To achieve an adequate respect of substantive rules regarding AI systems in Europe, any legislative framework needs to rely on a good enforcement architecture. In other words, an enforcement authority needs to be responsible to ensure the good application of the legal rules. It needs to be equipped with sufficient enforcement powers, staff, resources and independence to fulfil its task. A central guiding principle for the adequate enforcement of AI rules via an EU agency is something that I call AI-vigilance. Such a principle, which is inspired from pharmaceutical terminology (pharmacovigilance), regroups in essence all activities dedicated to the detection, assessment, understanding and prevention of adverse effects or any other AI-related problem for humans. AI-vigilance will help ensure the safety and non-discrimination of AI systems that are used in the EU market and at the same time guides the institutional choices for the enforcement framework, such as ex ante authorization procedures for specific AI systems.
This paper argues for the establishment of a strong enforcement authority, e.g., an EU Agency for Algorithms (EUAA) that will unite enforcement powers in relation to the regulation of AI systems and algorithms. It is inspired by the existing regulatory models used in competition law and medicines law. However, rather than entrusting different regulatory agencies, such as competition authorities , data protection authorities with powers in relation to the regulation of algorithms, it is suggested to combine regulatory powers in one dedicated agency that disposes of the relevant expertise in AI systems and algorithms. As can be seen in the risk-based approach underlying the EU AI Act , algorithms can pose risks for European citizens, notably when they predict or decide in professional or private situations. But at the same time, algorithms can be beneficial and AI innovations can facilitate work and life of humans. Bridging those two opposing side effects of AI systems needs to be at the heart of any EU regulatory agency. Such a dichotomy of beneficial impacts and negative side effects is a well-known phenomenon of drugs and medicines, which is managed within Member States and the EU within the framework of an authorisation procedure for new drugs and medicines. Due to these similarities between AI systems and drugs or medical devices, such an EU agency for algorithms shall be built on the regulatory model used for pharmaceuticals by the European Medicines Agency (EMA).
This paper will briefly set the context of the current elaboration of substantive rules on AI regulation at European and international level, notably at European Union and Council of Europe level with a particular focus on gender equality and non-discrimination. It aims at bringing substantive rules to the surface aimed at ensuring non-discriminatory algorithms (I.). On this basis, current proposals for the enforcement structure of the EU AI Act will be analysed and the advantages and disadvantages of different existing and suggested models will be discussed. In line with the analysis and findings of the present paper, it will be argued to use the European Medicines Agency as a model to design the EU agency for algorithms (II.). Building on the various arguments, the next section will give recommendations in relation to the design of a possible future EU agency for algorithms and sketch out in more detail its cooperation between the EU and Member States (within a system of shared competences between the EU and Member States), potential tasks, resources and staff as well as an illustrative organigram (III.). Finally, due to the global nature of AI, the creation of an EU agency will be positioned within the global context of the potential creation of an UN AI agency, which could be envisaged following the ongoing establishment of the UN multi-stakeholder High-level Advisory Body on Artificial Intelligence (IV.).