The ‘omnibus’ process to simplify digital legislation has officially been launched, and is already shaping up to be a confrontational one. The Cyprus Presidency of the EU Council has drafted an initial compromise on the part relating to the Regulation on Artificial Intelligence (AI Act), based on feedback from the Member States, a compromise which is due to be discussed on Friday 30 January by the EU Council’s Antici Group on Simplification (see EUROPE 13784/5).
The document, which Agence Europe was able to consult, reveals that the EU27 are very cautious about the changes proposed by the Commission. A number of key measures have been restricted or even scrapped in this first text, which augurs well for an intense legislative battle that will go far beyond the technical details.
Clear cuts in the Commission’s ambitions. This first compromise reinstates the obligation to register in the European database for suppliers of AI systems that are not considered to be high-risk. This obligation was removed by the Commission in its proposal, in a bid to simplify procedures.
The Member States are also reviving the concept of “strict necessity” in the new Article 4a, which allows suppliers of high-risk AI systems to process special categories of personal data and which has been heavily criticised (see EUROPE 13786/15).
They also amend the second paragraph of the same article, detailing that all providers and deployers of AI systems may exceptionally process special categories of personal data insofar as “processing is strictly necessary to ensure bias detection and correction in view of possible biases that are likely to affect the health and safety of persons, have a negative affect fundamental rights or lead to discrimination prohibited under Union law, especially where data outputs influence inputs for future operations and all of the conditions and safeguards set out in paragraph 1 are applied”.
In addition, the EU Council removes the possibility for the Commission to trigger the application of the obligations for high-risk AI systems prematurely, maintaining the original timetable of 2027 in the interests of legal certainty.
Finally, they clarify the role and powers of the AI Office, detailing its prerogatives and room for manoeuvre within the text itself, through the article 75a, entitled “investigations and coercive measures concerning AI systems supervised by the AI Office”.
In so doing, the EU Council removes the possibility for the Commission to introduce them by means of a delegated act. However, it leaves open the need to publish an implementing act to “further define the enforcement powers (...), including its ability to impose penalties, such as fines or other administrative sanctions”, for non-compliance.
By unravelling several of the pillars of the Commission’s proposal, this first compromise shows that the Member States are noticeably uneasy about the general approach of the ‘omnibus’ on AI. Presented as a simplification exercise, it actually opens up a delicate political sequence, far from a purely technical adjustment. (Original version in French by Isalia Stieffatre)