REMAX: Rigorous Evaluation Methods for AI Explainability

Explaining the decisions of complex AI models, like deep neural networks, is very hard. This is a key challenge for ethically responsible AI. Computer science develops technical solutions to this challenge within the field of Explainable AI. However, rigorous methods for evaluating the explainability of AI are currently lacking. Computer scientists use data-driven metrics. While they provide some insights for developers, it is unclear whether such metrics track ethically relevant properties. Conversely, AI ethics uses normative criteria. While based on principled ethical analyses of explainability, they are too abstract and provide little practical guidance. REMAX will bridge this gap by establishing a close collaboration between computer science and AI ethics. We will study how principled normative criteria can be translated into novel metrics of AI explainability, and how closer alignment with AI development practices can make these normative criteria more practically usable.