Two Gaps That Need to be Filled in Order to Trust AI in Complex Battle Scenarios

Abstract

In human terms, trust is earned. This paper presents an approach on how an AI-based Course of Action (COA) recommendation algorithm (CRA) can earn human trust. It introduces a nine-stage process (NSP) divided into three phases, where the first two phases close two critical logic gaps necessary to build a trustworthy CRA. The final phase involves deployment of a trusted CRA. Historical examples are presented to provide arguments on why trust needs to be earned, beyond explaining its recommendations, especially when battle complexity and opponent surprise actions are being addressed. The paper describes discussions on the effects that surprise actions had on past battles and how AI might have made a difference, but only if the degree of trust was high. To achieve this goal, the NSP introduces modeling constructs called EVEs. EVEs are key in allowing knowledge from varying sources and forms to be collected, integrated, and refined during all three phases. Using EVEs, the CRA can integrate knowledge from wargamers conducting tabletop discussions as well as operational test engineers working with actual technology during product testing. EVEs allow CRAs to be trained with a combination of theory and practice to provide more practical and accurate recommendations.