With high-stakes industries such as healthcare, finance, and autonomous systems, an increasing number of cognitive artificial intelligence are being utilized, which presents new challenges for developing calibrated trust. The complexity of trust involves balancing reliance and skepticism. Mistrust and misalignment cause complacency towards automation, premature skepticism, and outright rejection of the automation systems. Previous literature describes trust via isolated theoretical perspectives (cognitive load theory, expectancy-disconfirmation theory, algorithmic fairness). But these analyses examine trust in fragments and overlook important multidimensional dynamics, as well as failing to sufficiently quantify the principles of human-centered design and advanced ensemble architecture to an underdeveloped extent for capturing nonlinear sociotechnical phenomena. This research presents the first synthesized and integrated model to examine trust using a psychological, organizational, and computationally grounded approach. The integrated model uses engineered human-centered metrics such as Trust Stability Index (TSI), Bias Penalty Factor (BPF), and Cognitive Stress Aggregate (CSA), combined with rigorous mathematical formulations and advanced Sobolev-space functional analysis. Of 16 models analyzed with varying architecture and ensemble methods, the Stacking Ensemble achieved the best performance (R² = 0. 9489, RMSE = 2. 06, MAE = 1. 54). Overall analysis of the model yielded excellent generalization, a high level of noise tolerance (96. 98% retention under moderate perturbation), and model diagnostics, suggestive of the TSI model as a critical data quality constraining factor. The results of the analysis validate the first instance of integrated socio-technically engineered features with meta-learning, providing robust and meaningful trust calibrated human-AI collaboration predictive solutions to high-stakes industries.
Ramamurthy et al. (Fri,) studied this question.