Artificial Intelligence (AI) is transforming how humans do everything from getting to work to diagnosing illnesses to creating art. In all these applications, AI occupies a gray area between a tool (like a calculator) and a partner (like a colleague). AI is more than a tool because AI systems have goals, dynamically plan actions to achieve those goals, and adapt to the situation based on experience. However, humans can’t build the relationships with an AI system that they do with a trusted colleague. AI’s undeniable value in high-stakes, life-or-death decisions coupled with AI’s status as more than a tool but not yet a partner raises fascinating questions about how and how much humans should trust AI systems. These questions are especially critical for the training and simulation community, given its leading role in the deployment of AI.
This tutorial will review the science of trust across both the social and physical sciences and describe the three key aspects of AI trust: trustworthy, trustable, and trusted. Prominent theories and models of trust will be discussed and consideration of those applied throughout the human-AI lifecycle will be explored. Approaches to assessing AI trustworthiness will be explained including their relation to the DoD’s existing VV&A process. The technical requirements the AI system must meet to be capable of gaining a human’s trust will be detailed, including explainability, transparency, natural interaction and building common ground. Subjective and objective (behavioral and physiological) trust measurement approaches will be explained. All of this will surmount to a final discussion of human-AI trust calibration and the future of human-AI trust centered on the realm of the possible for standards (e.g. TRL equivalent for trust of a system, trustworthiness index for AI operational fielding decision). The tutorial addresses researchers, developers, and evaluators who create or use artificial intelligence. No technical knowledge is required.
Keywords
AI, AUTONOMY, CHARACTERIZING SYSTEM PERFORMANCE , COLLABORATIVE, EMERGING TECHNOLOGIES, MACHINE LEARNING