The Anonymisation of Personal Data in AI Models: EDPB Opinion 28/2024

1. Introduction

The European Data Protection Board (EDPB) recently issued Opinion 28/2024, a crucial document clarifying how personal data is handled in AI models under the General Data Protection Regulation (GDPR). This opinion, prompted by a request from the Irish Data Protection Commission (DPC), seeks to establish guidelines on AI model anonymisation, the legitimacy of processing personal data, and the legal implications of unlawfully obtained data in AI training. It follows a broader discourse on data protection and artificial intelligence, intensified by recent discussions and research on model memorisation risks.

2. AI Models and Anonymisation Under GDPR

A central question posed by the DPC concerns whether AI models qualify as anonymous under GDPR. The EDPB’s response underscores that anonymity is not a binary determination but rather depends on specific circumstances. If an AI model retains or reconstructs personal data used during training, it may not be considered anonymous, thereby falling within the scope of GDPR. The opinion highlights two key considerations:

1. AI models explicitly designed to recall specific personal data from training datasets cannot be deemed anonymous. This includes cases where developers integrate features that allow retrieval of identifiable information.
2. AI models inherently absorb training data within their mathematical parameters. If a model retains personal data in a way that allows extraction through queries, it remains subject to GDPR. However, the likelihood of such extraction varies, requiring case-by-case assessment.

In contrast to the stance taken by the Hamburg Commissioner for Data Protection and Freedom of Information (HmbBfDI), which suggested that LLMs do not necessarily contain personal data due to lack of direct linkage to individuals, the EDPB takes a more nuanced approach. It considers both probabilistic and query-based extraction risks in determining anonymisation status.

3. Assessing the Legal Basis for AI Processing

Another fundamental issue addressed in Opinion 28/2024 is whether the development and deployment of AI models can be justified under the legitimate interest basis outlined in GDPR. The EDPB provides a structured three-step test:

1. Purpose Assessment: The processing must serve a legitimate interest, such as enhancing cybersecurity or providing user-assistance tools.
2. Necessity Test: The AI model must strictly require personal data for its functioning, ensuring that alternative methods without personal data usage are not viable.
3. Balancing of Rights: The impact on individuals’ rights and freedoms must not outweigh the interests pursued by the data controller. Public awareness of data use, the nature of the controller-subject relationship, and mitigating measures must be considered.

This approach acknowledges that AI-driven services can be beneficial but mandates stringent safeguards. Importantly, the EDPB clarifies that AI models trained on publicly available personal data are not automatically exempt from compliance. The context of collection, user expectations, and intended future uses must be factored into the legitimacy analysis.

4. Consequences of Unlawful Data Processing in AI Models

One of the most debated aspects of Opinion 28/2024 concerns the fate of AI models developed using unlawfully processed personal data. While conventional wisdom suggests that outputs derived from illegal inputs remain tainted, the EDPB introduces a potential exception: if a model has been sufficiently anonymised post-processing, its deployment may still be considered lawful.

However, anonymisation is inherently contextual, meaning that an AI model’s compliance status can fluctuate depending on the entity controlling it. This interpretation offers data controllers a possible remediation pathway but does not absolve them of accountability. Supervisory authorities retain discretion to impose corrective measures tailored to each infringement scenario.

5. AI Model Compliance in Regulated Sectors

The implementation of AI models in financial institutions, such as banks and insurance companies, introduces additional complexities due to stringent regulatory requirements. Opinion 28/2024 underscores that financial operators cannot merely rely on assurances from AI model providers but must independently verify compliance. The key responsibilities for these institutions include:

5.1 Verification of the Legal Basis

Financial institutions must ensure that AI-driven data processing aligns with one of the legal bases outlined in GDPR Article 6, and in cases involving special categories of data, with Article 9. The chosen legal basis depends on the purpose of data use, whether it be customer profiling, fraud detection, or automated credit scoring. Importantly, these assessments must be continuously reviewed and documented, in accordance with GDPR’s accountability principle.

5.2 Organisational Safeguards

The use of AI models requires robust governance structures, including:

• Defining clear rules on model selection, compliance evaluation, and data processing.
• Comprising legal, technical, and compliance experts to oversee AI deployment.
• Conducting audits, obtaining compliance certifications, and enforcing contractual transparency obligations.

5.3 Continuous Monitoring and Verification

Compliance is an ongoing process. Financial intermediaries must:

• Regularly reassess the legal basis for data processing.
• Conduct performance evaluations to ensure AI model accuracy and fairness.
• Monitor re-identification risks, as evolving data aggregation techniques may compromise initial anonymisation efforts.
• Implement structured escalation procedures and recordkeeping for compliance assessments.

6. Operational Considerations and Compliance Risks

The EDPB’s opinion calls for structured governance frameworks to integrate AI risk management within financial institutions. Non-compliance can lead to severe penalties under GDPR Article 83, including fines of up to €20 million or 4% of global turnover. Risk areas include:

• Lack of a valid legal basis for data processing.
• Failure to implement sufficient technical and organisational safeguards.
• Inadequate response mechanisms to data subject complaints.

It could also be added that by proactively aligning AI deployment with regulatory requirements, financial institutions can transform compliance into a competitive advantage. In fact, ethical AI use not only ensures adherence to GDPR but also fosters consumer trust.