Why Your Smart Glasses Might Be Looking at Your Wall Instead of Your Fingers: How Explainable AI Exposes Shortcut Learning in Spatial Computing

When vision models in spatial computing achieve high laboratory accuracy, they often rely on spurious background features such as plain lab walls rather than true hand joint geometry. This paper explores how multi-layer Explainable AI (XAI) techniques, specifically Saliency Maps, Class Activation Maps (CAM), and Grad-CAM, expose shortcut bias in gesture control algorithms to ensure reliable performance in real-world environments.