Calibrated adaptive framework for trustworthy human and artificial intelligence decision systems

Abstract

The integration of artificial intelligence (AI) into safety-critical industrial decision-making promises substantial gains in efficiency and reliability, yet real-world deployment remains constrained by a deeper systemic problem: miscalibrated risk estimates, static trust assumptions, and unmodeled human cognitive biases jointly destabilize collaboration under operational pressure. In predictive maintenance and similar high-stakes settings, overconfident AI recommendations and pressure-biased human judgment can amplify, rather than mitigate, failure risk. Existing hybrid frameworks typically treat AI as a static advisor, neglect epistemic uncertainty, assume calibrated probabilities, and rely on fixed trust models; limitations that undermine robustness under distribution shift and dynamic feedback. We introduce Calibrated Adaptive Human-AI Teaming (CAHAT), a closed-loop decision architecture that integrates four complementary mechanisms: outcome-driven adaptive reliance updating ( \(\theta\) ) grounded in utility theory as the primary adaptation engine, post-hoc probability calibration through temperature scaling, epistemic uncertainty quantification via Monte Carlo Dropout (providing additional robustness under high-uncertainty conditions), and semantic explainability that translates technical outputs into managerial insights. By linking trust directly to empirical performance discrepancies, CAHAT enables dynamic self-correction rather than reliance on heuristics. The framework is evaluated in a physics-informed predictive maintenance simulator that models stochastic machine degradation and production pressure bias in a controlled, reproducible environment. Results demonstrate that adaptive fusion of calibrated AI and cognitively modeled humans outperforms the human-only and uncalibrated AI-only baselines, drastically improving cumulative reward over the human baseline (6,887.6 vs. − 1,982.8) and achieving near-zero failure rates (0.8%) under realistic stress conditions. While a perfectly calibrated AI-only model achieves marginally higher raw rewards in this controlled setting, the adaptive hybrid protocol provides superior robustness under distribution shift and dynamic cognitive bias. Ablation analysis confirms calibration as foundational (uncalibrated AI fails 12.4 % of the time) while sensitivity experiments identify human bias, not model capability, as the dominant failure mode. These simulation-based findings suggest design principles for trustworthy AI in safety-critical domains: quantify uncertainty, calibrate probabilities, and adapt reliance to observed performance. We emphasize that validation with real human operators and operational datasets remains an essential direction for future work. By reframing human-AI collaboration as a dynamically regulated control process rather than static automation, CAHAT provides a simulation-validated pathway toward intelligent, resilient teaming in industrial AI. We emphasize that these conclusions are derived from a controlled, physics-informed simulator; empirical validation with real-world datasets and human operators constitutes an essential direction for future work.