Trustworthy AI Systems

Abstract

Trustworthy AI -- AI systems that are reliable, safe, fair, explainable, privacy-preserving, and accountable -- has emerged as the central framing concept for responsible AI governance in Europe and beyond. The EU AI Act, NIST AI RMF, and ISO/IEC 42001 each operationalise trustworthiness through different dimensions and requirements, creating a complex compliance landscape for AI developers and deployers. This study develops and applies a comprehensive Trustworthy AI Assessment Framework (TAAF) covering seven trustworthiness dimensions -- reliability and robustness, safety, fairness and non-discrimination, explainability, privacy, accountability, and societal wellbeing -- to evaluate twelve deployed AI systems across four high-risk application domains: medical diagnosis support, credit scoring, recidivism risk assessment, and autonomous vehicle perception. Assessment instruments include automated technical tests (adversarial robustness, bias metrics, calibration, differential privacy audits) and structured expert review (legal compliance, ethical reasoning, stakeholder impact analysis). No evaluated system achieves passing marks across all seven dimensions. Medical diagnosis systems score highest on reliability (8.2/10) but weakest on explainability (4.8/10). Credit scoring systems show the largest fairness deficits (demographic AUC gaps up to 0.084). Recidivism assessment tools fail accountability requirements in 8 of 12 cases. A quantitative Trustworthiness Index (TI) is proposed and validated against practitioner expert ratings (Spearman rho = 0.82). Policy recommendations for regulators and practitioners are provided.