Why do explanations fail? A typology and discussion on failures in XAI

Explanation failures are complex and interconnected: The paper argues that XAI (explainable AI) problems don't stem from single issues but rather from overlapping failures across multiple dimensions. This means that simply fixing one technical problem with an explanation method may not solve the underlying issues users experience.

Both the system and the user contribute to explanation failures: The research distinguishes between "system-specific failures" (technical limitations in how explanations are generated) and "user-specific failures" (how people misinterpret explanations). Responsible AI requires attention to both sides—not just better algorithms, but also better user training and interface design.

Current ad-hoc approaches are insufficient: The paper critiques existing studies that investigate XAI problems in isolation, suggesting this fragmented approach misses the full picture of why explanations fail in practice. A more systematic, holistic framework is needed to understand and prevent real-world harms from poor explanations.

A typological framework can guide improvement: By categorizing different types of explanation failures, the paper provides practitioners with a structured way to diagnose problems in their XAI systems and identify which failures are most critical to address in their specific contexts.

Better understanding of XAI limitations is essential for responsible deployment: The research emphasizes that organizations deploying AI systems need to move beyond assuming explanations work as intended and instead actively understand where and how their explanation methods might fail users or stakeholders.