European legislation on artificial intelligence

A problem of legibility and flexibility in the regulation

Two elements deserve particular attention:

• The legal definition of artificial intelligence (AI) systems: is it adapted to the reality and upgradeability of these systems?
• The mechanisms for classifying systems as “high-risk” : the European Commission’s text provides for a pyramidal classification of AI systems based on a risk-based approach. It seems necessary to clarify these criteria and exceptions in the draft regulation (social scoring, real-time remote biometric identification systems, etc.).

In this respect, it would be appropriate to include relevant stakeholders in the procedures for reviewing this piece of legislation.

The multi-stakeholder governance approach must be reinforced

As drafted, the proposal allows Member States a certain degree of flexibility when designating the national authorities responsible for monitoring the application of the regulation. However, it would be desirable to allow for greater harmonisation in the implementation of the regulation within the European Union. In addition, it will be necessary to ensure that the designated authorities are capable of carrying out these tasks. These problems linked to the harmonisation of the implementation of European regulations have already been illustrated since the entry into force of the General Data Protection Regulation (GDPR).

The respective roles of the European Artificial Intelligence Board (EAIB), Member States and the Commission should also be specified. To date, the text seems to introduce several parallel governance systems, without a mechanism for clear communication between them. Finally, it will be necessary to give more autonomy to the EAIB in order to ensure its independence from the European Commission.

The AI Act poses implementation challenges

There are still many unknown variables regarding the assessment of these technologies. Given the diversity of AI systems and of their uses, determining the right scope of analysis is not always easy. At this stage, some of the proposed concepts in the draft regulation do not yet have a methodology for evaluation, nor a settled définition. These difficulties of interpretation will have to be resolved so that the actors at hand are able to grasp these concepts in their evaluation process.

In view of these interpretation challenges, and in line with the necessity to strengthen its role in terms of governance, the EAIB should be tasked with drawing up concrete recommendations in consultation with expert groups, relevant stakeholders, and the actors of the European artificial intelligence ecosystem.

Regulatory sandboxes : a lever for innovation on which the Union should have a strong ambition

The European Commission’s proposal relies on the assumption that the existence of a stable and clear regulatory framework would enable the development of the AI market in the European Union. However, the framework remains complex and will probably not be sufficient in itself to provide an incentive mechanism likely to create a market.

Thus, the regulatory sandboxes represent the “innovation” aspect of this text. However, in order for them to function and be a real lever for innovation, it seems essential to build a harmonised approach between the competent national authorities, and that these authorities dispose of the human, technical, and financial means to put them into action.

In this perspective, the functioning of regulatory sandboxes should be discussed in a collegial manner between the European Commission, the EAIB, national competent authorities, the AI Expert Group and relevant industry and civil society representatives. For now, the functioning of regulatory sandboxes varies from one Member State to another. Besides, regulatory sandboxes are most often an opportunity for data protection authorities (DPAs) to assess some compliance issues, rather than real experimental frameworks aimed to open up to innovation.