login
login
Image header Agence Europe
Europe Daily Bulletin No. 13301
SECTORAL POLICIES / Digital

AI Act’, European co-legislators still looking for solutions to break deadlock on foundation models

Negotiators from the Council of the European Union and the European Parliament parted company on Monday 27 November at the end of another technical meeting, without making any progress on the issue of foundation models in the context of legislation on artificial intelligence (‘AI Act’) (see EUROPE 13300/30).

While the co-legislators maintain the objective of reaching a political agreement during the next round of inter-institutional negotiations (‘trilogue’) – which will take place on 6 December – the situation is still blocked on the issue of foundation models since France, Italy and Germany opposed the approach of the Spanish Presidency of the Council of the EU on this issue (see EUROPE 13297/24).

The latter proposed opting for a gradual approach that would have involved stricter criteria and rules for high-impact foundation models. On the contrary, the three Member States mentioned above would like foundation models not to comply with rules defined in the text, but to be governed by a code of conduct and for future European sanctions not to be imposed in the first instance.

The European Parliament negotiators have left the door open to the introduction of a code of conduct. However, in their view, this should be accompanied by a set of other basic obligations, such as providing suppliers with technical information to ensure that they can comply with the ‘AI Act’.

In addition, MEPs would like models representing systemic risks to be subject to binding rules, particularly in the areas of cybersecurity, evaluation and testing of AI models, and compliance with standards designed to reduce the impact of these models on the environment.

According to the avenues explored by the European Parliament, it could be up to the future ‘Office for Artificial Intelligence’ to examine the technical documentation and evaluate the AI model concerned.

During discussions at the technical meeting on 27 November, the Presidency of the Council of the EU made it clear that it did not intend to go beyond the introduction of codes of conduct.

The Spanish Presidency’s revised mandate to try to relaunch the interinstitutional negotiations on a solid basis will be discussed by the Member States' ambassadors to the EU (Coreper) on Friday 1 December.

In addition, MEPs put forward the possibility of distinguishing between different models according to criteria that could be updated as the sector and technology evolve. Models active in fundamental research could be excluded from the scope of application.

Amnesty International calls on the three Member States not to “compromise” the text

The progress of the discussions – and their deadlock on the question of foundation models – has prompted a number of players from the NGO world and industry to react.

The EU must not waver in the face of this final obstacle. And EU Member States such as France, Germany and Italy must not jeopardise the ‘AI Act’ by giving in to the technology industry's claims that adopting this text would lead to heavy-handed regulation that would stifle innovation”, reacted Amnesty International's Secretary General, Agnès Callamard, on Monday.

It is imperative that France, Germany and Italy stop delaying the negotiation process and that EU legislators focus on ensuring that crucial human rights protections are codified in law before the end of the current EU mandate in 2024”, she added.

On the same day, six organisations from the technology and industry sectors, including BSA-The Software Alliance, CCIA and the Developers Alliance, voiced their concerns about “divergent and unclear scopes and definitions” in a letter to the co-legislators.

The document’s signatories also believe that the obligations inherent in AI foundation models should “take account of ongoing multinational and multi-stakeholder forums and allow for a process of co-regulation”. They are also concerned about proposals to “introduce additional requirements for the use of copyright-protected data to train AI systems, despite the comprehensive framework for copyright protection and enforcement that exists in the EU”. (Original version in French by Thomas Mangin)

Contents

BEACONS
FUNDAMENTAL RIGHTS - SOCIETAL ISSUES
SOCIAL AFFAIRS - EMPLOYMENT
SECTORAL POLICIES
EXTERNAL ACTION
SECURITY - DEFENCE
ECONOMY - FINANCE - BUSINESS
INSTITUTIONAL
COUNCIL OF EUROPE
NEWS BRIEFS
Kiosk