Measures of Performance

Literature on autonomous weapon systems (AWS) make frequent reference to measures of performance (MoP) like accuracy, reliability or robustness, often to make normative statements (e.g. regarding the prohibition on indiscriminate weapons). How users of AWS are expected to interpret and use such MoPs, however, is rarely addressed. This is problematic as many indicators, such as accuracy, are measured very differently for AI systems compared to conventional weapons. This chapter explores in depth four MoPs specific to AI—accuracy, robustness, reliability and understandability—that are crucial to consider when making legal decisions concerning the deployment of AWS. For accuracy, a warning is issued to not take the metric at face value, but instead to decompose the metric into Recall and Specificity, representing an AWS’s military utility and ability to discriminate respectively. The importance of understandability from a legal and practical perspective is also highlighted, and juxtaposed to modern AI models which often feature a degree of opacity. For MoPs in general, the chapter concludes that this is an element of Technical Awareness that is largely out of a field commander’s hands: instead, responsibility lies with the Providing Entity to engage in careful and effective development and testing to ensure legal and operational benchmarks are upheld. At the same time, end-users of AWS retain important supporting roles, e.g. by providing diligent reports after use to improve robustness and reliabillity predictions of systems.