Further Progress on Clarifying Obligations on AI Foundation Models in the Draft EU AI Act Negotiations
Summary
On November 7, 2023, the Spanish presidency of the European Council proposed a governance architecture for supervising foundation models’ obligations under the draft EU AI Act. Under the proposal, the European Commission would have exclusive powers to investigate and enforce certain foundation models’ obligations, either upon its own initiative or upon receipt of a complaint. To conduct its investigation, the Commission could require providers to give it access to the foundation model. For “high-impact” foundation models, the European Commission would be required, under the proposal, to use adversarial evaluations by internal or qualified external red teams—i.e., groups that pretend to be a threat actor to help organizations improve their operations. Upon a finding of a violation, the European Commission could order the provider to comply with the Act’s obligations or impose sanctions. It is not known whether the proposal would be accepted by the European Commission or European Parliament. However, its use as a compromise text could help the trilogue reach a breakthrough on one of the largest remaining sticking points in the negotiations.
