by Rodessa Marquez

Nearly ten years after the introduction of the GDPR, questions surrounding anonymisation, pseudonymisation, and the limits of data protection law continue to intensify, particularly in the context of medical artificial intelligence. Against this backdrop, colleagues from the Department of Innovation and Digitalisation in Law at the University of Vienna (Selen Yakar, Rodessa Marquez, Ibrahim Sabra, Kseniia Guliaeva, and Alexandra Marginean) organised and facilitated a workshop during CPDP 2026 that brought together legal scholars, privacy experts, industry representatives, and policy practitioners to debate one of the most pressing issues in contemporary data governance: can pseudonymised or synthetic health data ever truly be considered safe from re-identification in the age of AI?

Titled From Pseudonymisation to Synthesis: A Debate on Evaluating Privacy-Preserving Technologies and Re-identification Risks in Medical AI under the GDPR, the workshop explored the implications of the Court of Justice of the European Union’s EDPS v SRB judgment and examined how evolving AI capabilities are reshaping long-standing assumptions regarding identifiability, anonymisation, and acceptable privacy risks under European data protection law.

Rather than following a traditional panel structure, the workshop was organised as an interactive debate exercise centred around a fictional, but legally plausible, scenario involving the secondary use of health data for AI development. Drawing on findings and ongoing work from the PHASE IV AI and COMMUTE projects, the session encouraged participants not only to discuss the legal issues abstractly, but also to engage directly with them through practical argumentation and adversarial reasoning.

The workshop began with a short presentation introducing the research projects and the broader legal and technical questions surrounding privacy-preserving technologies in medical AI. Participants were then introduced to the fictional case scenario and provided with legal instructions before being divided into two groups: one representing Patient P, and the other representing Hospital Y. Each group was tasked with preparing and orally arguing either a complaint or a defence under Article 77 GDPR.

The fictional scenario centred on Hospital Y and its internal biobank, which had collected extensive patient information during the COVID-19 pandemic to support public health research. The data included sensitive health information, neurological symptoms, demographic details, and associated biological samples, all stored in pseudonymised form. In 2022, the hospital partnered with a technology company, BrainX, to develop an AI predictive model intended to assess the future risk of neurodegenerative diseases, particularly Alzheimer’s disease, in individuals previously infected with COVID-19.

To support the project, BrainX used historical real-world health records to generate synthetic datasets intended for AI model training. While synthetic data is commonly presented as privacy-preserving because it artificially reproduces statistical patterns without directly copying individual records, the scenario deliberately raised questions about whether such datasets can ever be fully detached from the real individuals whose data originally informed them. Participants were asked to confront the possibility that AI systems capable of pattern extraction, linkage, and inference may still reveal or reproduce sensitive personal information despite the use of pseudonymisation or synthetic generation techniques.

The legal conflict intensified when Patient P, whose health data had contributed to the research, was contacted regarding participation in a clinical trial linked to the AI system. At the same time, Patient P became aware of media allegations claiming that BrainX possessed internal software capable of re-identifying individuals from pseudonymised records. Although BrainX denied any unlawful re-identification and characterised the software as an internal anonymisation quality-assurance tool, the allegations raised concerns regarding the effectiveness of the privacy-preserving safeguards relied upon by the hospital and the company.

Against this background, participants debated several interrelated legal questions under the GDPR. One central issue concerned purpose limitation under Article 5(1)(b) GDPR and whether the reuse of COVID-19 cohort data for AI model training and clinical trial recruitment remained compatible with the original purpose for which the data had been collected. Closely linked to this was the question of whether the hospital could rely on Articles 9(2)(j) and 6(1)(f) GDPR for the secondary use of the data, or whether a new legal basis, or renewed consent, should have been required.

Another key area of discussion concerned the scope and effectiveness of pseudonymisation in light of the Court of Justice’s reasoning in EDPS v SRB. Participants explored the significance of the Court’s “means reasonably likely” standard for assessing identifiability and debated how contextual, contractual, and technical factors should influence the evaluation of re-identification risks. The discussion highlighted how advances in AI and machine learning complicate traditional assumptions that pseudonymisation necessarily provides a sufficient safeguard against identification.

Synthetic data similarly became a focal point of debate. While some participants argued that properly generated synthetic datasets may fall outside the scope of the GDPR because they no longer relate to identifiable individuals, others questioned whether this distinction remains convincing when synthetic datasets are derived from highly sensitive medical information and generated through models trained on real patient records. The possibility that synthetic datasets could reproduce rare combinations, correlations, or latent traits prompted broader reflection on whether current legal and technical understandings of anonymisation remain adequate in AI-driven environments.

The workshop also raised governance questions regarding accountability and institutional responsibility. Participants debated whether BrainX should be regarded as a processor, an independent controller, or a joint controller alongside Hospital Y, particularly given the collaborative nature of the AI development process and the commercial interests involved. The scenario further prompted discussion on the adequacy of existing oversight structures for public-private partnerships engaged in health AI development, especially where commercial innovation intersects with scientific research and sensitive health data processing.

One of the most valuable aspects of the workshop was the diversity of perspectives brought into the debate. Participants approached the scenario from different disciplinary, professional, and regulatory backgrounds, resulting in competing interpretations of both the technical risks and the legal obligations involved. Some defended the necessity of broad data access to support socially beneficial medical research and innovation, while others emphasised the cumulative privacy risks posed by increasingly sophisticated AI systems capable of inference and re-identification.

The debate format itself proved particularly effective in illustrating the practical tensions embedded within the GDPR framework. Rather than arriving at definitive answers, the workshop exposed the extent to which many foundational concepts in European data protection law, such as anonymisation, compatibility of purpose, and identifiability, remain deeply contested when applied to contemporary AI systems and large-scale health data ecosystems.

Ultimately, the workshop demonstrated that the boundaries between personal data, pseudonymised data, and supposedly anonymous or synthetic data are becoming increasingly unstable in the context of AI-driven medical research. As technical capabilities evolve, so too must the legal and governance frameworks intended to safeguard fundamental rights while enabling scientific progress. Nearly a decade after the GDPR entered into force, the workshop highlighted that many of the Regulation’s most difficult questions are not settled, but instead becoming more complex as AI technologies continue to advance.