Abstract:
Artificial intelligence (AI) has transformed professional translation, yet maritime terminology remains difficult to translate because of its complexity and standardisation. This paper examines how four machine translation (MT) systems, ChatGPT, Gemini, DeepL and Google Translate, render English maritime terminology into Romanian. The corpus comprises excerpts from COLREG and the IMO Standard Marine Communication Phrases, evaluated against their official Romanian translations. In the texts examined, large language model (LLM)-based systems reproduced codified terminology more reliably than dedicated MT engines, and their Rule 18 output improved between two rounds. All four systems, however, failed to translate fixed collocations and non-compositional expressions consistently, at times reversing their meaning. Where MT output follows the English wording more literally than the official translation, it misses the conventional Romanian equivalents. Since accuracy in this domain is a matter of operational safety, the paper calls for specialised terminology databases, improved contextual algorithms and structured user feedback.