생성형 AI의 번역문 구분 가능성에 대한 연구

Can Generative AIs Distinguish Translated from Non-Translated Texts?

This study explores the potential of Generative AI (GAI) models, including GPT-4, GPT-4o, and Gemini, as tools for distinguishing translated texts from non-translated texts. The analysis was conducted in two stages. First, the GAI models were tasked with describing Translation Universals (TUs) and then used to detect translated texts—initially without, and subsequently with, the TU characteristics of the texts included in the prompt. In the second stage, prompts were designed to score the extent of TUs in translated and non-translated texts, first without the TU characteristics and then with them included in the prompt. The results indicate that prompts requesting theoretical background enhance translation detection. When TU characteristics were explicitly included in the prompts, all models showed improved accuracy in distinguishing translated texts. This research contributes to expanding the application scope of GAI in translation studies, from generating translations to evaluating texts, and enhances understanding of the operational mechanisms of GAI.