This study examines the explainability requirements for artificial intelligence (AI) examination support tools used in substantive design examinations from an explainable AI (XAI) perspective. Substantive examination combines prior design search, similarity of articles, overall aesthetic impression, and ease of creation. Hence, AI-generated similarity scores, candidate rankings, and visualized similarity-contributing regions must support, rather than replace, legal judgment. Drawing on the Design Protection Act, Design Examination Guidelines, Korean AI legislation, recent design and AI policy research, and domestic and international materials on XAI and trustworthy AI, this study identifies three discrepancies between: image similarity and overall aesthetic impression; search objects and units, and the legal relevance of retrieved candidates; and AI similarity-contributing regions, and legally significant or dominant features. As such, it argues that explainability should not require the full disclosure of model internals. Instead, it should be designed as a procedural structure which enables examiners to connect AI outputs with legal criteria, and review and record case-specific input scope, search objects and units, recommendation reasons, similarity-contributing regions, and decisions to accept, modify, or reject AI outputs. The proposed framework includes differentiated logging, quality management, and retraining controls.