Who's afraid of AI-powered translation?

Abstract This study compares normalisation effect in Chinese–to–English translations produced by DeepL, ChatGPT and human translators, and examines its implications for the (in)visibility of human translators in human–machine collaboration. Moving beyond a monolingual comparable corpus design, the study integrates source-language features and applies principal component analysis, random forest modelling and t-SNE visualisation to better capture the impact of translation on linguistic outcomes. The results showed that DeepL outputs exhibited the strongest normalisation tendencies, while ChatGPT–generated translations demonstrated greater likelihood of deviating from target-language conventions. DeepL translation displayed a higher degree of mechanicalness with less autonomy in handling the typological differences between Chinese and English. While ChatGPT translation demonstrated an advantage in cohesion, its tendency to overuse certain linguistic features might create unnaturally elevated tone. These findings highlight the double-edged nature of AI-driven translation technologies: while their consistency and adaptability may support post-editing workflows, their potential influence on how users perceive translation and the role of translators warrants critical attention. We argue that ethical frameworks should not only promote fair recognition of human input but also convey the limitations of translation technology to the end-users, ensuring human intelligence remains central in an increasingly automated translation landscape.