Explainability and AI Governance

Abstract

To counter the opacity of algorithmic and artificial intelligence (AI) decision-making, a burgeoning discourse around the aim of ‘explainable AI’ has emerged. Most efforts directed at this aim treat AI opacity primarily as an epistemic challenge, which may have ethical ramifications. For instance, if an AI-driven decision concerning a loan application is biased, explaining this decision may prevent discriminatory outcomes. As such, explainability primarily fulfils an instrumental purpose to remediate ethical concerns for individual users. However, recent work has hinted toward three changes in perspective. First, from considering explainability as being instrumental to considering it as having constitutive value in a discursive context, captured by notions like deliberative agency. Second, from considering the effect of the technology (AI) on the individual to considering the larger decision-making ecosystem. Third, accordingly, from considering explainability as a matter of (individual) ethics to a matter of (political) governance. This chapter explores this tripartite direction by explicating the outlines of the governance of explainable AI. Drawing from the work of Paul Ricoeur, it argues that the primary end of the explainability criterium is to bridge the ethics of conviction and the ethics of argumentation in a context in which humans and AI systems discursively relate. This chapter will develop its argument in three sections. In the first section, this chapter will develop a critique of the current discourse of explainable AI. In the second section, this chapter will present a new, governance perspective, drawing from Ricoeur’s dynamic between the ethics of conviction and argumentation. In the third section, this chapter will apply this view to the organization context of fraud detection, demonstrating how explainability is a function of the governance in which this activity unfolds and showing how the framework developed helps shed light on this dynamic.

Ähnliche Arbeiten