Trustworthy Artificial Intelligence in High-Stakes Decision Systems: A Cross-Domain Systematic Review of Explainability, Fairness, Privacy, Robustness, and Governance

Abstract

Artificial intelligence (AI) systems are increasingly entrusted with consequential decisions in finance, healthcare, cybersecurity, public infrastructure, and online platforms. As these systems move from research prototypes to operational deployment, their value depends not only on predictive accuracy but on whether stakeholders can trust them to behave transparently, fairly, privately, robustly, and accountably. This paper presents a cross-domain systematic review of trustworthy AI in high-stakes decision systems, synthesizing 196 recent studies that span financial risk and regulatory technology, clinical and biomedical analytics, network and software security, social-media information integrity, energy and sustainability, and commerce and mobility. We organize the literature along five trustworthiness pillars-explainability, fairness, privacy preservation, robustness and safety, and accountability and governance-and three cross-cutting methodological enablers: multimodal and multi-source data fusion, graph-based learning, and large language models (LLMs) with agentic orchestration. For each domain we characterize the dominant tasks, data modalities, and evaluation practices; for each pillar we summarize representative techniques and their tensions. Our coverage analysis reveals an uneven landscape: explainability and robustness dominate finance and security, privacy concentrates in healthcare and consumer platforms, and fairness remains comparatively underexplored beyond credit scoring. We further discuss recurring trade-offs-privacy versus utility, fairness versus accuracy, transparency versus performance-and the emerging risks introduced by LLM-based and multi-agent systems, including hallucination, memory poisoning, and over-refusal. We conclude with open challenges and a research agenda emphasizing standardized evaluation, lifecycle governance, and the integration of trustworthiness guarantees into agentic AI.