Practical Dilemmas, Ethical Risks and Human-Machine Collaboration Pathways for AI-Empowered Legal Translation
Abstract
The rapid iteration of generative artificial intelligence (AI) and neural machine translation has brought efficiency innovation to legal translation practice, while posing new challenges to the conversion of legal English terminology, transmission of concepts across legal system and translation quality control. Different from general text translation, legal translation features dual attributes of linguistic transformation and transplantation of legal cultural connotations. AI translation has inherent limitations in the accurate matching of legal terms, disambiguation of legal senses of polysemous words, reproduction of legal stylistic features, and reconstruction of legal logic. It also gives rise to multiple ethical risks including translation distortion, data bias, algorithmic black-box effects, technological alienation and blurred accountability. Against the backdrop of global expansion of cross-border commercial transactions and international judicial cooperation, indigenous Chinese rule-of-law discourse requires reliable translation output to eliminate cross-jurisdictional misunderstandings. Nevertheless, off-the-shelf large language model outputs frequently produce superficially fluent yet legally inaccurate renderings, a phenomenon known as model hallucination, which aggravates risks in high‑stakes legal scenarios. Drawing on empirical cases of legal terminology translation, this paper figures out the strengths and practical bottlenecks of artificial intelligence applied to legal translation, analyzes the root causes of translation dilemmas, and constructs a human-machine collaborative workflow adapted to legal translation scenarios. Optimization suggestions are put forward from the dimensions of talent cultivation, industrial norms, technical supervision and terminology database construction, so as to facilitate the accurate cross‑cultural communication of indigenous Chinese rule-of-law concepts.