What Do We Want From Explainable Artificial Intelligence (XAI)? -- A Stakeholder Perspective on XAI and a Conceptual Model Guiding Interdisciplinary XAI Research

Different stakeholders need different explanations: The paper identifies multiple classes of stakeholders (such as users, regulators, developers, and affected individuals) who have distinct needs and expectations from AI explanations. A single explanation approach cannot satisfy everyone, meaning organizations must consider who specifically needs to understand an AI system and tailor explanations accordingly.

XAI research lacks coordination across disciplines: The paper notes that XAI literature is "vast" and "spreads out across multiple largely disconnected disciplines," making it unclear how different explainability approaches actually help meet stakeholders' needs. This fragmentation suggests the field needs better communication between computer scientists, ethicists, social scientists, and domain experts to develop more effective solutions.

Current XAI approaches often miss the mark: The paper emphasizes that it "often remains unclear how explainability approaches are supposed to achieve the goal of satisfying stakeholders' desiderata," indicating that many technical XAI solutions may not actually address what stakeholders really want or need from AI systems.

A common framework is essential for progress: The authors propose a conceptual model that explicitly defines key concepts and relationships for evaluating explainability approaches. This shared framework could help researchers across different fields work together more effectively and ensure that XAI development is guided by actual stakeholder needs rather than technical capabilities alone.

Interdisciplinary collaboration is critical: The paper highlights significant "interdisciplinary potential" in developing and evaluating explainability approaches, suggesting that responsible AI requires input from multiple fields beyond computer science to truly serve stakeholders' interests.