Enhancing Human-Artificial Intelligence Alignment: A Calibration-Based Approach

Artificial Intelligence (AI) systems are typically designed to align with ground truth or expert consensus. Complex domains like triage, where ground truth may not exist and experts disagree, create alignment challenges, making it less likely that human users will sufficiently trust and delegate to those AI systems. One approach to mitigating these challenges is to capture and model individual decision-making tendencies so AI can align to the human it assists, gaining the trust necessary for decision delegation. Past work has identified Key Decision-Making Attributes (KDMAs)-measurable traits influencing decisions-that can be codified to enhance AI-human alignment. Traditional KDMA scoring relies on manual values, introducing bias and limiting reliability. Our approach replaces manually assigned KDMA values with data-driven probability distributions derived from real-world decision tendencies. When compared to the manual scoring approach in a triage scenario, our method improves KDMA resolution, enhancing alignment between AI systems and human decision-makers. These improvements were achieved without altering the overall alignment scoring framework, demonstrating our approach's effectiveness.