Advanced artificial intelligence models for skin lesion classification often suffer from performance disparities when applied to images of patients with darker skin tones, largely due to underrepresentation of dark skin tone images in training datasets. In this study, we investigate this issue by evaluating a previously proposed explainable framework, MultiExCAM, trained on the widely used ISIC2018 dataset. We test its performance on Pipsqueak, a previously proposed dataset composed by skin lesion images on dark skin tones. As expected, we observe a significant drop in classification performance when the model is applied to Pipsqueak. To better understand the source of these failures, we employ explainable artificial intelligence techniques to visualize and analyze the model’s decision-making process on both datasets. Our results highlight clear differences in attention patterns and decision rationale, revealing how the lack of dark skin tone representation in the training data leads to poor generalization and biased behavior. This work emphasizes the critical role of explainable analysis in exposing and understanding model bias in clinical applications, and the necessity of inclusive datasets for fair and reliable skin lesion classification.