Pressed by a tight timetable - the revision of the regulation on artificial intelligence (AI) must, ideally, be completed by 1 August - the EU27 are continuing to fine-tune the text, striking an increasingly subtle balance between European autonomy and national safeguards (see EUROPE 13807/5). The second compromise of the Cyprus Presidency of the Council, which Agence Europe was able to consult, illustrates this approach: a few more restrictions and a few less flexibilities offered to the European Commission and AI deployers (see EUROPE 13794/8).
This second compromise seems to have focused on one rather specific point: the AI Office’s powers to enforce obligations and monitor certain AI systems. Whereas, in the first compromise, the Office “had to adopt” a decision to initiate proceedings in the event of reasonable suspicion, the new text now only allows it the “possibility” of “launching an investigation”.
The automatic obligation is thus replaced by a discretionary power: the AI Office would no longer be required to act on reasonable suspicion, but would have a wider margin of discretion, which could manually reduce the automatic nature of the agency’s intervention.
Several new articles have been added, detailing the procedures for requesting or deciding on access to information, when issued by the AI Office. The EU27 exempt AI providers from an obligation to respond when the request is a “simple request”, but expose them to fines in the event of data being “distorted”.
On the other hand, during potential physical inspections, the Office’s representatives are given extended powers, comparable to those found in competition law, with an extended right of access to the “premises, documents, employees and data” of AI operators.
In addition, the EU27 have substantially amended Article 60a on the real-life testing phase for AI systems. For certain systems to be able to carry out these tests, the member countries concerned will have to put in place precise “regulatory frameworks” to set out the “detailed conditions, requirements, governance and liability provisions”. The aim is to “drastically minimise possible risks to health, safety or fundamental rights”.
The new compromise also introduces a deadline - 2 December 2027 - for the Commission to adopt guidelines on the post-marketing surveillance plan for the systems. The timetable for implementing the obligations for high-risk systems has not changed since the first compromise.
Through these adjustments, the EU27 seem to be seeking to strike a more prudent balance, reducing certain margins for manoeuvre while consolidating European control instruments.
The AI Office, entrusted by the Commission with a great deal of centralised power, is in danger of being deprived of certain freedoms, in exchange for a stronger basis for its investigative work. (Original version in French by Isalia Stieffatre)