Characterizing Human Explanation Strategies to Inform the Design of Explainable AI for Building Damage Assessment

Humans use diverse explanation strategies that AI systems should mirror: The study identified six major strategies that people naturally use when explaining damage assessments from satellite imagery, suggesting that effective XAI systems should be designed to match these human communication patterns rather than imposing AI-centric explanation formats.

One-size-fits-all XAI approaches may not serve high-stakes tasks well: The research demonstrates that existing XAI techniques often lack grounding in task-specific human needs, which is particularly problematic for critical applications like disaster response where incorrect damage assessments could affect resource allocation and emergency response decisions.

Human input remains essential for visual detection tasks: The study reinforces that fully-automated approaches to damage detection are "unlikely to be perfectly safe and reliable," indicating that human-AI collaboration—supported by appropriate explanations—is necessary for responsible deployment in high-stakes scenarios.

Understanding human explanation needs should precede XAI system design: By studying how 60 crowdworkers naturally explained their damage assessments, the researchers established an evidence-based foundation for XAI design, suggesting that responsible AI development requires first understanding what explanations humans actually need rather than what technologists assume they need.

Context matters for explanation design: The focus on satellite imagery damage assessment highlights that XAI systems should be tailored to specific visual detection contexts, as explanation needs likely vary significantly across different application domains and task types.