Abstract:
Large language models (LLMs) are now widely used to translate Bangla into English, and their output usually reads
smoothly. This study asks whether that fluency preserves pragmatic and cultural meaning. It compares the translations of five LLM
platforms (Gemini, ChatGPT, DeepSeek, Grok and Claude) with published human translations of 38 Bangla segments. Most segments
come from three canonical texts: Jibanananda Das's "Banalata Sen", the opening chapter of Bibhutibhushan Bandyopadhyay's Pather
Panchali and Sheikh Mujibur Rahman's 7 March speech. The remaining segments come from Shazed ul Hoq Khan Abir's short story
"Mofiz-er Relation-e Spark Nai" (2026), a recent work that the models are unlikely to have encountered during training.". Drawing on
implicature, speech act theory, House's model of register and Venuti's notion of domestication, the study reads each version closely
against its source. The LLMs proved most reliable with names and famous texts, but least reliable with idioms, polysemous words and
traditional counting units. In the speech, four of the five systems lost or reversed the force of a key threat. Changed numbers and
substituted everyday objects were concentrated in the recent story, which suggests that strong results on canonical texts may partly
reflect familiarity. The human translations were more idiomatic, yet they also domesticated the texts through added explanation and
converted cultural units. The study therefore argues that human and machine translators both transform Bangla texts, though in
different directions, and that fluent English is no guarantee that meaning has survived.