Hiroki Shima (Columbia U) has posted “Ethical AI in Multilingual Institutions: Semantic Drift in Fairness Principles and the Hybrid Justice Model” on SSRN. Here is the abstract:
Multilingual public institutions such as the United Nations, the European Union, and ASEAN increasingly deploy AI systems for recruitment, service delivery, and regulatory tasks under ethics frameworks that emphasize fairness, transparency, and non-discrimination. These principles, however, are usually drafted in English and then translated into multiple legal and working languages. Research in legal and institutional translation shows that normative terms such as “fair,” “equity,” and “accountability” rarely map cleanly across languages and legal cultures, creating systematic “semantic drift” between source texts and target versions. At the same time, global AI governance instruments—from UN-system resolutions to regional guides such as the ASEAN Guide on AI Governance and Ethics and the OECD AI Principles—are increasingly used as reference points for national and organizational policy. This paper conceptualizes “semantic drift” in translated AI ethics and fairness policies as a distinct layer of risk for organizational justice and trust in AI-supported processes. Building on the Hybrid Justice Model (HJM), it treats semantic drift as a mechanism that can widen the Justice Gap Index (JGI) and generate “semantic trust drift” in the Perception Shift Index (PSI) across language communities, even when underlying models satisfy standard statistical fairness constraints. Drawing on insights from legal translation, institutional translation, and cross-cultural value research (e.g., GLOBE, Schwartz), the paper proposes a Multilingual Fairness Drift Framework that links three elements: (1) textual drift in fairness‐related concepts across languages, (2) implementation drift as policies are operationalized in local procedures and explanations, and (3) perception drift as candidates and citizens experience AI-mediated decisions. Using UN, EU, and ASEAN examples as comparative reference points, the paper illustrates how misaligned translations of fairness principles can produce divergent expectations and justice perceptions across language groups, even within a single institution. It then outlines a research agenda that uses HJM indicators—particularly μ (cultural and linguistic alignment) and PSI—to measure and govern semantic drift over time. The contribution is twofold: it reframes multilingual translation not as a technical afterthought but as a core dimension of ethical AI governance, and it extends organizational justice theory to cover the textual and semantic infrastructures through which AI ethics principles are communicated and implemented in multilingual institutions.