Contrastive explanations—a contribution to the framework for social design of explainable AI

Please find the abstract and an introductory paper linked below. During the meeting Prof. Rohlfing will present her research in the context of the explainability in AI problem and we will have a chance to discuss both this topic and the ideas on how to make our research connected to societal issues, while keeping them quite basic in nature.

Technological advancements in machine learning affecting humans’ lives on the one hand and also regulatory initiatives fostering transparency in algorithmic decision making on the other hand drive a recent surge of interest in explainable AI (XAI). Explainability is discussed as a solution to sociotechnical challenges such as intelligent software providing incomprehensible decisions or big data enabling fast learning but becoming too complex to fully comprehend and judge its achievements. With explainable AI, more insights into the functions, decisions, and usefulness of algorithms are expected. If an explanation is successful, it results in an understanding. Current XAI research is centering around one-way interaction from which solutions to achieve understanding are derived. In the presentation, I will point to an important resource for achieving understanding that has been overlooked so far: the interaction with the addressee. The A05 project of the TRR 318 gives insights into how the cognitive processes should be considered in the design of interaction with the addressee.