From Curiosity to Caution: How Expertise Shapes the Use of Interpretable Machine Learning

Interpretability tools are increasingly used to make ML models more transparent, yet their effectiveness has been limited in practice despite significant work on improving their design. While individual factors (e.g., mental models, cognitive biases) shape tool use, the research community has also highlighted the influence of socio-organizational factors (e.g., job roles, team dynamics, policies). We trace the impact of one such factor that operates at a hierarchical level: proficiency differential, i.e., the development of skill over time as people evolve from novices to experts. To investigate the value of this proficiency differential, we conducted contextual inquiries and semi-structured interviews with expert and novice data scientists (N=23). Our work contributes empirical evidence of how proficiency shapes interpretability use, with novices driven by curiosity and experts by efficiency and caution. We present a framework for understanding expert-novice differences, and identify design implications for interpretable ML that scaffold both expert-like reasoning and novice-like exploration.