Clean or simplistic? Testing a lower granularity algorithmic explainer for users’ comprehension, trust, and preference

The growing demand to reduce the opacity of algorithmic systems has led to the development of numerous Explainable Artificial Intelligence (XAI) methodologies that provide clear and coherent explanations for specific algorithmic predictions. In several real-life domains, the need for rapid explanation motivates the use of models, such as AcME (Accelerated Model-agnostic Explanations), that significantly improve time efficiency in both computation and comprehension compared with reference models. However, the lower granularity of their output visualizations can be perceived as simplistic, negatively impacting their perceived trustworthiness among experts. In this study, we evaluated AcME against a benchmark explainer (SHAP, SHapley Additive exPlanation) for users’ comprehension, trust and preference. We conducted two studies with expert users (N = 146) and collected both quantitative and qualitative data via the Qualtrics survey platform. The results suggest that AcME's explanations are understood, and their lower granularity does not compromise their trustworthiness relative to the benchmark. Furthermore, users’ preferences split evenly between the two methods, with AcME preferred for its clean layout and SHAP for its rich information.