The Czech Presidency of the Council of the EU will present to the Member States, on Wednesday 20 July at a meeting of the EU Council Working Party on Telecommunications, the second compromise proposal on the Artificial Intelligence (AI Act) (see EUROPE 12992/12).
Firstly, a number of important changes have been made to the definitions. On this point, the new version of the compromise text retains the basic concepts of the OECD definition while refocusing it on “systems developed through machine learning techniques and knowledge-based approaches”.
This change of approach follows concerns expressed by a number of Member States that the definition proposed by the previous French Presidency of the EU Council was “too broad and ambiguous” and that the distinguishing criteria were not sufficient to distinguish AI systems from some “conventional” software.
The Commission’s powers reviewed
In addition, the text also extensively reviews the powers of the Commission. The latter’s power to update the definition of AI systems via delegated acts has been removed and replaced by recitals clarifying what is meant by “machine learning” and “logic- and knowledge-based approaches”.
However, the Commission would retain the power to adopt implementing acts to “specify and update” techniques under machine learning and logic- and knowledge-based approaches.
The new version of the compromise text also largely revisits the classification of high-risk AI systems. Here again, a number of Member States had expressed doubts about the proposed method of classification, considering that the use of overly broad terms could lead to AI systems that do not pose a risk to fundamental rights being included.
To address these concerns, the Czech Presidency of the Council of the EU has focused on the introduction of a new ‘horizontal level’ to complement the high-risk classification of systems, based on the provisions issued by the High-Level Expert Group on AI and the OECD framework for classifying AI systems.
This framework states that “the significance of the AI system’s output relative to the decision or action taken by a human being and the immediacy of the effect should also be taken into account when classifying AI systems as high risk. The significance of the output of an AI system would be assessed based on whether or not it is purely accessory in respect of the relevant action or decision to be taken by the human”.
Here, the Commission would be empowered to adopt implementing acts to specify further the “purely accessory nature of the output across the relevant high-risk AI systems”, the document says.
A stronger role for the European AI Board
The Czech Presidency of the EU Council has also reviewed the list of high-risk use cases. The Commission could add high-risk use cases to the list via delegated acts, but could also delete them.
The issue of governance had also come up repeatedly in the discussions, both in the EU Council and in the Parliament (see EUROPE 12991/35). Some Member States were concerned that excessive decentralisation to the national level would undermine the proper implementation of this legislation due to the lack of expertise and capacity of national authorities.
To address this issue, the Czech Presidency of the Council proposes to strengthen the role of the European Artificial Intelligence Board so that it is better able to support Member States in the implementation of the text. Its composition has been reviewed and the European AI Board would only include representatives of the Member States. A permanent sub-group, with expert input, would also be established.
Finally, several clarifications have been made regarding the exclusion of national security from the scope of application. While a majority of Member States welcomed this decision, several stated their interest in providing clarity on what is excluded from the scope and what is not. In concrete terms, the text proposes, among other things, that the obligations should not apply where military, defence or national security activities are concerned.
Member States will be asked to provide their comments and recommendations on this second compromise proposal by 2 September.
See the document: https://aeur.eu/f/2p0 (Original version in French by Thomas Mangin)