A mental models approach for defining explainable artificial intelligence

Abstract

Abstract Background Wide-ranging concerns exist regarding the use of black-box modelling methods in sensitive contexts such as healthcare. Despite performance gains and hype, uptake of artificial intelligence (AI) is hindered by these concerns. Explainable AI is thought to help alleviate these concerns. However, existing definitions for explainable are not forming a solid foundation for this work. Methods We critique recent reviews on the literature regarding: the agency of an AI within a team; mental models, especially as they apply to healthcare, and the practical aspects of their elicitation; and existing and current definitions of explainability, especially from the perspective of AI researchers. On the basis of this literature, we create a new definition of explainable, and supporting terms, providing definitions that can be objectively evaluated. Finally, we apply the new definition of explainable to three existing models, demonstrating how it can apply to previous research, and providing guidance for future research on the basis of this definition. Results Existing definitions of explanation are premised on global applicability and don’t address the question ‘understandable by whom?’. Eliciting mental models can be likened to creating explainable AI if one considers the AI as a member of a team. On this basis, we define explainability in terms of the context of the model, comprising the purpose, audience, and language of the model and explanation. As examples, this definition is applied to regression models, neural nets, and human mental models in operating-room teams. Conclusions Existing definitions of explanation have limitations for ensuring that the concerns for practical applications are resolved. Defining explainability in terms of the context of their application forces evaluations to be aligned with the practical goals of the model. Further, it will allow researchers to explicitly distinguish between explanations for technical and lay audiences, allowing different evaluations to be applied to each.