After a total of 32 hours of discussions since Wednesday, exhausted negotiators from the Council of the European Union and the European Parliament were still pulling out all the stops on Friday 8 December, at the time of going to press, to reach an Interinstitutional agreement on the EU legislation governing artificial intelligence ('AI Act'). A number of points had been cleared up the previous day, but a balance still had to be found on certain issues, such as prohibited practices (see EUROPE 13309/3).
Negotiators from the Council and Parliament spent the day, interspersed with technical meetings and examinations of the various proposals in separate rooms, trying to reach agreement on the safeguards for law enforcement uses of remote biometric identification.
At the end of these negotiations, the European Parliament would like to see a ban on this practice, while the Council of the EU would like more freedom in this area.
Progress has also been made on bans, but a balance still needs to be struck as the European Parliament pushes for a substantial list of banned practices. The Council of the EU could make concessions on this issue, but wants to give as little ground as possible when these practices touch on security issues.
Provisional agreement on foundation models
The day before, after an initial session of 22 hours of negotiations – between Wednesday 6 and Thursday 7 December – the representatives of the European Parliament and the Council had managed to resolve, at least provisionally, a number of points, including the thorny issue of foundation models.
The approach adopted in this area would be close to the proposal made by the European Commission when it tried to find a compromise between the gradual approach favoured by the Spanish Presidency of the EU Council – with stricter rules for high-impact models – and the French, German and Italian wishes, which argued in favour of replacing rules with codes of conduct. These three Member States also wanted sanctions not to be imposed in the first instance (see EUROPE 13297/24).
In accordance with the text proposed by the Commission, the codes of conduct have also been included. However, these would complement the existing rules, not replace them.
The provisional agreement on foundation models provides for all models to be subject to requirements, such as the obligation to publish the summary of training data, or the fact that all AI-generated content must be easily identifiable.
Under the approach based on the impact of AI models, models representing a higher risk would be subject to a battery of additional requirements, including model assessment, monitoring of systemic risks, cybersecurity provisions and, as the European Parliament wanted, the obligation to publish a report on the energy consumption of the AI models in question.
For the time being, a balance has also been struck on the European AI Office. This would come under the European Commission and could have its own budget line. It would be made up of the relevant national competent authorities, and a group of experts would be tasked with advising on the implementation of EU law and warning of risks. (Original version in French by Thomas Mangin)