How the AI Act Frames the Provider’s Role and What It Means for DPIAs

← Alle Beiträge
Künstliche Intelligenz · EN

How the AI Act Frames the Provider’s Role and What It Means for DPIAs

One of the key concepts introduced is the provider: any organisation that develops an AI system or model, or places it on the market or into service under its own name or trademark. This includes well‑known players developing general‑purpose models as well as companies integrating or commercialising AI systems developed by others.

Developing and deploying AI models typically involves at least two types of processing operations: AI development and training (e.g., collecting and pre‑processing data, model training, fine‑tuning, updates), and AI service provision (user‑facing deployment and ongoing operation).

Each of these stages constitute a separate processing activity, potentially involving different roles, controllers and risk profiles. Even models not “designed” to reveal personal data may still retain information within their parameters making genuine anonymisation a high bar to meet.

For organisations, this raises a central question: when is a DPIA required?

Given the nature of AI training, large datasets and evolving use cases, many AI‑related processing operations will meet GDPR high‑risk criteria. In practice, this means companies may need to conduct multiple DPIAs, or consolidate assessments into a structured, lifecycle‑based approach. Model cards, clear identification of processing purposes and early collaboration between engineering, privacy and security teams become essential.

A practical approach includes:

1) Identifying processing operations across both development and deployment.
2) Assessing whether personal data is involved, including the risk of information extraction from the model itself.
3) Evaluating whether a DPIA is required under GDPR and EDPB guidance.
4) Integrating outputs from training‑related DPIAs into service‑provision assessments.
5) Publishing results where appropriate and conducting periodic follow‑up audits.

There is no one‑size‑fits‑all model for DPIAs in the AI context. But there is a clear direction: privacy‑by‑design must start at the earliest stage of the AI lifecycle, not at deployment.

Zu diesem Beitrag gibt es auf LinkedIn ein Bild. Auf LinkedIn ansehen →

Auf LinkedIn diskutieren