Interdependence Analysis for Artificial Intelligence System Safety

Engineers responsible for evaluating tactical and weapons systems for system safety will need a new approach for evaluating emerging artificial intelligence (AI)-enabled systems, since these systems leverage machine learning (ML) techniques. For many reasons, ML algorithms are often difficult to diagnose for safety purposes. For instance, they did not lend themselves easily to codebase inspections, thus necessitating the reduction in "autonomy" of the ML-enabled component. By modifying Interdependence Analysis (IA) techniques, a more rigorous approach to evaluating AI/ML-enabled weapons can be found. The IA process produces a rigorous exploration based on observability, predictability, and direct ability, highlighting the key requirements that encapsulate all interactions between human and machine. This paper explores using IA to define the interaction requirements for human–machine teaming, employs those results to identify key critical functions, and leverages those findings to reveal how "autonomy" reduction might be employed.Prepared for the Naval Postgraduate School, Monterey, CA 93943.Naval Postgraduate SchoolApproved for public release; distribution is unlimited.Approved for public release; distribution is unlimited