WHERE HUMAN AND MACHINE TRANSLATIONS DIVERGE: LOCATING THE PROBLEM OF ENGLISH–UZBEK TRANSLATION

English–Uzbek translation now happens mostly through two very different channels at once: professional or student human translators, and freely available machine translation (MT) and large language model (LLM) tools such as Google Translate and ChatGPT. The second channel has grown far faster than research has been able to follow it. Uzbek is an agglutinative, morphologically rich, low-resource language in which grammatical relations — case, number, possession, tense — are encoded through long, ordered suffix chains rather than free-standing words or prepositions, as English mostly does [1]. This typological distance from English is already known to cause systematic difficulty: a recent large-scale error analysis of Uzbek text generated by GPT-5, Gemini 2.0 Pro, and LLaMA 4 found that morphological errors — faulty suffix ordering, broken vowel harmony, agreement failures, and confusion of formality register — account for more than a third of all non-target-like forms produced by current models [2]. If state-of-the-art systems still fail this often when simply generating Uzbek text, it is reasonable to expect comparable or greater difficulty when translating into Uzbek from a structurally distant source language such as English, where a translator or system must additionally preserve the meaning, register, and cultural content of the original. Whether human translators face the same difficulties, different ones, or avoid them through strategies unavailable to a machine, has not been systematically asked. The relevance of the present study lies in this open and increasingly consequential question: as MT and LLM tools become the default first option for English–Uzbek translation in education, publishing, and everyday communication, it is necessary to establish, on principled linguistic grounds, where their output is likely to diverge from human translation and why.