Faithful or evasive? An empirical study on translation norm preferences of Chinese and American LLMs in Chinese official political and policy discourse

Abstract

Introduction Large language models now handle a share of cross-lingual political translation, but no study has directly measured whether they follow stable normative preferences when doing so. Methods We target Chinese-to-English translation of Chinese political and policy terms embedded in authentic official discourse. Mapping 52 publications across translation theory, LLM empirics, and AI alignment yields a five-dimensional Translation Norm Orientations (TNO) framework: Faithfulness (FN), Fluency (FLN), Cultural Adaptation (CAN), Ethical Hedging (EHN), and Foreignization Retention (FRN). Eight mainstream Chinese and American LLMs complete a Forced-Choice Translation Ranking Task (FCTRT) on 20 ideologically sensitive political terms, ranked across three repeated trials. Kendall’s W assesses ranking stability, Mann–Whitney U tests with BH-FDR correction test between-group differences, and hierarchical clustering identifies latent normative prototypes. Results All eight models rank FN first; CAN, EHN, and FRN settle into a shared low-priority band beneath it. None of the five dimensions shows a statistically significant CN–US split. FN alone carries a medium effect size (Cliff’s δ = 0.50), which points to a difference in how strongly, not which way, the two groups lean. Cluster membership cuts across national lines rather than following it: four transnational profiles emerge, separated by foreignization tolerance rather than developer geography. Discussion These results tie translation norm theory to AI evaluation and supply a benchmark other researchers can reuse to examine LLM-mediated political communication.