AI 공정성과 형사사법의 규범적 한계 : 형사판례 비교 분석

This article critically examines the limitations of approaches that reduce AI fairness to performance metrics or statistical parity, reconstructing it instead as a multidimensional legal concept. It draws on four key judicial decisions: State v. Loomis (United States) on algorithmic risk assessment in sentencing; R (Bridges) v. Chief Constable of South Wales Police (United Kingdom) on live facial recognition; decisions of the German Federal Constitutional Court on predictive policing; and D.H. and Others v. Czech Republic (European Court of Human Rights), which established indirect discrimination doctrine based on statistical evidence. Drawing on these cases, the article proposes a reconceptualization of AI fairness along five dimensions: distributive equality, relational equality, freedom as non-domination, procedural capability, and structural justice. It argues that algorithmic fairness cannot be assessed by accuracy or error rates alone, but must also address how automated systems classify individuals, exercise opaque power, and entrench structural disadvantage. The article concludes by outlining normative criteria for the lawful and legitimate use of AI in criminal justice, with particular attention to implications for jurisdictions such as South Korea.