content the answer depends on is exactly what you would remove. There, careful pseudonymisation with a visual check is the realistic route, combined with a check whether you may enter the data into the chosen tool at all.

What does it mean when a company sends a pseudonymised prompt to an AI provider?

The company holds the key and stays the controller of personal data. For the company the data is personal, because it can link it to its own mapping table.

• Legal basis: you still need one for your own processing of the prompt, and pseudonymising is itself a processing operation.^[1][6]
• Transparency: information duties towards the people concerned remain.^[11] According to the CJEU ruling, the controller's own point of view at the time of collection counts for the duty to inform.^[3]
• Provider: whether the provider can re-identify the person is a separate question. It depends on context, volume and the provider's means. Pseudonymisation remains a sensible safeguard, but it does not replace a check of plan, contract and legal basis. For the plan side see Is ChatGPT GDPR compliant?
• Protect the key: the table does not belong in the chat, in the same document or in the same folder as the pseudonymised texts (Art. 4(5) requires separate storage).^[1]

In practice, set an internal rule: who may keep the mapping table, where it lives and when it is deleted. Record in your register of processing activities that prompts with pseudonyms still contain personal data. And train staff that "pseudonymised" does not mean "approved": it describes a safeguard, not the outcome of a legal assessment.

Why does it matter that you cannot follow a prompt after sending it?

Because pseudonymisation is the only safeguard in this chain that you apply with your own means, and you apply it before the data leaves. After sending, you cannot observe where a provider processes or stores a prompt; you have its documentation and contract, and what has been sent cannot be recalled. Deleting a chat or changing a setting acts on the future. That is why Articles 25 and 32 GDPR name pseudonymisation as a measure^[8][9]: it reduces what is exposed if the path turns out to be wider than expected. It does not make the prompt anonymous and does not replace contract, legal basis or the human check, and no automatic detection finds every identifier. The provider side of the question, plan by plan, is in where your data goes when you use an AI assistant.

Which mistakes happen most often?

• "Names gone, so anonymous." See above: the context often still identifies.
• Key and data in the same place. If the mapping table sits next to the pseudonymised document, the separation is gone.
• "Anonymous because of the provider." That the provider cannot recognise the person does not make the prompt anonymous for you. The ruling says pseudonymised data are not personal data "in all cases and for every person", so the answer depends on who holds the data and with what means, not on a general status.^[3]
• Anonymisation without a test. Whoever claims data is anonymous should be able to show it. Document which details you removed or coarsened and which means of re-identification you considered.
• Forgotten parts of files. Headers, comments, metadata and file names often contain names.

This article is general information and not legal advice. It does not replace an assessment of your own case by your data protection officer or a lawyer. Last updated: 13 November 2026.