A letter from the black box: interpreting artificial intelligence through example-based Twin systems

Deep learning models have achieved impressive performance across domains such as computer vision, natural language processing, speech recognition, and predictive analytics. However, their decision-making processes remain largely opaque, functioning as “black boxes” that hinder transparency, interpretability, and trust—particularly in high-stakes settings like healthcare, finance, and autonomous systems. To address this limitation, we propose a post-hoc explanation framework based on the Twin-Systems paradigm, which pairs a high-performing deep learning model with a case-based reasoning (CBR) system. While the deep model ensures strong predictive accuracy, the CBR component provides intuitive explanations through similar examples, enabling more human-understandable interpretations. We evaluate our approach across multiple data modalities, including tabular, image, and text data, using representative architectures such as multi-layer perceptrons, convolutional neural networks, and transformer-based models. We further assess different feature-weighting strategies to examine their ability to align CBR explanations with the black-box model’s reasoning. To enhance interpretability, we integrate feature activation maps to visually highlight influential input regions. Experimental results demonstrate that the proposed Twin-Systems framework effectively improves transparency and supports the development of more interpretable, trustworthy, and user-centered explainable artificial intelligence systems for real-world decision-making.