EDPB Data Anonymisation Guidelines: New Rules for AI Web Scraping

In a watershed moment for digital privacy, the European Data Protection Board (EDPB) recently enacted a double-pronged regulatory strike that fundamentally redefines the boundaries of digital identity in the machine learning era. During its 122nd plenary meeting on July 7, 2026, the board adopted the draft data anonymisation guidelines (Guidelines 02/2026) alongside a sweeping set of rules targeting automated data extraction (Guidelines 03/2026 on Web Scraping in the Context of Generative AI). Officially published on July 8, 2026, these frameworks establish an aggressive, legally binding baseline for how digital footprints are processed, tracked, and permanently erased across the European Union and beyond.

The timing of these releases is highly strategic. As artificial intelligence developers continue their relentless pursuit of training data, and ad-tech conglomerates deploy increasingly sophisticated profiling techniques, the regulatory gap between technical “de-identification” and legal “anonymity” has widened. By simultaneously raising the threshold for what constitutes true anonymisation and classifying automated web scraping as active personal data processing, the EDPB has effectively boxed in companies that have long operated in the regulatory gray zones. The implications are profound, marking a massive victory for digital sovereignty and providing consumers with a robust legal toolkit to force the total erasure of their online histories.

Decoding the EDPB’s New Data Anonymisation Guidelines

For nearly a decade under the General Data Protection Regulation (GDPR), “anonymous data” has served as the ultimate corporate escape hatch. Recital 26 of the GDPR explicitly states that the principles of data protection do not apply to anonymous information—meaning that if a company can successfully anonymise its datasets, it is entirely free from the law’s stringent administrative, security, and consent-related obligations. However, this loophole has led to widespread industry abuse, with data brokers, analytics firms, and healthcare processors routinely using basic pseudonymisation techniques as a proxy for true anonymity.

The EDPB’s Guidelines 02/2026 on Anonymisation directly target this corporate sleight of hand by drawing an uncompromising, technically rigorous line between pseudonymisation and anonymisation. Pseudonymisation, as defined under Article 4(5) of the GDPR, merely masks direct identifiers (such as replacing a user’s name with a cryptographic hash or a randomized token) while maintaining the structural linkability of the underlying records. The EDPB reiterates that pseudonymised data remains legally classified as “personal data” because the potential for re-identification is preserved, either through the possession of an administrative “key” or through external data correlation.

This clarification builds directly upon the landmark Court of Justice of the European Union (CJEU) ruling in European Data Protection Supervisor (EDPS) v. Single Resolution Board (SRB) (Case C-413/23 P), decided on September 4, 2025. In that case, the CJEU established a “relative” or “contextual” approach to personal data, ruling that pseudonymised information does not automatically constitute personal data for a third-party recipient if that recipient has absolutely no reasonable, legal, or technical means to re-identify the data subjects. While some industry groups hoped this ruling would weaken the GDPR’s scope, the EDPB’s new data anonymisation guidelines operationalize the judgment with extreme caution. The board clarifies that the existence of a link between data and a natural person is often non-obvious and highly dynamic, requiring controllers to perform exhausting, case-by-case threat modeling before asserting that shared data has achieved true anonymity.

The Three-Part Litmus Test and the Dual Assessment Framework

Under Guidelines 02/2026, the EDPB mandates that data is only considered truly anonymous if re-identification is mathematically and technically impossible. To verify this, the board codifies a strict three-part litmus test that organizations must apply to their datasets. A dataset only exits the scope of the GDPR if all three of the following conditions are met:

• No Record Isolation (Singling Out): It must be impossible for any entity to isolate or identify a single individual’s record within the dataset. If a specific user’s behavioral profile can be separated from the crowd—even if their name and IP address are missing—the dataset fails this test.
• No Linkage: It must be technically impossible to link two or more records relating to the same data subject, either within the same database or across disparate external datasets. For instance, if an anonymous advertising dataset can be cross-referenced with a public voter registration file to re-establish a user’s identity, the data remains personal.
• No Inference: It must be impossible to deduce, infer, or predict any information about an individual using contextual clues, behavioral patterns, or proxy variables. For example, if an algorithm can infer a user’s specific medical condition by analyzing a combination of zip codes, age brackets, and purchase histories, the dataset is legally personal data.

