| Abstract: |
Deep neural networks (DNNs) have achieved state-of-the-art performance in numerous high-stakes domains, yet their opacity undermines trust, accountability, and compliance with emerging regulations on automated decision-making. Explainable artificial intelligence (XAI) has emerged as a critical discipline for rendering the internal reasoning of deep models comprehensible to human stakeholders without sacrificing predictive performance. This paper presents a systematic review of past research on designing XAI frameworks for transparent and ethical decision-making in DNNs. It synthesizes the existing body of work spanning ante-hoc interpretable architectures, post-hoc attribution methods, surrogate models, concept-based explanations, counterfactual reasoning, and human-centred evaluation protocols. A meta-analytic perspective is adopted to compare findings across studies, to identify methodological inconsistencies, and to distill convergent evidence regarding the faithfulness, stability, and usability of explanation techniques [1][2]. The review further examines the ethical dimensions of XAI, including fairness auditing, bias detection, accountability, and the risks of misleading or manipulable explanations. Drawing on a critical analysis of past work, the paper identifies persistent gaps, notably the absence of standardized evaluation benchmarks, the trade-off between model performance and interpretability, and the limited integration of XAI into operational deployment pipelines, alongside the scarcity of longitudinal field studies. A consolidated framework is proposed that couples explanation generation with quantitative validation, ethical auditing, and human oversight. The review concludes with a research agenda for building transparent, ethically aligned DNN systems that engender stakeholder confidence in real-world applications, and it closes with a research agenda prioritizing standardized evaluation, robust validation, and human-centred design. |
