Generative artificial intelligence (GAI) is reshaping the fundamental paradigm of translation. Its gains in efficiency starkly contrast with unsettled ethical practices. Drawing on Chesterman’s ethics of commitment, communication, and service, this paper identifies three persistent paradoxes in the GAI-mediated translation arena: suspended responsibility attribution, dissipation or distortion of cultural information, and the erosion of translators’ professional agency. Specifically, blurred duties give rise to dual disputes over error liability and intellectual-property attribution; mechanized communication increases the risks of contextual collapse and cultural misreading; and service instrumentalization obscures translators’ value while weakening user transparency and informed consent. To restore balance, the study proposes a governance agenda comprising: (a) robust mechanisms for delineating accountability across developers, platforms, translators, and clients; (b) cultivation of AI-enabled cultural translation competence to mitigate bias and preserve diversity; and (c) translator-centered human-AI collaboration that embeds expert oversight across the workflow. The paper argues that the sustainable evolution of GAI in translation hinges on re-balancing instrumental rationality with value rationality, thereby establishing a trustworthy ethical framework that safeguards quality, fairness, and public confidence.