In the third draft by the Cyprus Presidency of the Council of the European Union of the revision of the AI Act, dated Thursday 5 March and consulted by Agence Europe, the EU27 begin to consolidate their approach: fixed timetable; less flexibility for artificial intelligence (AI) deployers; framed and specified prerogatives for AI Office, shared responsibilities in terms of AI Literacy and safeguards for real-life testing (see EUROPE 13808/5).
The new Cypriot version does not make any substantial changes, but refines certain points already modified in previous drafts and supports others (see EUROPE 13794/8).
This draft includes a few additional safeguards for the testing phases of AI models in real-life conditions, and retains the timetable for the application of the obligations for high-risk systems: 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems.
The text adds the need for member countries to put in place “at least one regulatory sandbox” for the test phases, to be operational by 2 December 2027 at the latest, again in line with the timetable for the obligations of high-risk systems.
The EU27 are also adding a number of exceptions to the AI models that fall within the scope of the AI Office’s supervisory powers: the high-risk systems in Annex I and III, as well as those “provided by law enforcement authorities, border management authorities, judicial authorities and financial institutions”.
They also mandate the European Commission to be fully and solely “responsible for conformity assessments and pre-market testing of AI systems that are classified as high risk and subject to third party conformity assessment”.
There is no mention, however, of a potential ban on AI models likely to generate deepfakes and non-consensual sexually explicit content. This request from France and Spain does not appear to have been accepted, despite the heated debate on the issue (see EUROPE 13807/6).
This latest Cypriot compromise reverses certain sensitive parts of the European Commission’s initial proposal, striking a more cautious balance between European prerogatives to control AI models and the desire to simplify the legislative framework. (Original version in French by Isalia Stieffatre)