Designing Targeted Explanations for Heterogeneous Stakeholders - AI for Collaborative Use in the Railways

The opaqueness of AI continues to represent a significant challenge when designing safe and reliable high risk-AI-systems, particularly for collaborative use by multiple heterogeneous stakeholders. Explainability approaches are used to address this opacity by increasing transparency and interpretability of AI-based systems and thereby establishing human control over the AI-system. In order to be effective, these explainability approaches need to be targeted to needs of all involved stakeholders. We show with an example use case from the railway sector how work design can be combined with technology design for successful human-system integration. Through observation of work processes and interviews with stakeholders at an early stage of system development we can support the design of targeted explanations for high-risk AI systems used in complex collaborative settings by gaining in depth understanding of the respective work systems. We identified five relevant explainability approaches for a visual train inspection decision support system. Exploratory testing then showed preferences for some explanations over others, fostering the discussion on choosing appropriate and effective explanations for visual tasks and informing the design of a final visual train inspection interface prototype for future research and application in the railways.