Autonomous Weapons Systems and the ai Alignment Problem

Abstract

The challenge in deploying Autonomous Weapons Systems ('aws') is not that it can kill people and destroy objects, but ensuring that it only kills the right people and destroys the right objects. In this paper, we use a hypothetical recently discussed at a military ai conference as a springboard to introduce important dimensions of the 'Alignment Problem' into the discourse concerning aws. This paper will consider important dimensions of what is known as the Alignment Problem, why it is difficult to specify smart goals for autonomous systems, why intelligent systems can pursue dumb goals, and the legal implications for assurance of aws. We begin with some preliminary definitions and conceptual analyses. We then outline the Alignment Problem including introducing the concept of objective functions and rewards. We then turn to an exploration of what the Alignment Problem implies for aws testing, and why apparently simple solutions may not be effective. From here we discuss the implications that the Alignment Problem has for international law applicable to aws, addressing legal obligations relating to the responsibility of states to respect and ensure respect for with international humanitarian law (ihl) and international human rights law (ihrl).