On 8 July 2026 the European Data Protection Board (EDPB) announced its draft Guidelines 02/2026 on anonymisation, an update to the 2014 opinion that businesses have relied on for years. Public comments are open until 30 October 2026. The subject sounds technical, but it decides something every organisation depends on: the point at which data stops being personal data and the GDPR stops applying to it.
If you share analytics with partners, publish statistics, keep “anonymised” customer exports, or train and fine-tune AI on customer data, this draft changes how confidently you can use the word “anonymous”.
What the EDPB actually published
The draft follows the Court of Justice's ruling in EDPS v SRB (C-413/23 P, 4 September 2025), which held that pseudonymised data is not automatically personal data for every recipient. The same dataset can be personal data in one organisation's hands and anonymous in another's. The EDPB's guidelines turn that principle into a working method, and they are candid that the answer “may vary from one entity to another”.
The EDPB announced them together with separate guidelines on web scraping for generative AI, both open for comment until the same 30 October date. We covered the scraping side in our article on GDPR and AI training data. This article is about the other half: when data, or a model, can honestly be called anonymous.
“Anonymous” now depends on who holds the data
Under the draft, data is anonymous if it does not relate to an identified or identifiable person. A person is identifiable if they can be distinguished from others in a specific context using means reasonably likely to be used, judged from the perspective of the “relevant entity” and all objective factors. Three consequences stand out in the law-firm analyses published since July:
- The same data can give different answers. A recipient with no realistic way to re-identify people can treat a dataset as anonymous even if the organisation that produced it still could. That opens practical doors for analytics and research sharing.
- Processors don't change the picture. Commentators read the draft as saying data cannot be personal in a controller's hands but anonymous in its processor's. The controller's perspective governs.
- Contracts are not enough. A clause forbidding a recipient from re-identifying people is treated as a complement to technical measures, not a substitute for them.
The EDPB also offers two ways to run the assessment. The contextual approach looks at what each relevant entity could actually do. The simplified approach ignores those differences. It is more convenient and gives more confidence, but it can treat data as not anonymous even where it would be anonymous for some entities. Ashurst reports that the EDPB suggests starting with the simplified approach and switching to the contextual one when you need a more precise answer.
The three-part test
The EDPB's framework uses three criteria. If all three are met, the data can safely be considered anonymous. If one fails, that does not make the data personal automatically, but it means more analysis is required.
- No record isolation. Nobody can pick out one person's record. Freshfields reports the draft's example that combining “finance department” with “leaves after 8 pm” may single out one employee.
- No linkage. The records cannot be matched with other data about the same person. Birth year, gender and location, for instance, become linkable if a partner's marketing database holds the same attributes.
- No inference. Nobody can draw a specific, meaningful conclusion about an individual, whether from a single record or from aggregate figures.
Zero risk is not the standard. Law-firm summaries describe the test as whether the chance of distinguishing a person is insignificant in practice. Ashurst adds a point that catches many businesses out: the test looks at capability, not motivation, so “nobody would bother” is not a defence.
Where AI changes the answer
Commentators note that the draft acknowledges AI, and agentic AI in particular, is reducing the cost of re-identification. A dataset that is anonymous today may not stay anonymous. That has three practical effects:
- Anonymisation is not a one-off. Re-identification risk typically rises over time, so assessments should be repeated, and a security incident can be a trigger to look again.
- Models can leak too. According to one analysis of the draft, inferences can be drawn by querying or prompting a model, and training data can sometimes be extracted from a supposedly anonymous one, so a model cannot simply be declared anonymous.
- This matches the EDPB's earlier AI position. Its December 2024 Opinion 28/2024 said an AI model trained on personal data is not anonymous in every case, and that both the likelihood of extracting personal data and of obtaining it through queries must be insignificant.
If you fine-tune a model on customer conversations, or feed support tickets into a tool that “learns”, do not assume the output is anonymous because names were removed. Our piece on why an AI chatbot isn't anonymising anything shows how quickly that assumption fails.
What doesn't go away when data is anonymised
- Anonymising is itself processing. Commentators agree the draft treats the act of anonymising as processing that needs a lawful basis under Article 6, an Article 9(2) condition where special-category data is involved, and transparency towards the people concerned.
- Don't over-claim. Calling data “anonymous” in a privacy notice when people can still be identified is a transparency problem, not a marketing detail.
- Mixed datasets are unsettled. Law firms read the draft differently on what happens when only some records can be anonymised: one reads it as making the whole dataset personal data, another as an assessment made record by record. Until the final text, plan on the stricter reading.
There is one piece of reassurance. Ashurst notes that organisations which have already assessed datasets as anonymous under the 2014 opinion do not have to reassess them purely because of this draft.
Editable DPA pack, SAR response toolkit, and Article 30 RoPA register: the records that show what data you hold, why you hold it and how you assessed it. €39–€59 each or €99 for all three.
A practical checklist before 30 October
- List every dataset you call anonymous. Exports, shared statistics, dashboards, machine-learning training sets, test data.
- Name the relevant entities. Who could get the data: the recipient, their partners, your own staff, an attacker? What else could each of them combine it with?
- Run the three tests and write down the result. Isolation, linkage, inference. If you are unsure, start with the more conservative simplified approach.
- Back every contract clause with a technical measure. A no-re-identification promise on its own will not carry the claim.
- Test AI outputs instead of declaring them anonymous. If you fine-tune on customer data, check for extraction and regurgitation, keep the results, and cover it in your DPIA.
- Set a re-check date and consider commenting. Put a review in the calendar, and if this affects your sector, submit feedback through the EDPB's public consultation page before 30 October. Trade associations are the usual route for smaller businesses.
Remember that this is a draft. The final text may change, and nothing here adds obligations beyond what the GDPR and the Court's case law already require. What it does is show how regulators will read a claim of anonymity.
The bottom line
The direction is clear even while the details are open. Anonymity is not a stamp you apply once. It is a conclusion you reach for a particular holder of the data, test against three criteria, document, and revisit as technology improves. Organisations that can show that work will be in a far stronger position than those that simply write “anonymous” on the file.
Book a free consultation with GDPRGard → and we will help you review which of your datasets and AI tools really qualify.
Sources
- EDPB: EDPB sheds light on anonymisation and web scraping for generative AI (8 July 2026)
- EDPB: Opinion 28/2024 on data protection aspects of AI models
- Ashurst: Data Bytes 68, EMEA Data Privacy Update, September 2026
- Freshfields: Anonymous or not? The EDPB's new draft guidelines on anonymisation
- Sidley: EDPB Publishes Draft Guidelines on Anonymisation
- Gaming Tech Law: EDPB Anonymisation Guidelines, What They Mean for AI Systems
Also read:
- The "It Was Public" Excuse Is Over: GDPR Now Formally Covers AI Training Data
- Your AI chatbot isn't anonymising anything
- The EDPB just standardised how every DPA in Europe calculates your fine