Degen's XAI content model defines 14 categories for AI explanations
Six user studies across industries yield a reproducible framework for explaining AI decisions
Helmut Degen's new human-centered explanation content model addresses a fundamental gap in XAI: what specific content should local, post-hoc explanations actually contain? Using a hybrid inductive-deductive qualitative content analysis on 325 meaning units from six user studies spanning building technology, manufacturing, AI software development, and hospital cybersecurity, Degen identified 12 initial codes. Theory-informed coverage assessment and expert review then added two additional codes—Rule base and What-if backward—resulting in a final 14-code model organized into four groups: rule-based explanations, causal explanations, epistemic (actual), and epistemic (similar).
An eleven-member expert panel confirmed content adequacy across all codes (I-CVI ≥ 0.82; scale-level agreement of 0.93 for relevance, 0.92 for boundary clarity, and 0.94 for understandability). A stratified subsample of 82 units coded independently by two researchers achieved Krippendorff's α = 0.920 and Cohen's κ = 0.920, demonstrating high reproducibility. While behavioral validation of downstream effects remains future work, the model is immediately applicable for elicitation, specification, and evaluation of explanation content in industrial AI systems. The paper is currently under submission to the International Journal of Human-Computer Studies.
- 14-code explanation model derived from 325 meaning units across 6 user studies in multiple industrial domains
- Expert panel validation achieved I-CVI ≥ 0.82 and scale-level agreement of 0.93 for relevance, 0.92 for boundary clarity
- Intercoder reliability of Krippendorff's α = 0.920 and Cohen's κ = 0.920 on a 25% stratified subsample
Why It Matters
Gives AI teams a standard, validated framework for designing explanations that users actually need in high-stakes industrial contexts.