A Multidisciplinary Survey and Framework for Design and Evaluation of Explainable AI Systems

Different disciplines approach explainability differently: The paper finds that machine learning, visualization, and human-computer interaction researchers focus on fairly independent topics when designing explainable AI systems, which can create fragmentation in how we think about making AI interpretable and accountable.

There's no one-size-fits-all explanation: The framework identifies that different user groups (such as developers, domain experts, and end users) have different needs for understanding AI decisions, requiring tailored design goals and evaluation methods rather than universal approaches to explainability.

Evaluation methods must match design goals: The paper emphasizes that how we test whether an AI system is truly explainable should directly connect to what we're trying to achieve—meaning organizations need to align their evaluation strategies with their specific objectives for transparency and accountability.

Multidisciplinary collaboration requires shared language: By mapping design goals to evaluation methods across disciplines, the framework provides a common vocabulary and structured approach that helps teams from different backgrounds (engineers, designers, social scientists) work together more effectively on XAI projects.

Iterative design and testing is essential: The paper advocates for cyclical processes where XAI systems are continuously designed, evaluated, and refined based on feedback, rather than treating explainability as a one-time feature added to an AI system.