EDPB Anonymization Guidelines 2026: What Changes and What to Do Now

When is data truly anonymous?

That is the question the European Data Protection Board (EDPB) answers in its new guidelines on anonymization (spelled anonymisation in the EU and the UK).

The draft was adopted in July 2026, together with guidelines on web scraping to train generative AI.

Both public consultations closed on October 30, 2026.

Here is what changes, and what to do now.

Why these guidelines matter

Anonymous data falls outside the scope of the GDPR.

Data that only looks anonymous is still personal data, with every obligation that comes with it.

Until now, the reference was Opinion 05/2014 of the Article 29 Working Party, the EDPB’s predecessor.

Guidelines 02/2026 build on it, taking into account twelve years of case law and the rise of AI.

The text is still a draft: the final version may change after the consultation.

3 criteria for anonymous data

The EDPB sets three criteria.

If all three are met, the data can be considered anonymous:

1. No singling out.
   No combination of values points to a single person in the dataset.
2. No linkage.
   No record can be linked, with certainty or high likelihood, to another source about the same person.
3. No inference.
   No specific information that could affect a person’s rights can be deduced about them.

If any of the three is not met, further analysis is needed before calling the data anonymous.

These are the three criteria of 2014 (singling out, linkability, inference), made more precise.

Anonymity depends on the recipient

This is the biggest change.

The EDPB accepts that the same dataset can be anonymous for a recipient and still personal data for the organization that shared it.

It all depends on the means each party can reasonably use to identify people.

This follows the SRB ruling of the Court of Justice of the EU, issued on September 4, 2025.

We explain it in our guide to pseudonymization vs anonymization.

The EDPB offers two ways to reason:

1. The contextual approach.
   You assess each recipient’s actual means: their other data, their tools, their access.
2. The simplified approach.
   You ignore those differences and reason as if anyone could receive the data. It is more cautious: data may be treated as personal even when it would be anonymous for some recipients.

The two can be combined: a simplified first pass, then a contextual analysis where needed.

Key point:
Before saying “this data is anonymous”, you need to be able to say “for whom”.

What AI changes for anonymization

The EDPB brings AI into the picture in two ways.

1. Re-identification gets easier.
   Advances in AI, including agents that can chain searches together, reduce the time and cost of re-identifying people.
2. Models can reveal information.
   Querying a model, or using synthetic data, can allow specific inferences about individuals.

Two practical consequences:

1. Reassess anonymity over time.
   Anonymization that is good enough today may not be in two years.
2. Do not assume model outputs are anonymous.
   Neither AI responses nor synthetic datasets are anonymous by default.

To prepare training data, see how to anonymize data before sharing it with an AI vendor.

Web scraping for generative AI

The second set of guidelines covers web scraping to train generative AI.

The EDPB recalls that the GDPR applies as soon as scraping involves personal data.

It stresses data minimization, reliable sources, and special category data, which is prohibited in principle unless a GDPR exception applies.

For organizations building their own datasets, the message is the same: anonymize what identifies people before training, not after.

Anonymization must be documented

The EDPB recalls that anonymizing is itself a processing of personal data.

It expects three things:

1. Document the process.
   Which data, which methods, which settings.
2. Document the testing.
   How you checked that the three criteria were met.
3. Document the risk assessment.
   For which recipients, assuming which means.

One last point on wording: data that remains identifiable should not be called “anonymous”.

In the UK: the ICO’s guidance

The EDPB guidelines do not apply in the UK.

Since Brexit, organizations there follow the UK GDPR and the guidance of the Information Commissioner’s Office (ICO).

The ICO published its final anonymisation and pseudonymisation guidance in March 2025.

It asks the same core question, whether a person can still be identified, and assesses it with the “motivated intruder” test: could a determined person, without special skills, identify someone from the data?

For organizations working on both sides of the Channel, the practical steps below fit both texts.

In the US: when the GDPR applies

The GDPR reaches beyond Europe.

It applies to US companies that offer goods or services to people in the EU, or monitor their behavior.

It also applies when a US company processes data on behalf of a European client.

For these companies, the EDPB guidelines show how European regulators will judge whether shared data is truly anonymous.

What to do now

Without waiting for the final version:

1. List the datasets you describe as anonymous.
   Exports, test extracts, data shared with vendors, training data.
2. Check them against the 3 criteria.
   Singling out, linkage, inference, including free text, where identities often hide.
3. Specify for whom they are anonymous.
   A vendor with no other source does not have the same means as an internal team with access to the CRM.
4. Keep a record.
   What was detected, what was processed, and when.
5. Schedule a reassessment.
   Especially for data shared over a long period or published.

For the methods themselves, see our guide to PII redaction.

With NymData

NymData detects and anonymizes personal data in your files and databases, on your own servers:

1. Mapping first.
   A scan that changes nothing, to see which personal data is present, including in free text.
2. The method you choose.
   Anonymization by type ([PERSON], [EMAIL]), realistic pseudonymization, or a constant identifier, depending on the recipient.
3. Audit report.
   Every run records what was detected and what was done: a basis for your documentation.

See how NymData works