Toward a Hermeneutics of Explainability: Unraveling the Inner Workings of AI Systems

Abstract

For all the enthusiasm and sheer volume of research in explainability recently, there is curiously little consideration of the interpretive theory that forms the backdrop of many of the proposed methods. Virtually all accounts of explainability presume that there are right interpretations into which researchers should guide a particular audience when asked to provide a clear explanation of an outcome. In this paper, we argue that it is time for technical researchers to look to the humanities and social sciences traditions surrounding interpretation (with roots in the work of Gadamer in the 1960s and the hermeneutic circle) in order to ground our explainability efforts in a more informed, critical, and self-reflexive context. Indeed, in doing so, we will shed a more critical view of what is likely a commonplace task for human researchers that should not be taken lightly even when machine-based support strategies are deployed. Our take is that the core ideas of hermeneutics provide a template for understanding the relationality of interpretive acts. With these ideas in hand, AI researchers should be able to reason more coherently, and with greater humility and sensitivity, about what interpretative acts mean, and about how we might design systems and support strategies that help to realize specific ends in interpretive situations.