What do we want from Explainable Artificial Intelligence (XAI)? – A stakeholder perspective on XAI and a conceptual model guiding interdisciplinary XAI research

Analysis

Why This Paper Matters

This paper addresses a critical gap in Explainable AI (XAI): the disconnect between technical explainability methods and the actual needs of diverse stakeholders. As AI systems proliferate in high-stakes domains (e.g., healthcare, finance, criminal justice), the demand for explanations grows, yet many XAI approaches are developed without clear understanding of who the explanations are for and what they require. By systematically reviewing stakeholder desiderata from literature across computer science, philosophy, law, and social sciences, the authors provide a much-needed synthesis that can guide both researchers and practitioners.

The paper's significance lies in its explicit recognition that XAI is inherently interdisciplinary. It moves beyond technical metrics (e.g., fidelity, interpretability) to consider human-centered factors such as trust, fairness, and regulatory compliance. This stakeholder perspective is crucial for ensuring that XAI systems are not just technically sound but also practically useful and ethically aligned.

Technical Contributions

• Stakeholder Taxonomy: The paper categorizes stakeholders into groups such as developers, end-users, domain experts, regulators, and affected parties, each with distinct desiderata (e.g., debugging for developers, justification for users, auditability for regulators).
• Conceptual Model: A formal model that links stakeholders → desiderata → explainability approaches → evaluation criteria, providing a structured vocabulary for interdisciplinary discourse.
• Interdisciplinary Framework: The model integrates insights from philosophy (e.g., what constitutes a good explanation), law (e.g., right to explanation), and human-computer interaction (e.g., user mental models).
• Gap Analysis: Identifies under-explored areas, such as the lack of empirical studies on how different stakeholders actually use explanations in practice.

Results

The paper does not present experimental results or quantitative metrics. Its primary output is the conceptual model itself, which is derived from a qualitative synthesis of existing literature. The model's utility is demonstrated through illustrative examples (e.g., how a developer's need for debugging differs from a patient's need for understanding a medical diagnosis). The authors argue that the model can serve as a checklist for evaluating XAI methods and as a blueprint for designing new ones.

Significance

This paper has broad implications for the AI field. It provides a common language for researchers from disparate disciplines to collaborate on XAI, potentially accelerating progress toward more human-centered AI. For practitioners, it offers a framework to ensure that explainability investments are aligned with actual stakeholder requirements, reducing the risk of building explanations that are technically impressive but practically irrelevant. The work also highlights the need for empirical validation of XAI methods in real-world contexts, a direction that could shape future research agendas.