The Belgium Data Protection Authority’s recent report, Artificial Intelligence Systems and the GDPR: A Data Protection Perspective, is a timely analysis exploring the interplay of the General Data Protection Regulation (GDPR) and the EU Artificial Intelligence (AI) Act. It seeks to provide insights into where the EU AI Act adds additional compliance requirements to aspects of the GDPR in light of novel challenges posed by AI as well as to recommend high-level strategies to meet these requirements.
However, despite its comprehensive scope, the report falls somewhat short in addressing certain critical distinctions and practicalities essential for a holistic understanding of the compliance obligations emerging from the interplay of the two laws. This article examines the report’s insights and points out some omissions that lead to remaining open questions, particularly regarding the applicability of the GDPR to general-purpose AI, the roles of AI model providers and deployers, data deletion processes, and other nuanced compliance challenges.
GDPR Principles: Transparency, Accountability, Fairness, Lawfulness, Data Minimization, and Purpose Limitation
The report highlights key GDPR principles, underscoring the necessity for transparency, fairness, lawfulness, and accountability when processing personal data, and it provides some insights into how the EU AI Act builds on these principles and applies them in the AI context. It helpfully adds that under the EU AI Act, there are now outright prohibitions of certain AI systems that do not draw on principles relating to the data they’re trained on but rather consider the use of the system to pose unacceptable risks per se. Interestingly, though, all of the prohibited AI practices would involve training on data pertaining to individuals, such as facial images, personality characteristics and behavioral data, sentiment indicators, and biometric data. Note that it makes no difference under the EU AI Act whether training of such systems would be possible on fully anonymized data for the prohibition to apply.
While outlining the GDPR principles of data minimization and purpose limitation, the report nowhere mentions general purpose AI, which poses particularly thorny questions with regard to these principles, given the large amount of data needed to train these models, their general purpose and the difficulty that creates for purpose limitation, transparency, and consent.
Automated Decision-Making
The report addresses the respective requirements for transparency and explainability, human involvement and oversight, and accountability in automated decision-making and explains well the additional obligations the EU AI Act adds to the GDPR. It would have been helpful, however, to provide insights into what these requirements mean in practice for the development and implementation of large language models, whose idiosyncrasies are such that the explainability of decisions these models make cannot always be achieved due to the complexity of the models.
Security Framework: Addressing AI-Specific Risks
The report commendably addresses AI-specific security risks, acknowledging that AI systems introduce unique vulnerabilities not seen in traditional data processing, such as potential bias in training data and susceptibility to data manipulation. The brochure’s user story for a car insurance AI system illustrates this well, outlining additional safeguards like data validation, anomaly detection, and human oversight to maintain data integrity and fairness.
Missing Distinctions: The Roles of Model Providers vs. Deployers
A significant gap in the report lies in its omission to distinguish between the responsibilities of AI model providers and deployers. The AI Act and GDPR assign unique obligations to different actors in the AI ecosystem, yet the report provides only a generalized framework for compliance. Model providers—those who design and train AI algorithms—bear distinct duties related to transparency, risk assessment, and ensuring non-discriminatory training data. Deployers, on the other hand, implement these models within specific operational contexts and are responsible for continuous oversight, data accuracy, and bias monitoring.
The report would benefit from clarifying these roles, as a one-size-fits-all approach fails to account for the technical and operational nuances in the respective environments of these actors. Providers need guidance on integrating model training data requirements with GDPR principles, while model deployers require explicit instructions on what they need to include in information provided to users. A more differentiated approach would ensure that each actor in the AI supply chain can comply with GDPR and AI Act requirements effectively.
Data Deletion in AI Models
The Belgium DPA report barely touches on data retention and deletion obligations and does not address a core challenge: how to delete personal data embedded within AI models. AI systems often use personal data during training, leading to potential residual traces in model weights or outputs. The GDPR mandates that personal data should be deletable upon request, but the technical feasibility of this within AI models remains an area of active research and currently a largely unsolved problem. The absence of guidance on this matter in the report leaves organizations navigating compliance requirements for data deletion with significant uncertainty.
Conclusion: Moving Towards Comprehensive AI Compliance Guidance
The Belgium DPA’s report provides valuable insights into the intersection of AI regulation and data protection. However, to serve as a truly practical compliance guide, the report should address critical gaps in its recommendations, specifically in distinguishing obligations between model providers and deployers, clarifying data deletion methods, and introducing AI-specific security measures. Organizations require actionable, role-specific guidance to navigate the complex regulatory landscape of AI and data protection.
By addressing these areas, future editions of the report could improve its utility for organizations striving to develop and deploy AI responsibly. A more nuanced approach to compliance would benefit both organizations and data subjects, fostering AI applications that are not only compliant but also trustworthy and transparent.
To explore how Private AI can support your organization in complying with GDPR and AI Act requirements, specifically data minimization, try our web demo or get an API key to test de-identification solutions on your data.