Trustworthy and Explainable AI Across Critical Sectors: From Medical Diagnosis to Cyber-Physical and Business Systems

Abstract

Artificial intelligence is increasingly deployed in critical decision environments, healthcare, assistive technologies, cyber-physical systems, agriculture, business analytics, cybersecurity, and distributed infrastructure, were inaccuracy, opacity, or unreliability may cause severe harm. While predictive accuracy has driven much of the field's progress, it is now broadly recognized as insufficient: trustworthy AI must also be explainable, robust, privacy-preserving, secure, fair, accountable, and practically deployable under real-world constraints. This structured critical review synthesizes using a six-axis taxonomy comprising critical sector, data modality, architecture family, explainability function, and trustworthiness concern. The review identifies six critical sectors, healthcare and biomedical AI, human-centered and assistive AI, cyber-physical systems and infrastructure, agriculture and sustainability, business and enterprise analytics, and cybersecurity and distributed intelligence—and eight architecture families, from conventional machine learning and CNNs through vision transformers, graph neural networks, Bayesian physics-guided models, generative AI, and federated learning systems. Synthesis reveals that while explainability mechanisms—visual, attentional, post-hoc, and knowledge-structured, are increasingly integrated across sectors, they are rarely validated against formal standards or shown to support trustworthy human oversight in deployment. Key gaps include explanation validation protocols, cross-sector benchmarking, privacy-preserving inference at scale, uncertainty quantification, and governance-aligned reporting. A structured research agenda is proposed that prioritizes validated explainability, federated deployment, robustness under distribution shift, fairness, and evidence maturity across all critical sectors.