No. A peer-reviewed legal analysis concludes that token ID sequences from language model inputs containing personal data are pseudonymised personal data by default under the GDPR, not anonymous data. Tokenisation, hashing or encryption reduce risk but do not automatically make data anonymous, so all GDPR duties continue to apply until irreversibility is proven.
On 4 October 2026 the International Review of Law, Computers & Technology published the study Qualifying LLM tokenisation under the GDPR: pseudonymisation by default. The core point: when a language model converts text containing personal data into token IDs, that sequence of numbers can, under a fixed version of the tokeniser, retain a traceable relationship to the original text. Legally that is pseudonymisation, not anonymisation. This means the data remains within the scope of the GDPR.
This touches on a widespread misconception. Many organisations treat a tokenised, hashed or encrypted text as if it thereby falls outside the privacy rules. The study shows why that assumption is incorrect for token-ID sequences derived from personal-data inputs in typical language-model projects.
What exactly does the new research state about tokenisation and the GDPR?
The study qualifies tokenisation in common language model projects as pseudonymisation by default. Tokenisation is the technique whereby text is broken up into small units and translated into numerical identifiers that the model processes. Because the vocabulary, normalization rules and tokenizer metadata can preserve a referable relationship under a fixed tokenizer version, a link between the token IDs and the source text can remain.
The authors align with the existing legal distinction: pseudonymisation keeps identification possible via additional information, anonymisation requires that identification is no longer reasonably possible. They describe an anonymisation claim as exceptional, and tenable only with recipient- and context-specific proof that the link back to the source is reasonably irreversible.
Why do encryption, hashing or labels not automatically make data anonymous?
Because each of these techniques reduces risk without removing the link to the person. A key, a tokeniser version, a source file or the behaviour of the model itself can restore the relationship with the original text. When a relevant party can reasonably use that additional information to identify individuals, it remains pseudonymised personal data.
The European Data Protection Board (EDPB) applies the same assessment framework in its guidance on anonymisation and pseudonymisation: pseudonymised data falls under the data protection rules, anonymous data only when identification is reasonably excluded. The European Commission elaborated on this in November 2026 with a practical example in which names, addresses and reference numbers are replaced locally by labels before text goes to an AI assistant. That too, according to the Commission, is pseudonymisation, not anonymisation, and context, language and remaining unique characteristics continue to weigh in.
The same holds more broadly across the chain. An earlier analysis in the Computer Law & Security Review on synthetic data under the GDPR already warned that legal risks do not disappear through transformation alone. In our assessment, the combined source set extends that same line of reasoning to prompts, training and fine-tuning: the transformation itself is not proof of anonymity.
- Tokenisation: can preserve a referable link under a fixed tokeniser version.
- Hashing: link restorable via lookup or brute force where value ranges are limited.
- Encryption: reversible as soon as the key is available.
- Labels instead of names: remaining free text and rare combinations can still isolate a person.
When may I treat a token ID sequence as anonymous after all?
Only after a separate, context- and recipient-specific assessment showing that tracing back to the source is reasonably irreversible under the concrete access situation. Our analysis is that the sources give no general threshold or template, and that a party claiming anonymity must substantiate that claim with proof under the concrete access situation; the sources moreover treat anonymisation as exceptional.
The question the sources leave open is precisely which concrete proof is sufficient in practice. As long as that is uncertain, an anonymity claim on a token ID sequence without substantiation is, in our assessment, a legally vulnerable position. A workable verification protocol:
- Classify the input as personal data until proven otherwise.
- Separate keys, tokeniser versions and identity information from the transformed data.
- Document which parties hold which additional information.
- Test the transformed text for unique free text, rare combinations and contextual identifiers.
- Assess tokenisation and model output separately, because an answer too can be traceable.
- Record why anonymisation would be irreversible under the concrete access situation.
What does this mean for directors, lawyers and CISOs deploying language models?
In our analysis, because token-ID sequences containing personal data should be treated as pseudonymised personal data by default, organisations sending prompts to a language model should assess the continuing applicability of GDPR duties—including processor arrangements, a legal basis, retention controls and data-subject rights—and should document tokenisation in their records as pseudonymisation rather than assume anonymisation. In our analysis, an unsupported anonymity claim in a DPIA or contract creates a legally vulnerable position, particularly if a breach or supervisory query later exposes remaining linkability. Lawyers should therefore review existing anonymity assertions about AI data and, where the evidence is insufficient, document the data as pseudonymised until a defensible anonymity assessment has been completed. In our analysis, because a key, tokeniser version or source file can preserve a link, CISOs should separate that additional information organisationally and technically from the transformed data and record who has access; without appropriate separation and controls, practical linkability can remain high. In our analysis, model output should also be assessed separately because it may contain or reveal identifying information. Organisations should consider traceability testing for both input and output before a confidential workflow goes live, so that output-related risks are not overlooked in production.
This analysis is our interpretation, not a ruling by a Dutch supervisory authority. The cited sources offer no guarantee that tokenisation or synthetic data is never traceable, and do not constitute an enforcement decision. For the broader statutory duties that continue to apply to pseudonymised data we refer to the rules and duties for pseudonymisation under the GDPR and to the EDPB test for anonymisation of AI training data. Anyone processing sensitive prompts locally will find additional points of attention in our explanation of why running locally does not by itself protect sensitive prompts. More background is available in the topic hub on AI privacy and GDPR obligations.
Sources and references
Sources: The article draws on the peer-reviewed study in the International Review of Law, Computers & Technology, the EDPB guidance on anonymisation and pseudonymisation, a practical contribution from the European Commission and an analysis in the Computer Law & Security Review.