Regulating Non-High-Risk AI Systems under the EU’s Artificial Intelligence Act, with Special Focus on the Role of Soft Law

Abstract

The remit of this research paper is twofold. First, it provides a concise overview of how the EU’s Artificial Intelligence Act regulates non-high-risk AI systems-that is, systems deemed by the Act as not presenting a high or unacceptable risk. Second, the paper elucidates the role that the Act assigns to codes of practice and codes of conduct for governing these systems. In tackling the first-mentioned task, the paper draws up a typology consisting of four types of non-high-risk AI systems: (1) those that appear to be high-risk systems but are actually not high-risk (type 1 systems); (2) those presenting limited risk due to their potential for deceiving or manipulating humans (type 2 systems); (3) general-purpose AI systems (type 3 systems); and (4) those that are otherwise low risk (type 4 systems). Regarding the role of codes of practice and codes of conduct, the paper highlights how the Act assigns the former type of code different fields of application and different degrees of legal effect than the latter type of code. Whereas codes of practice are reserved for governing type 2 and type 3 systems, and are probably compulsory for type 3 systems (at least as a ‘stop-gap’ measure in advance of the adoption of harmonised standards), codes of conduct are purely voluntary and capable of governing all non-high-risk systems (i.e. types 1 to 4). Additionally, codes of conduct are envisaged as governing all AI systems (also high-risk systems) in respect of certain matters specified in the Act. A key argument advanced in the paper is that codes of practice may be regarded as instruments of meta-regulation that are truly embedded in the Act, whereas codes of conduct are simply instruments for potential meta-regulation. As such, codes of practice will likely play a much more crucial role under the Act than will codes of conduct. Extending this argument, the paper predicts that the meta-regulatory potential of codes of conduct will likely be undercut by the Act’s failure to provide a compelling incentive structure for their creation. This will weaken, in turn, the Act’s ability to promote various considerations that are relatively peripheral to its central thrust yet societally important-such as AI literacy, participatory design, gender equality and environmental sustainability. Hence, the paper rounds off with several suggestions for strengthening the traction of codes of conduct under the Act.