AI translation of Gaza war news coverage between Arabic and English: biased or unbiased?

Abstract

This paper investigates whether three major AI translation engines, Gemini, ChatGPT, and DeepSeek, render politically sensitive Gaza-war terminology from Arabic into English in a biased or neutral manner. Using random thirty-five certified benchmark terms, the paper conducts a mixed-methods analysis combining quantitative error with qualitative examination of framing, context, and ideological shifts. Results reveal that although all three AI systems demonstrate high lexical accuracy in general terms, their performance diverges significantly on politically charged expressions. Gemini tends to replace Arab-centric terminology with Western media frames, while ChatGPT frequently produces softened or legally neutral alternatives. DeepSeek, by contrast, shows a pattern of literal translation that overlooks culturally and historically fixed proper nouns. Quantitatively, Gemini and ChatGPT each committed (3) deviations (8.58%), while DeepSeek committed (6) deviations (17.16%). Qualitatively, however, Gemini’s errors carried the strongest ideological weight, particularly in terms such as "suicide bombings" and "Separation Wall." The study concludes that all three engines exhibit varying forms of bias, ideological, euphemistic, or literal indicating that current AI translation systems cannot yet ensure fully neutral representation in conflict-based linguistic contexts.