Data storytelling meets interpretable machine learning: Decoding AI decisions for non-experts without revealing sensitive data and model details

As automated decision-making systems become ubiquitous, the 'black-box' nature of high-performance AI models creates a significant barrier to trust for non-experts. While Interpretable Machine Learning (IML) methods like SHAP provide technical insights, they are often too complex for general users to interpret effectively. This study addresses this gap by proposing a framework that integrates data storytelling with IML to translate technical model outputs into clear, relatable narratives without exposing sensitive input data or proprietary model architectures.

The authors introduce a comprehensive methodology grounded in Design Science Research (DSR). This includes the DIST (Decision-Interpretation-Storytelling-Trust) Pyramid, which maps the progression from raw model output to user trust. To operationalize this, the study develops a reference workflow that uses an 'And-But-Therefore' (ABT) narrative structure. This structure is chosen for its simplicity and logical coherence: the 'And' sets the context, the 'But' introduces the problem or conflict (such as a 'what-if' or 'why-not' query), and the 'Therefore' provides the resolution based on model insights.

The researchers implemented this workflow using the Boston Housing dataset and an XGBoost model. By applying differential privacy to sensitive features and using Large Language Models (LLMs) to generate narratives based on SHAP explanations, they created automated 'what-if' and 'why-not' stories. An empirical study with 379 participants demonstrated that these data stories were significantly more comprehensible and useful than traditional SHAP visualizations. Participants consistently rated the narrative-based explanations higher in accessibility, contextualization, and narrative coherence.