Should We Trust (X)AI? Design Dimensions for Structured Experimental Evaluations

XAI evaluation lacks standardized methodology across disciplines: The paper reveals that machine learning, human-computer interaction, and visual analytics fields each evaluate explainable AI differently, targeting different parts of the XAI process. This fragmentation means there's currently no unified framework for assessing whether AI explanations are actually trustworthy.

Current XAI studies don't adequately address real-world complexity: The authors identify significant gaps between controlled comparative studies and real-world applications, noting that a "holistic evaluation of XAI in real-world scenarios" is needed but largely missing from current research. This suggests many XAI systems haven't been properly tested in the messy conditions where they'll actually be deployed.

Trust in AI explanations requires systematic bias identification: The paper proposes a conceptual model that characterizes different sources of bias in the XAI process and how explanations build (or fail to build) trust. This suggests that simply providing explanations isn't enough—organizations need to systematically identify where bias can enter the explanation pipeline.

Precise terminology is essential for responsible AI governance: The authors emphasize the need for "precise terminology required for a rigorous study of XAI," implying that vague or inconsistent language about AI explainability has hindered progress. Clear definitions are foundational for developing standards, regulations, and best practices around trustworthy AI.

Significant research gaps remain in evaluating XAI effectiveness: The paper identifies multiple unresolved questions about how to properly evaluate whether explanations actually help users make better decisions or appropriately calibrate their trust in AI systems.