Is pseudonymised use of AI automatically GDPR-compliant?

What is the difference between pseudonymised and anonymous?

Art. 4(5) GDPR describes pseudonymisation as follows: personal data is processed in such a way that it can no longer be attributed to a person without additional information, and this information is kept separately and protected. A placeholder such as “PATIENT_17”, whose real name is known from a table, is exactly that. Under Art. 32 GDPR, pseudonymisation is a recognised safeguard, but it does not make the data anonymous.

Under Recital 26 GDPR, data is only anonymous when the person can no longer be identified by the means reasonably likely to be used. The GDPR then does not apply to this data.

What did the CJEU decide in C-413/23 P?

The case between the European Data Protection Supervisor and the Single Resolution Board (SRB) concerned comments from shareholders that the SRB had passed on to a consultancy with code numbers instead of names. On 4 September 2025, the Court established three points:

• Pseudonymised data is not personal data in every case and for every person. Effective pseudonymisation can in fact prevent persons other than the controller from identifying someone (para. 86). The condition is that the recipient cannot reverse the measures and that other means do not lead to identification either (para. 77).
• For the controller, it remains personal data. Anyone who knows the key continues to process personal data.
• The duty to inform depends on the controller. Whether data subjects must be informed about recipients is assessed at the time of collection and from the controller’s point of view (para. 111), not from the recipient’s.

The judgment concerns the data protection regulation for EU institutions; its terms correspond to the GDPR. The Court referred the case back to the General Court of the EU for further examination.

What does this mean for a proxy in front of an AI model?

Every use of a proxy involves three perspectives:

Party	Is the data personal?	What follows
Your organisation	yes, you know the key	legal basis under Art. 6 GDPR, for health data an exception under Art. 9(2), informing data subjects, usually a data protection impact assessment
Proxy provider	yes, if it processes originals or the key table	usually a processor, data processing agreement under Art. 28 GDPR
Model provider	case-by-case assessment	depends on the means realistically available to it for identification

Two things matter for the model provider. First: under the draft Guidelines 02/2026 (para. 15), for a body that processes on behalf of another, the controller’s perspective is taken as the basis. If the data is personal for you, it is also personal for your processor. A model provider that computes on your behalf is in this role. It can only be different for a recipient who decides independently on purposes and means; the draft gives the example of a research institute that receives extracts from patient records. Second: even for an independent recipient, separating the key is not enough on its own. Rare diagnoses, location details or connections across several requests can make a person identifiable.

How do the data protection authorities assess anonymity?

The European Data Protection Board (EDPB) adopted the draft Guidelines 02/2026 on anonymisation on 7 July 2026; the public consultation runs until 30 October 2026. The draft considers anonymity in relation to the specific recipient and assesses three criteria:

1. Singling out: Can a record be attributed to a single person, for example through a rare combination of characteristics?
2. Linkability: Can records be linked with other data about the same person?
3. Inference: Can specific statements about a person be derived from the data?

The French CNIL also uses these three criteria. In its guidance on AI (May 2024), the German Data Protection Conference (DSK, the joint body of the German data protection authorities) points out that removing names and addresses from an input is regularly not enough, because the reference to a person can arise from the context. We are not aware of any general approval of a particular proxy software by a supervisory authority.

What do you have to do despite pseudonymisation?

• Determine the legal basis for each processing step, for health data the exception under Art. 9(2) GDPR.
• Inform data subjects about recipients under Art. 13 and 14 GDPR, even if these only receive pseudonyms.
• Conclude contracts: a data processing agreement with the proxy provider and, depending on its role, with the model provider (Data processing agreements with AI providers).
• Carry out a data protection impact assessment and assess the remaining risk of re-identification (DPIA for AI).
• Assess transfers to third countries separately (Third-country transfers).
• Observe § 203 StGB (German Criminal Code, professional secrecy) where professional secrets are involved.

How does this work with basebox?

With basebox, the platform and AI model run on a server at your organisation or in our cloud in the Munich-East data centre; by default, your data goes to no AI model of another provider. The question of whether a model provider may treat pseudonymised data as anonymous therefore does not arise. The data remains personal data and is processed on the basis of the data processing agreement and, for professional secrecy holders, the agreement under § 203 StGB. If you use the optional app with an external model, you decide with your data protection officer for which data (details).

Not legal advice: This article gives an overview as of 11 October 2026. Guidelines 02/2026 are a draft under consultation and may change; case C-413/23 P is not yet concluded before the General Court of the EU. Check your use case with your data protection officer.