Opinion 28/2024 on Data Protection Aspects of AI Models
European Union · Guideline · European Data Protection Board
Our summary
The European Data Protection Board issued this opinion to address how data protection laws apply when organizations use personal data to develop and deploy artificial intelligence systems. The guidance examines key privacy considerations and legal obligations that arise in AI model contexts. It responds to a request from the Irish supervisory authority.
Written by the Index in our own words. It is not legal advice; the official text is the authority.
Inside this policy
Every section of this policy in order. Our notes summarise each one in plain language and do not quote the official text, which is the authority.
Drafted by the Index An automated summary of the official text, not yet reviewed by a person.Reviewed Checked by a person against the official text.
12 sections
Jump to a chapter
1.1 Summary of facts
1.2 Admissibility of the Request for an Article 64(2) GDPR opinion
The EDPB considers the request admissible as a matter of general application under Article 64(2) GDPR, addressing systemic and novel issues about applying personal data, lawfulness and legitimate interest concepts to AI model development and deployment. The matter also meets the condition of producing effects in more than one EEA member state, as AI models impact many data subjects across the region and SAs have raised concerns. The EDPB found the request reasoned according to its Rules of Procedure and determined that it had not previously provided an opinion or answers on the same questions, making the request admissible for analysis under Article 64(2) GDPR.
2.1 Scope of the Opinion
This opinion provides a framework for supervisory authorities to assess specific cases involving anonymity of AI models, controllers' demonstration of legitimate interest as a legal basis, and whether unlawful processing in development affects subsequent processing, without attempting complete coverage of all scenarios, while preserving all controller and processor obligations, particularly the accountability principle requiring controllers to demonstrate compliance with processing principles. The opinion does not analyse in detail special categories of data under Articles 9(1) and 9(2), automated decision-making under Article 22, purpose compatibility under Article 6(4), data protection impact assessments under Article 35, and data protection by design under Article 25(1), all of which remain relevant considerations for authorities assessing AI model processing. The opinion is informed by prior WP29 guidelines on automated decision-making and profiling, and on data protection impact assessments, as well as relevant CJEU case law.
2.2 Key notions
The opinion defines first-party data as personal data collected by the controller from data subjects, third-party data as personal data obtained from a third party or via web scraping, and clarifies that development phase covers all stages before model deployment including data collection and training, while deployment phase covers all stages after development including operations. Training refers to the stage where models learn patterns from data.
2.3 AI models in the context of the Opinion
AI models are components of AI systems that learn from data through training processes and are determined by algorithmic elements such as layer types, weights, and activation functions; they include deep learning models such as neural networks, large language models, and generative AI. This opinion covers AI models developed through training, particularly those resulting from training with personal data.
3.1 On the nature of AI models in relation to the definition of personal data
Personal data is defined as anything relating to an identified or identifiable person, and according to Recital 26 GDPR, data protection does not apply to anonymous information, meaning data that was never linked to an identifiable individual or personal data made anonymous so individuals are no longer identifiable. The EDPB addresses whether an AI model resulting from training on personal data should always be considered anonymous.
3.2.1 General consideration regarding anonymisation in the context at hand
3.2.2 Elements to evaluate the residual likelihood of identification
3.3.1 General observations
personal data is retained in the model and is subsequently processed by the same controller (for instance in the context of the deployment of the model)
Where a controller unlawfully processes personal data (violating Articles 5(1)(a) and 6 GDPR) to develop an AI model that retains information about identifiable persons, and then the same controller processes this model subsequently, supervisory authorities must conduct case-by-case analysis considering whether development and deployment phases constitute separate purposes. When subsequent processing relies on legitimate interest, the unlawfulness of initial processing must factor into the legitimate interest assessment and may affect whether the subsequent processing is lawful.
personal data is retained in the model and is processed by another controller in the context of the deployment of the model
When personal data embedded in an AI model is processed by a different controller during deployment, both the controller that developed the model and the controller that deploys it must ensure their respective processing is lawful and demonstrate compliance. Supervisory authorities should assess whether the deploying controller conducted adequate checks to verify that the model was not created through unlawful processing, considering factors like whether the source data involved a breach or whether prior infringements were established by authorities; the depth of this assessment may vary based on the type and magnitude of risks associated with processing the data subjects' information during model deployment. When assessing whether a controller lawfully used legitimate interest as a basis for deploying an AI model, supervisory authorities must consider whether the model's development involved unlawful processing, evaluating technical safeguards like access restrictions built into the model and legal factors such as the nature and severity of the initial unlawful processing.
then ensures that the model is anonymised, before the same or another controller initiates another processing of personal data in the context of the deployment
Supervisory authorities can intervene regarding unlawful personal data processing during AI model development and the anonymisation process itself. If the deployed model truly does not process personal data, the GDPR would not apply to its operation and the initial unlawful processing would not affect it, though mere claims of anonymity are insufficient and must be verified case-by-case. When personal data is subsequently processed during deployment after anonymisation, the GDPR applies to those activities and their lawfulness is independent of the initial development.
Controls
What this policy requires, forbids, grants or penalises, one control at a time and in our own words. Each control links to the official text, which is the authority. This is not legal advice.
Drafted by the Index Drafted automatically from the official text, not yet reviewed by a person.Reviewed Checked by a person against the official text.
Themes covered 0 of 35
Choose a theme to see only its controls. About the themes
Risk and scope 0 of 4
- Risk classification, not covered
- Prohibited practices, not covered
- Risk management system, not covered
- Impact assessment, not covered
Data and privacy 0 of 4
- Data governance and quality, not covered
- Data subject rights and consent, not covered
- Cross-border data transfer, not covered
- Copyright and training data, not covered
Transparency and records 0 of 4
- Transparency and disclosure to users, not covered
- Labelling of synthetic content, not covered
- Technical documentation, not covered
- Record keeping and logging, not covered
Technical safeguards 0 of 4
- Human oversight, not covered
- Accuracy, robustness and cybersecurity, not covered
- Bias and non-discrimination, not covered
- Testing and evaluation, not covered
Assurance and lifecycle 0 of 4
- Conformity assessment and certification, not covered
- Registration and notification, not covered
- Incident and breach reporting, not covered
- Post-market monitoring and corrective action, not covered
Rights and protected groups 0 of 4
- Automated decision-making rights, not covered
- Children and vulnerable groups, not covered
- Biometric identification and surveillance, not covered
- Whistleblowing, complaints and redress, not covered
Roles and organisation 0 of 6
- Supply chain and provider-deployer duties, not covered
- General purpose and foundation models, not covered
- Accountability and governance roles, not covered
- Training and AI literacy, not covered
- Public sector use and procurement, not covered
- Innovation support and sandboxes, not covered
Enforcement and sectors 0 of 5
- Enforcement, penalties and regulators, not covered
- Sector rules: health, not covered
- Sector rules: finance and insurance, not covered
- Sector rules: employment, not covered
- Sector rules: education, not covered
The full list of controls could not be loaded, so only the first 0 are shown.
Change history
No changes recorded since we began monitoring this policy.