TRANSLATING UZBEK REALIA: A COMPARATIVE ANALYSIS OF HUMAN AND AI TRANSLATIONS OF HAMID ISMOILOV'S JINLAR BAZMI

Realia are words for objects, institutions and customs that belong to one culture. They are hard for any translator and a known weak point of machine translation. This article compares how culture-specific items in Hamid Ismoilov's Uzbek novel Jinlar bazmi (2012) are rendered in the published English translation by Donald Rayfield, The Devils' Dance (2018), and by a large language model, Claude Opus 5.5. The novel was the first written in Uzbek to be translated into English and won the 2019 EBRD Literature Prize. Ten passages (2,492 words), one sampled from each chapter by a fixed rule, yielded 122 realia items. Items were classified by thematic category (Vlahov & Florin, 1980). Their renderings were coded with Davies's (2003) and Pedersen's (2011) taxonomies, placed on Venuti's (1995) domestication–foreignization continuum, and checked for errors with an adapted MQM scheme. Religious terms (20.5%) and forms of address (18.0%) were the largest categories. Of the 113 items with a counterpart in both versions, Rayfield domesticated 43.4%, chiefly by substitution, omission and generalisation, while the AI system domesticated 9.7% and retained the Uzbek word in 24.8% (Stuart–Maxwell χ²(2) = 33.04, p < .001). The AI rendering contained four major errors, all failures to recognise a historical or literary referent, and four cultural shifts that imported Christian terms; the human translation contained two minor errors. The findings suggest that, contrary to the common view that machine translation flattens cultural difference, the AI system kept more of the source culture's surface than the human translator, but without the knowledge needed to identify allusions or to explain what it kept.