Reconciling the AI Value Chain with the EU’s Artificial Intelligence Act

The EU Artificial Intelligence Act (AI Act), proposed by the European Commission in April 2021, is an ambitious and welcome attempt to develop rules for artificial intelligence, and to mitigate its risks. The current text, however, is based on a linear view of the AI value chain, in which one entity places a given AI system on the market and is made accountable for complying with the regulation whenever the system is considered ‘high risk’. In reality, the AI value chain can present itself in a wide variety of configurations. In this paper, in view of the many limitations of the Act, we propose a typology of the AI value chain featuring seven distinct scenarios, and discuss the possible treatment of each one under the AI Act. Moreover, we consider the specific case of general-purpose AI (GPAI) models and their possible inclusion in the scope of the AI Act, and offer six policy recommendations. First, the AI Act should discourage application programming interface access for GPAI use in high-risk AI systems, in order to avoid cases in which providers place systems on the market that they cannot fully observe, let alone control. Second, the AI Act should envisage soft commitments for GPAI model providers to strengthen legal certainty and reduce transaction costs. Third, for high-risk AI applications, the AI Act should discourage value chain types in which a vendor builds software for a specific intended purpose that includes code for training machine learning models, but does not provide the data itself or pre-trained AI models (‘software with AI model’ in our typology). Fourth, the AI Act should explicitly exempt the placing of an AI system online as free and open-source software. Fifth, there is a need to clarify ambiguities concerning the identity and obligations of the providers of high-risk AI systems (PHRAIS) in several of the common business models discussed in the AI value chain typology. Sixth, the proposals made in the European Parliament’s IMCO-LIBE draft report, adding specific new user requirements, should be incorporated into the final text of the AI Act.