AI Safety: Performance Metrics And Failure Modes

Abstract

Aim of this work is to update the safety assessment in order to extend the actual safety rigor to Artificial Intelligence (AI) systems. The results are based on the European Union Aviation Safety Agency (EASA) publications, part of the EASA AI Roadmap (Ref. 1), and the starting point of the entire discussion is the Concept Paper: First usable guidance for level 1&2 application (Ref. 2). In the guideline, the Anticipated-MOC-SA-01-7 asks to specify a link between AI performance metrics and safety assessment: this work shows a direct connection between the performance metrics of an AI binary classification algorithm and its Failure Mode, allowing to provide the link required but also to set safety requirements on the performance metrics that will guide the AI development.