Human-AI Collaboration in High-Stakes Decision-Making Environments

This sequential mixed-methods dissertation investigated human-AI collaboration in high-stakes decision-making environments, examining how situational awareness (SA) should be distributed between human operators and AI systems, factors influencing calibrated trust, effects of AI explanation modalities on team performance, and training intervention effectiveness. Grounded in distributed cognition theory, joint cognitive systems frameworks, Endsley's three-level SA model, and Lee and See's trust in automation framework, this research addressed critical gaps in empirical validation of human-AI teaming configurations for safety-critical contexts. Four complementary studies were conducted: a cross-sectional survey of 512 professionals in aviation, healthcare, and security domains; a factorial experiment with 240 participants manipulating AI capability, explanation type, and training condition; a qualitative field study with 48 participants in operational environments; and a randomized controlled intervention study with 180 participants. Key findings revealed that optimal SA distribution depends on task characteristics and AI capability levels, with individual SA significantly predicting team performance (β = .52, p < .001). Trust calibration emerged as a four-factor construct encompassing perceived competence, trust alignment, uncertainty recognition, and metacognitive monitoring. Level 3 projection-focused SA explanations produced significantly superior team performance compared to perception or comprehension-focused explanations, F(2, 237) = 28.76, p < .001, η²ₚ = .21. The training intervention yielded large effects on trust calibration (d = 0.94) and team performance (d = 0.80), with benefits maintained at eight-week follow-up. These findings advance theoretical understanding of human-AI collaboration while providing evidence-based guidelines for AI system design, organizational training programs, and regulatory frameworks governing AI deployment in safety-critical domains. Keywords: human-AI collaboration, situational awareness, trust calibration, high-stakes decision-making, distributed cognition, explainable AI, mixed methods, safety-critical systems