France, Germany and Italy circulated, on Saturday 18 November, a joint document in which they call for a review of the approach to foundation models as part of the interinstitutional negotiations (‘trilogues’) on artificial intelligence (AI) (see EUROPE 13293/6).
The three Member States, which had already disassociated themselves from the approach proposed by the Spanish Presidency of the EU Council on this issue in recent weeks (see EUROPE 13291/9), want foundation models not to be subject to rules defined in the text, but to be governed by a code of conduct and for sanctions not to be imposed in the first instance.
“The inherent risks lie in the application of AI systems rather than in the technology itself”, say the three countries, which are calling for a code of conduct and mandatory self-regulation, in line with the principles defined by the G7 members as part of the Hiroshima AI Process (see EUROPE 13282/10).
The legislative work leading to future AI legislation should therefore, in their view, focus on general-purpose AI systems and their uses, rather than on the foundation models themselves.
The code of conduct sought by Paris, Rome and Berlin would entail the publication of technical data sheets containing information on foundation models, the results of bias studies and intended uses. The document also proposes that a European body should be given responsibility for AI governance and that it should draw up guidelines.
In addition, France, Germany and Italy want to give companies some leeway when it comes to possible sanctions. These could only be imposed after repeated breaches of the code of conduct.
This U-turn by the three Member States comes at a time when the issue of foundation models is at the heart of negotiations between the Council of the EU and the European Parliament. The latter, for its part, wants rules to be put in place concerning foundation models and for these provisions to be part of the text in the final agreement.
The possibility of a compromise emerging based on a gradual approach – according to the potential impact of an AI system – therefore seems, for the time being, to be receding.
“Today, the foundation models are part of the text. The situation could perhaps be unblocked by not putting it in the text, by putting it in the hands of the future AI Council, and by implementing acts. But delegating is not the desired scenario”, commented the dossier’s co-rapporteur, Brando Benifei (S&D, Italian), on 15 November.
The issue of foundation models will be discussed by the Member States on Tuesday 21 November at a meeting of the Telecommunications Group of the Council of the European Union. The next technical meeting between the teams from the Council of the European Union and the European Parliament is due to take place on Friday 24 November, prior to a trilogue scheduled for 6 December. (Original version in French by Thomas Mangin)