Test and Evaluation for Artificial Intelligence
Laura Freeman
Abstract
ABSTRACT Incorporating artificial intelligence (AI) leveraging statistical machine learning (ML) into complex systems poses numerous challenges to traditional test and evaluation (T&E) methods. As AI handles varying decision levels, the underlying ML needs confidence to ensure testable, repeatable, and auditable decisions. Additionally, we need to understand failure modes and failure mitigation techniques. We need AI assurance–certifying ML and/or AI algorithms function as intended and are vulnerability free, either intentionally or unintentionally designed or inserted as data/algorithm parts. T&E provides a process for AI assurance. This article highlights existing test and evaluation methods, the key challenges embedded‐AI exacerbates, and themes based for how T&E will evolve to provide AI system assurance.