Draft Commission Guidelines on the qualification of AI systems as high risk
The European Commission has published draft Guidelines in 2026 clarifying the criteria for determining when an IA system is to be considered high risk under Article 6 of the IA Regulation ("RIA"). The non-binding document takes a practical approach and aims to facilitate a homogenous application of the regime by providing legal interpretation and examples of use.
Key points:
- Risk-based approach: the RIA uses a classification model based on the potential impact on security and fundamental rights, reserving high risk for cases with greater potential for harm.
- Two ways of classification: a system will be high risk either because of its integration into regulated products or security functions (Art. 6.1 RIA), or because of its use in the specific cases in Annex III (Art. 6.2 RIA).
- Relevance of the purpose of the system: the classification depends on the intended use as defined by the provider, whereby the way the system is presented and marketed is decisive.
- Functional concept of IA system and safety component: only systems that fit the definition of the RIA can be classified, and in the scope of Annex I it is key whether they fulfil safety functions or generate risks in case of failure.
- Closed list of high-risk areas: Annex III exhaustively includes sensitive sectors (biometrics, employment, essential services, public sector, etc.), with classification conditional on the specific use case.
- Limited exclusion mechanism: certain systems may be excluded from high risk if their function is merely ancillary or not decisive, unless they involve profiling or materially influence decisions.
- Irrelevance of human control in the classification: human intervention does not change the risk rating, although it is relevant for regulatory compliance.
- Initial responsibility of the provider: classification is in the first instance the responsibility of the provider, without prejudice to subsequent control by supervisory authorities.
These Guidelines are interpretative in nature and may be updated, in particular by amending the cases in Annex III.
Article written by the TMT area of ECIJA Madrid.
Links: