Machine Semiology in Practice: Clinician Strategies for Interpreting AI-Generated Visual Explanations

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The increasing adoption of AI-based decision support systems in high-stakes professional settings has intensified interest in eXplainable Artificial Intelligence (XAI), particularly in domains where human accountability remains central. In medical imaging, visual explanation techniques such as saliency maps are widely promoted as a means of rendering opaque models interpretable to clinicians. Despite their popularity, however, little is known about how such explanations are actually interpreted and made meaningful within radiological practice. This study addresses this gap by introducing and empirically examining the concept of machine semiology, defined as the study of the repertoire of signs produced by AI systems and the situated interpretation work through which practitioners attribute meaning to these signs and reconcile them with established professional vision and diagnostic practice. Here, we report a qualitative study examining how clinicians engage with and interpret saliency maps produced by an AI-based decision support system for spinal fracture detection. Based on eight in-depth semi-structured interviews with orthopedists, spine surgeons, and radiologists using real radiographs and Grad-CAM explanations, we find that saliency maps function as underspecified, indexical sign systems that often conflict with radiological semiology. Clinicians performed extensive interpretive work to resolve semantic and pragmatic misalignments, developing provisional “theories of machine” to explain system behavior and errors. These findings suggest that explainability emerges as a sociotechnical accomplishment, with implications for the design of interactive, practice-aligned XAI systems.