Briefing to the Air Force Scientific Advisory Board: T and E Contributions to Avoiding Unintended Behaviors in Autonomous Systems
[Technical Report, Briefing Charts]
INSTITUTE FOR DEFENSE ANALYSES ALEXANDRIA VA
Pagination or Media Count:
Testers rely on making valid inferences about a systems performance in untested situations in order to evaluate and certify the system. However, when systems are black boxesas is frequently the case for AI-enabled technologiesthese inferences are not possible. Avoiding unintended behaviors in these systems, however, requires us to make valid predictions about behavior. This means we must have models of system decision makingan understanding of what causally drives systems to make one decision over another.