Identifying Cultural and Semantic Translation Errors in Pashto–English Proverbs Translation: A Comparative Study of ChatGPT, Gemini, and Google Translations

Abstract

Machine Translation (MT) has advanced rapidly with the emergence of neural and AI- powered systems, yet translating culturally embedded figurative language particularly proverbs continue to pose significant challenges, especially in low-resource languages such as Pashto. This study examines how accurately these three major AI translation tools such as Google Translate, ChatGPT, and Gemini interpret and translate 20 culturally rich Pashto proverbs into English. Therefore, by applying a qualitative research design supported by basic quantitative analysis, the current study evaluates the semantic, figurative, and cultural accuracy of the AI-generated translations. Moreover, the study uses Vinay and Darbelnet’s (1958) Model of Translation Strategies as the analytical framework for identifying the strategies employed and the types of errors made by each tool. The analysis of the study reveals substantial variation in performance by highlighting that ChatGPT achieved the highest accuracy of 75%, effectively applying equivalence and modulation strategies to preserve figurative meaning, Gemini performed moderately with the frequency of 65% with occasional semantic shifts whereas, Google Translate showed the lowest accuracy of 15%, frequently producing literal, fragmented, or culturally inappropriate translations. The findings of the study highlighted the limitations of AI systems in handling proverbs from low- resource languages, demonstrating that figurative and culturally bound expressions require oblique translation strategies that current MT tools often fail to apply. The present study contributes to MT research by providing empirical evidence from a low resource language such as Pashto and underscores the need for culturally informed, context-sensitive AI translation models.