Abstract

Recent research on large language model translation has largely focused on product-oriented comparisons between human and machine output, in which human-AI collaboration is typically narrowed to post-editing. This article argues that an approach overlooks how translation quality is shaped through decision-making in the translation process. Drawing on process-oriented perspectives in translation studies (Toury, 2012), it reconceptualizes human-AI collaboration as a redistribution of decision-making between human translators and large language models. Using the UN Women Strategic Plan (2022–2025), translated bidirectionally across English, French, and Chinese as its corpus, the study examines this collaboration through a controlled API-based method. Human translators design prompts, set parameters, and structure workflows, thereby constraining and guiding model execution, while LLMs perform large-scale text generation within these conditions. Evaluation using COMET and BERTScore shows consistently mid-to-high semantic quality across language directions. From a process-oriented perspective, the study shows how translation quality in multilingual institutional contexts is shaped by human-defined constraints on model execution.

This work is licensed under a Creative Commons Attribution 4.0 License.