GDPR for generative AI services

Mapping starts at the interface and continues through the API, model provider, vector database, logs, observability tools and feedback systems. User data, data of people mentioned and data generated through inference must be distinguished.

The provider’s setting not to use prompts for training is important but does not resolve retention, security access, transfers or data stored by other components.

The controller, processor or, where relevant, joint controllers are established for each purpose. The commercial agreement alone does not determine the role if the facts show different control.

• Providing the response and managing the account.
• Retaining conversations and personalising the experience.
• Safety, abuse prevention and incident investigation.
• Evaluating and improving the model or service.
• Training or fine-tuning using user interactions.
• Analytics, marketing and correlation with other sources.

A DPIA may be required where processing is likely to result in high risk, particularly for sensitive data, monitoring, systematic assessment or decisions with significant effects. An AI Act assessment does not replace a DPIA.

Measures may include prompt filtering, redaction, tenant isolation, restricted retention, extraction testing, plugin controls, human review and prohibiting output as the sole basis for sensitive decisions.

The procedure must locate conversations, RAG documents, profiles and logs relating to the requester. Rectifying a false output does not always mean changing model parameters; source corrections, filters, blocking reuse and informing the person may be needed.

For erasure, operational data, backups, datasets and potentially the model are checked. The response must explain technical and legal limits without generically refusing because ‘AI cannot forget’.