“Not just language conversion but cultural transmission”: comparing human-first and AI-first workflows for translation quality, efficiency, and speed in the Anthropocene

Abstract

Artificial intelligence (AI) has reshaped translation workflows, moving from manual, sentence-level crafting to integrated pipelines using translation memories, neural machine translation, and large language models, yet education struggles to prepare translators for these shifts, risking deskilling and ethical challenges. This study aimed to optimize training by comparing two AI-assisted workflows in a food-related menu translation task, evaluating quality, time, edits, and trainee perceptions to enhance efficiency, accuracy, and self-efficacy. Class 1 ( n  = 29) used an AI-first workflow, where AI generated initial drafts followed by human post-editing to correct errors and refine outputs. Class 2 ( n  = 27) employed a human-first workflow, with translators drafting texts and AI polishing to enhance style and coherence. Quality (fluency, adequacy), time, edit behaviors (total, necessary, harmful), and perceptions were assessed via quantitative metrics and qualitative feedback. Class 2 outperformed Class 1 in adequacy ( M  = 43.02 vs. 39.22, p  = 0.002, d  = −0.85) and total score ( M  = 89.93 vs. 86.67, p  = 0.012, d  = −0.68), with no fluency difference. Class 1 was faster ( M  = 2763.69 s vs. 3494.11 s, p  < 0.001, d  = −2.40) but made more harmful edits ( d  = 1.07). Class 1’s time-quality correlation ( r  = 0.905, p  < 0.001) showed effort-intensive post-editing, unlike Class 2’s efficiency. Human-first workflows improve quality and confidence, while AI-first prioritize speed. Training should emphasize human-first workflows, prompt design, and selective post-editing to align with workplace curation roles, countering deskilling.