An Overview of Multidisciplinary Research on Explainability: Concepts, Challenges, and Design Considerations

Intelligent systems, including those powered by Artificial Intelligence (AI), are being increasingly used in our everyday life and there is a growing demand for making them transparent, understandable, predictable, and ultimately, acceptable. This can be partially achieved through the integration of explanation capabilities. However, practically speaking, the implementation of useful explanation capabilities for end-users has always been a difficult task, and the difficulty grows as intelligent systems use more complex reasoning and learning procedures. This paper provides a brief overview of studies on explanations and describes important explanation concepts across different disciplines and through the history of their integration in knowledge-based and later Machine Leaning (ML)-based systems. It discusses the general challenges of explanation design, and those that are unique to dynamic and/or distributed environments. Finally, it argues for a human-centered perspective for explanations where characteristics of good explanations and design considerations are discussed.