Translation Strategies of AI Models in Rendering Jordanian Proverbs into English: A Comparative Analysis of DeepSeek, Claude, and Grok
TL;DR
It is demonstrated that preserving source imagery does not necessarily result in idiomatic equivalence and that AI models differ considerably in their strategic approaches to culturally embedded proverbial language, highlighting the need for AI translation models that can better balance the preservation of cultural meaning and rhetorical imagery with natural and idiomatic expression in the target language.
Abstract
This study investigates the translation strategies used by three AI models, namely DeepSeek, Claude, and Grok, in rendering Jordanian proverbs into English. The study examines 190 Jordanian proverbs, producing a corpus of 570 AI-generated translations. It employs Baker’s (1992) classification of idiom translation strategies as the main theoretical framework, supplemented by Literal Idiomatic Transfer (LIT) as an additional analytical category. A quantitative and qualitative approach was used to identify the frequency of the translation strategies and examine representative examples of how the models handled the linguistic, metaphorical, and cultural features of Jordanian proverbs. The findings showed that LIT was the dominant strategy, accounting for 76.14% of all translations, followed by translation by paraphrase at 17.54%. Idiomatic equivalence was relatively uncommon, with Strategy A accounting for 3.33% and Strategy B for 2.63%, while Strategies E and F each occurred only once (0.18%), and Strategy C was not observed. Clear differences also emerged among the models. Grok and DeepSeek relied heavily on LIT, at 88.95% and 85.26%, respectively, whereas Claude used LIT less frequently (54.21%) and showed a considerably stronger preference for paraphrase (38.42%). The qualitative findings further revealed that the models frequently preserved metaphorical and culturally embedded images without producing familiar English idioms. Paraphrase generally increased clarity but sometimes reduced the rhetorical imagery and proverbial form of the source expression. The findings demonstrate that preserving source imagery does not necessarily result in idiomatic equivalence and that AI models differ considerably in their strategic approaches to culturally embedded proverbial language. The study highlights the need for AI translation models that can better balance the preservation of cultural meaning and rhetorical imagery with natural and idiomatic expression in the target language.