Explainability for Safety-Critical Decision Support: A Case Study on Stakeholders and Method Reliability in Dynamic Alternate Airport Selection

Kurzfassung

The use of artificial intelligence (AI) in safety-critical domains such as aviation increasingly depends on explainable and interpretable methods to support transparency, trust, and certification. However, the reliability and suitability of explanation techniques across different model types and operational contexts remain open questions. This paper examines these issues through an aviation decision-support use case based on the Dynamic Alternate Airport Selection (DAAS) problem, which involves ranking operational alternatives under time pressure and multi-criteria constraints. Linear models, tree-based models, and neural networks were applied to the DAAS task, and feature-importance explanations were generated using both model-specific and general-purpose methods, including SHAP and Integrated Gradients. While model-specific explanations produced stable and domain-consistent results for linear and tree-based models, neural networks exhibited inconsistent and implausible attributions when using general-purpose methods, despite achieving the highest predictive performance. In contrast, permutation-based importance measures yielded more stable and operationally plausible explanations for neural networks. The results highlight the strong dependence of explainability outcomes on model architecture and intended usage context, and emphasize the need to distinguish between explainability for system development and for operational decision support in safety-critical human–AI systems.