In the context of the rapid evolution of digital finance, artificial intelligence (AI) models are deeply integrated into critical business processes such as risk assessment, asset pricing, and anti-fraud. The resultant lack of model interpretability has progressively become a significant source of risk, constraining financial stability and public trust. This study aims to comprehensively explore the causes, harms, identification, and governance of AI model interpretability risks. It finds that the interpretability risks of AI models primarily stem from the high complexity of algorithmic structures, implicit biases within data samples, inconsistency between modeling objectives and interpretability regulatory goals, and failure of explanations due to continuous model iteration. Building upon this, the study systematically reveals the multi-layered harms of AI model interpretability risks across four key dimensions: financial stability, social inclusion, legal compliance, and technical security. Concurrently, an identification framework for AI model interpretability risks is constructed, centered on the core methodology of transparency quantification, bias identification, compliance validation, and security detection. Finally, we propose a comprehensive governance system encompassing model engineering optimization, data governance and feature management, multi-party auditing and regulatory coordination, and construction of standards systems and responsibility delineation. This framework seeks to achieve a dynamic balance among technological efficiency, regulatory controllability, and social trust in the collaborative development of digital finance and AI.