Explainability, Transparency, and Accountability in AI Systems

<p><span>Background</span><b><i><span>.</span></i></b><span> Explainability, transparency, and accountability have evolved from research aspirations to binding regulatory requirements for AI systems deployed in high-stakes domains. Article 86 of the EU AI Act, clarified by the political agreement on the Digital Omnibus reached on 7 May 2026, establishes a binding right for affected individuals to an explanation in high-risk AI decisions. The Court of Justice of the European Union's judgment in Case C-203/22 has confirmed the right to explanation under Article 15(1)(h) of the General Data Protection Regulation. The technical landscape of explainable AI, anchored by methods including LIME, SHAP, and the DARPA Explainable AI programme, has matured substantially but continues to face structural challenges in producing explanations that are simultaneously faithful, comprehensible, and actionable.</span></p> <p><span>Purpose</span><b><i><span>.</span></i></b><span> This paper synthesises the explainability, transparency, and accountability literature applicable to deployed AI systems, examining post-hoc and intrinsic explainability methods, adversarial attacks on explanation and fairness mechanisms, regulatory transparency requirements and their operational feasibility, accountability architectures including audit trails and human oversight, and the specific challenges of explaining agentic AI behaviour as distinct from single-prediction explanation.</span></p> <p><span>Approach</span><b><i><span>.</span></i></b><span> The paper adopts a narrative literature review methodology drawing on authoritative primary sources, including foundational explainable AI research (Ribeiro, Singh, &amp; Guestrin, 2016; Lundberg &amp; Lee, 2017; Gunning et al., 2019), recent agentic explainability research, the EU AI Act with the May 2026 Digital Omnibus implementation timeline, the EDPB Opinion 28/2024, the CJEU judgment in Case C-203/22, and operational guidance from ISO/IEC 42001 and the NIST AI Risk Management Framework.</span></p> <p><span>Findings.</span><span> Three findings are advanced. First, post-hoc explainability methods, including LIME and SHAP, are operationally available but do not guarantee the faithfulness, stability, or comprehensibility of their explanations, and are themselves subject to adversarial manipulation. Second, regulatory transparency obligations under the EU AI Act Article 86, the GDPR, and equivalent frameworks impose operational requirements that current explainable AI technology cannot fully satisfy, creating an implementation gap that organisations must bridge through governance practice. Third, explaining agentic AI behaviour requires explaining sequences of actions and their justifications, rather than single predictions, a frontier for which current explanation methods are inadequate.</span></p> <p><span>Implications</span><b><i><span>.</span></i></b><span> Practitioners deploying high-risk AI systems require integrated explainability architectures that combine technical XAI methods, audit logging sufficient to support post-hoc investigation, human oversight that meets the meaningfulness requirements of Article 14 of the EU AI Act, and continuous monitoring designed to detect explanation degradation and adversarial manipulation. The paper proposes an explainability control architecture mapping technical, governance, and regulatory dimensions for practitioner use.</span></p>