XSPLAIN: XAI-enabling Splat-based Prototype Learning for Attribute-aware INterpretability

Prototype-based explanations improve user trust in AI systems: A user study with 51 participants found that people strongly preferred XSPLAIN's "this looks like that" explanations (selected 48.4% of the time as best) over existing methods (p<0.001), suggesting that grounding AI decisions in representative examples makes systems more transparent and trustworthy.

Interpretability can be achieved without sacrificing accuracy: XSPLAIN maintains the original classification performance while providing explanations, demonstrating that transparency and model performance are not inherently in conflict—a key consideration for deploying AI in critical domains.

Current explanation methods fail for advanced 3D technologies: Existing explainability approaches designed for other 3D representations produce ambiguous saliency maps that don't capture how 3D Gaussian Splatting actually works, highlighting the need for interpretability methods tailored to specific AI architectures rather than one-size-fits-all solutions.

Interpretability barriers limit adoption of high-performing models: The paper identifies lack of interpretability as a key obstacle preventing 3D Gaussian Splatting—a state-of-the-art reconstruction technology—from being used in critical applications, illustrating how explainability gaps can prevent beneficial AI systems from being deployed responsibly.