Grad-CAM for Vision Transformers: A Systematic Taxonomy and Audit of Methodological Ambiguity in Explainable AI

Explainability methods lack standardization: Most of the 175 papers applying Grad-CAM to Vision Transformers don't fully explain their technical choices, making it difficult to understand or reproduce how these "explanation" visualizations are actually created—a critical gap when these tools are used to justify AI decisions.

Different architectural designs require different explanation approaches: Vision Transformers work fundamentally differently from older convolutional neural networks (using tokens and attention instead of spatial feature maps), yet researchers often treat explanation methods as simple plug-and-play tools without accounting for these architectural differences, potentially producing misleading visualizations.

Hidden methodological choices affect interpretation: The paper identifies multiple decision points (feature locations, gradient targets, spatial reconstruction, and aggregation methods) that researchers often leave implicit, meaning two papers using "Grad-CAM" on the same model could produce different visualizations and reach different conclusions about why the model made a decision.

Reproducibility and rigor are at risk: When methodological choices aren't clearly documented or justified, other researchers cannot reliably reproduce results or verify whether an AI system's explanations are actually trustworthy—undermining efforts to audit and validate AI systems before deployment.

Explainability tools need clearer standards: The paper calls for a more systematic approach to adapting explanation methods for new AI architectures, suggesting that responsible AI development requires explicit documentation of technical choices rather than treating explanation methods as interchangeable across different model types.