A Generalisable Alignment System (GAS) for Humanartificial Intelligence (AI) Alignment

Current developments in the world of Artificial Intelligence (AI) seek to create AI systems that behave in line with human decision-making preferences. Such alignment is thought to increase trust and delegation to AI in high-stakes decision-making situations. A central component of this question is what we align an AI to. Specifically, what factors within an individual predict the decisions that they will make. This is especially important when looking at decisions that do not have “right” answers and where even trusted experts may disagree. In this study we present a Generalizable Alignment System (GAS) based on an underlying framework that governs human choice. Using data from 234 members of the United States Armed Forces we present a framework of choice preference and then demonstrate how individual scores on this framework predict decision-making. The implications for AI design, alignment and trust are then discussed.