Negotiators from the Council of the EU and the European Parliament met again on the evening of 24 October to attempt to reach a political agreement on the future legislation on artificial intelligence (the AI Act) (see EUROPE 13263/15). While progress has been made, the co-legislators will meet again in just over a month, on 6 December, for what they hope will be the final round of inter-institutional negotiations (‘trilogue’).
For example, the co-legislators addressed the central issue of general-purpose AI systems and foundation models, which should be subject to enhanced transparency requirements. Work has also focused on defining more precise criteria for differentiating and categorising general-purpose systems and foundation models.
These criteria could also be used to define whether an artificial intelligence system could potentially be considered ‘high-impact’. The definition of these criteria could include the computing power of a system or the amount of training data required.
A number of other points were also discussed in this fourth trilogue. In particular, the discussions enabled the negotiators to agree on the exemptions available to suppliers of AI systems who themselves agree that the system in question is risk-free.
As such, the compromise proposes that these systems should improve the result of an action completed by a human or, in the context of a decision-making model, that they should not “replace or influence the human evaluation previously carried out”. (Original version in French by Thomas Mangin)