To help organizations navigate these rigorous technical hurdles, the EDPB introduces a dual-pathway assessment framework. Under the Contextual Approach, which directly mirrors the EDPS v SRB jurisprudence, controllers must thoroughly evaluate the real-world capabilities, security controls, and technological resources of any recipient who might attempt to re-identify the individuals. This requires ongoing threat modeling and contractual restrictions. Alternatively, for companies seeking maximum compliance certainty and administrative simplicity, the EDPB offers the Simplified Approach. Under this pathway, the controller bypasses recipient-specific nuances and assumes a worst-case scenario. If the data fails any of the three criteria under any plausible condition, it is treated as personal data across the board.

Dismantling the AI Scraper Wild West: Guidelines 03/2026

While Guidelines 02/2026 redefine how data is stored and shared, Guidelines 03/2026 on Web Scraping in the Context of Generative AI target how data is collected. Led by the Swedish Authority for Privacy Protection (Integritetsskyddsmyndigheten – IMY), this framework addresses the systematic, automated harvesting of public digital footprints by generative AI developers seeking to train large language models (LLMs) and neural networks.

For years, AI developers have operated under the assumption that because information is “publicly available” on social networks, blogs, and public directories, it can be scraped with impunity. The EDPB has formally dismantled this assumption, ruling that large-scale web scraping constitutes the active processing of personal data (encompassing collection, organization, retrieval, and storage). Consequently, any scraping activity targeting personal data must strictly adhere to the GDPR.

The guidelines place severe limitations on the legal bases AI companies can use to justify scraping:

• The Failure of Legitimate Interest: AI developers can no longer rely on “Legitimate Interest” (Article 6(1)(f) GDPR) as a blanket justification. The EDPB emphasizes that the balancing test required under Article 6 highly favors the fundamental rights of data subjects. Internet users do not reasonably expect their casual forum posts, personal photos, or professional profiles to be ingested into permanent, proprietary AI models.
• The Special Category Prohibition: Web scraping inherently harvests sensitive personal data (such as political affiliations, religious beliefs, health details, or sexual orientation), which is strictly prohibited under Article 9(1) GDPR. The EDPB clarifies that simply publishing this information on a semi-public web platform does not satisfy the exception of “manifestly making data public” (Article 9(2)(e)) for the purpose of AI training.
• Pre-, During-, and Post-Scraping Safeguards: AI firms are now legally obligated to implement robust technical measures. Before scraping, they must employ state-of-the-art filters to exclude sensitive domains. During scraping, they must utilize real-time classifiers to detect and discard personal data. Post-scraping, they must execute rigorous sanitization and de-identification pipelines.

Importantly, these obligations do not merely apply to the developers operating the scraping bots. The EDPB rules that any enterprise or third-party developer that purchases or acquires pre-scraped, pre-packaged training datasets from external vendors inherits full liability as a data controller. This establishes an unbroken chain of compliance, forcing the entire AI supply chain to police its inputs.

Reclaiming the Right to Be Forgotten: Digital Footprint Erasure

The synergy between these two guidelines represents a massive paradigm shift for digital footprint erasure. By combining a near-impossible threshold for anonymisation with a strict regulatory ban on unconsented AI harvesting, the EDPB has stripped data brokers, ad-tech platforms, and AI conglomerates of their favorite legal shields. Under the new frameworks, unless a digital footprint has been rendered entirely unlinkable and mathematically impossible to re-identify, the individual retains their absolute Right to Erasure (Article 17 GDPR).

Historically, when users submitted deletion requests to data brokers, they were routinely met with evasive responses claiming that their profiles were “anonymised” or “aggregated,” thereby exempting the company from compliance. Under the new data anonymisation guidelines, these claims are legally indefensible unless the company can prove it meets the three-part litmus test across all contextual vectors. If any residual capability to reconstruct a user’s identity or isolate their behavior exists, the information remains personal data, and the company is legally obligated to execute a complete and permanent deletion upon request.

Furthermore, these rules empower individuals to proactively opt-out of generative AI training databases. Users can demand that their scraped digital footprints be purged from active training pools, and AI developers can no longer hide behind the complexity of their model architectures as an excuse for non-compliance. Both Guidelines 02/2026 and Guidelines 03/2026 are open for public consultation until October 30, 2026, allowing stakeholders a brief window to offer feedback. However, the regulatory trajectory is clear: the European Data Protection Board has officially signaled the end of the data harvesting Wild West, putting the power of digital self-determination back into the hands of the individual.