인공지능 공정성의 이론과 실제

The emergence of ChatGPT has led to a significant resurgence of social interest in artificial intelligence (AI). While many scholars express concerns about the potential risks associated with ChatGPT, there seems to be a decreased focus on fairness issues compared to the past. However, some perceive this as a positive development, given the increasing quantity and quality of research on AI fairness and the corresponding technical tools. Ultimately, a comprehensive understanding of the field of AI fairness is essential to provide an accurate normative evaluation of this situation. With this concern in mind, this paper aims to explore the discourse on AI fairness from both theoretical and practical perspectives. First, from a theoretical standpoint, we examine the role of AI fairness within the broader discussions on trustworthy AI governance. Additionally, we investigate why fairness is favored over competing concepts like equality and justice. Subsequently, we delve into the subcategories of AI fairness, namely distributive fairness and representational fairness, and thoroughly explore the relationship between bias and unfairness, which share conceptual similarities within the AI context. Next, this paper delves into the technical methodologies for the practical implementation of AI fairness theory. These methodologies can generally be classified into input-centered and output-centered approaches. We analyze the current emphasis on the latter over the former. Furthermore, while examining the specific details of both approaches, we focus on gaining a deeper understanding of the AI fairness metrics, which serve as an engineering implementation of the output-centered approach. It is worth noting that the existence of various fairness metrics for AI and the potential conflicts between them highlight the fact that technical regulation is not omnipotent. Furthermore, this parallel can be seen as reminiscent of ongoing discussions on fairness within human society.