Japanese haiku features an extremely concise form, profound cultural connotations embedded in kigo (seasonal words), and sophisticated grammatical functions of kireji (cutting words). It has long been a difficult and widely discussed topic in the field of literary translation. This paper analyzes the constraints on haiku translation from linguistic, cultural, and poetic perspectives, and points out that conciseness, seasonal markers, and imagery transitions can hardly be fully preserved simultaneously in cross-linguistic conversion. This paper also discusses the practical performance of neural machine translation and large language models in haiku translation. It is found that artificial intelligence boasts remarkable efficiency in rapidly generating initial drafts and maintaining consistency of terminology, yet it encounters difficulties in handling the intertextual implications of seasonal words and the textual tension created by cutting words. On this basis, the paper proposes a structured human-machine collaboration workflow. It holds that artificial intelligence shall undertake basic tasks requiring speed and consistency, while human translators retain the power of aesthetic judgment and cultural decision-making. The research concludes that although artificial intelligence can assist with a large number of technical translation tasks, human translators are still required to calibrate and confirm the core parts of haiku translation related to poetic perception and cultural memory.