Whose Fairness? Value Gaps in AI Ethics and the Institutional Conditions for Embedding Fairness in AI Systems

Description This study investigates how different social groups understand the concept of fairness in the context of generative AI and how these value gaps shape the institutionalization of AI ethics. As AI systems increasingly influence decision-making processes across society, questions of fairness have become central to discussions of AI governance. However, fairness is often treated as a single, universally shared normative concept, despite the likelihood that its meaning varies across social positions, professional backgrounds, and lived experiences. This research challenges that assumption by empirically examining how fairness is perceived by AI practitioners and general users. Drawing on a web-based survey of 600 respondents—including members of the Japanese Society for Artificial Intelligence, members of the Japan Deep Learning Association, and non‑AI professionals—the study analyzes differences in fairness perceptions using a framework derived from Rawls’ theory of justice. Fairness is conceptualized across three layers: (1) fairness of opportunity and procedure, (2) legitimacy in judgment and allocation, and (3) justice as a social consequence. The analysis reveals that AI practitioners tend to emphasize procedural fairness, algorithmic consistency, and transactional legitimacy, reflecting a professional habitus shaped by technical training and rationalist norms. In contrast, general users prioritize outcome equality, protection of vulnerable groups, and corrective justice, expressing expectations that AI should contribute to improving social fairness. The study also identifies gender-based differences, with women across groups more likely to perceive current social conditions as unfair. These findings indicate that fairness perceptions are not merely individual preferences but are structurally shaped by social attributes and professional identities. Such divergences pose risks when AI systems are implemented without acknowledging these differences. If AI design processes rely solely on the fairness assumptions of technical experts, they may unintentionally embed narrow value frameworks into systems that affect diverse populations. The research argues that effective institutionalization of AI ethics requires recognizing the coexistence of multiple fairness conceptions and designing governance mechanisms that incorporate diverse stakeholder perspectives. This includes establishing participatory oversight structures, implementing external audits that can scrutinize the assumptions embedded in AI systems, and ensuring institutional flexibility that allows fairness criteria to be revised in response to social change. Rather than embedding a single definition of fairness into AI systems, the study emphasizes the importance of building governance frameworks that can accommodate value pluralism while maintaining a minimum standard of dignity-preserving fairness. By reframing fairness gaps as structural issues rather than individual differences, this study contributes both theoretically and empirically to ongoing debates on AI governance. It provides a foundation for designing institutional mechanisms that align AI development with societal expectations and support the sustainable and legitimate integration of AI into social systems.