Issues in XAI #5: Understanding Black Boxes: Interdisciplinary Perspectives 

TU Dortmund, 5 – 7 September 2022 (in-person event)

Artificial intelligent (AI) systems are used to great success in many contexts, e.g. for medical diagnosis or autonomous driving. They often depend on extremely complex algorithms designed to detect patterns in large amounts of data, which makes it difficult to discern their inner workings. However, especially in high-stakes situations, it is important to be able to evaluate desirable system properties, e.g. their safety, fairness or trustworthiness, and so to understand how they work. To address this concern, the field of explainable artificial intelligence (XAI) develops methods to make opaque AI systems understandable by means of explanations.

But what does it mean to understand an (opaque) system or its outputs? What is the link between explanations and understanding? How do different contexts affect the explainability and understandability of AI systems? How important is understanding really and how does this depend on the specific context of use, e.g. on legal vs. medical contexts? How can we make complex systems understandable at all and how do different explainability methods fare in this regard (e.g., surrogate models or perturbation-based methods)? What can theories of understanding and explanation from psychology and philosophy contribute to XAI, and how do these insights mesh with specific explainability approaches from computer science?

Clarifying relevant concepts such as understanding and explanation is an important fundamental step towards developing human-centered XAI methods; vice versa, researchers from philosophy and psychology can gain a deeper understanding of the concepts by reflecting on their application in XAI. In this light, our workshop will bring together researchers from philosophy, psychology, computer science, and law to push forward research on understanding and XAI.

Funded by the