Contrasting Chinese-English translation strategies by ChatGPT and DeepSeek in subtitling magic culture-specific items: a case study of Nezha 1 and Nezha 2

Artificial intelligence is increasingly used in audiovisual translation, yet limited research has examined how large language models translate culture-specific items (CSIs) in fantasy films. This study investigates the Chinese-English translation strategies employed by ChatGPT-5 and DeepSeek-V3.1 for magic CSIs and the influence of prompt design on strategy selection. Drawing on Franco Aixelá’s taxonomy, a mixed-methods analysis was conducted on 90 magic CSIs from Nezha 1 and Nezha 2. Translations generated under basic, context-enhanced, and role-specific prompts were compared with official human subtitles. Results show that both AI systems predominantly employ source-oriented conservative strategies, with linguistic translation most frequent. DeepSeek-V3.1 exhibits stronger conservative tendencies and greater reliance on single-strategy rendering, whereas ChatGPT-5 uses more mixed and substitutive strategies. Human subtitles show a more target-oriented profile, particularly through absolute universalization, autonomous creation, and deletion, and have the lowest mean character count among the English conditions. Prompt variation has limited influence on strategy patterns, although the context-enhanced prompt slightly increases target-oriented renderings. These findings demonstrate the potential of large language models for subtitling while highlighting the continuing importance of human post-editing for cultural adaptation and context-sensitive translation.