The Politics of AI Ethics: How Normative Principles Translate into Law and Policy in South Korea

Abstract

A growing global consensus around AI ethics principles has not produced equivalent convergence in law and policy. The process through which normative principles are translated into governance, however, remains understudied: we know comparatively little about why some principles survive, change, or disappear as they move into policy, how different institutional pathways shape that process, or how translation itself should be measured. We develop a framework for studying ethics-to-governance translation and examine several possible sources of variation, including principles’ demands, institutional fit and state capacity, policy instrument ‘pathways,’ and political administration. We examine South Korea, a high-capacity developmental state and early mover in comprehensive AI regulation, analyzing 15 ethical principles across 69 government documents issued between 2017 and 2025: 23 ethics guidelines, 27 non-binding executive policy documents, and 19 enacted statutes and enforcement decrees. Our analysis combines principle-level keyword prevalence, keyword-in-context and embedding-based semantic validation, a translation-gap measure comparing ethics guidelines with binding and non-binding governance pathways, and article-level coding of whether statutory provisions specify obligations, duty-bearers, procedures, penalties, and remedies. We find that internationally recognized and Korean-salient principles begin at similar levels of prominence in ethics guidelines but diverge substantially as they move into governance, especially binding law. The pattern is not simply one of international versus domestic values: principles such as governance and auditing, which map readily onto established functions of administrative coordination, monitoring, and verification, survive more consistently than principles requiring less-developed rights-enforcement or other institutional machinery. Binding law also filters normative content more strongly than non-binding policy. Yet when rights-oriented principles such as justice/fairness and transparency reach statutory text, they are comparably operative; the main difference is legal coverage, not drafting strength. These findings are consistent with institutional fit as one explanation for selective translation while leaving substantial room for political, cultural, strategic, and other institutional explanations. More broadly, we conceptualize ethics-to-governance translation as a multidimensional process requiring explanation in its own right and offer a replicable framework for studying which ethical commitments enter governance, through which pathways, in what form, and under what conditions.