Abstract:
Despite its importance, the concept of algorithmic transparency has yet to be fully explicated. By asking what is transparent or opaque, we propose a comprehensive framework dividing transparency into four forms: use transparency, which discloses algorithm goals and uses; data transparency, which informs sources, processing, and data quality; model transparency, which explains how the model functions; and decision transparency, which clarifies the rationale behind decisions. We demonstrate that various explainable AI methods aim to enhance model or decision transparency within this framework.
