Abstract

This study aims to examine the feasibility of using domestic sovereign AI models and global large language models (LLMs) for automated creation of library metadata by comparing their performance in MARC record generation. To this end, six generative AI models (GPT, Gemini, Grok, HyperCLOVA, EXAONE, and A.X) were used to generate MARC records for 40 domestic and foreign monographs, and their field-level performance was evaluated using three criteria: completeness, correctness, and rule compliance. The analysis showed, first, that the three global LLMs (GPT, Gemini, Grok) generally outperformed domestic sovereign AI models, with fewer missing fields and more stable handling of formal elements such as indicators and codes. However, their performance tended to decline when the cataloguing target shifted from English-language to Korean books, as errors increased in field configuration and statement of responsibility. Second, the domestic sovereign AI models (HyperCLOVA, EXAONE, A.X) exhibited relatively low overall performance in both MARC21 and KORMARC, and did not show clear performance gains even for Korean books. Third, at the field level, most models generated relatively stable results for title and statement of responsibility (245), whereas rule-dependent fields such as series statements (490/830) and the choice of main entry showed large performance gaps between models and revealed structural misunderstandings of cataloguing rules for example, mechanically transferring MARC21 practices for series treatment to KORMARC. These findings suggest that, at present, generative AI should be introduced into library metadata workflows primarily as an assistive tool for generating draft records and supporting error detection and correction, rather than as a fully automated cataloguing system. The results also indicate that, in order to ensure stable performance of domestic sovereign AI models, systematic training on Korean bibliographic data, including KORMARC records, is required. Furthermore, the careful selection and curation of training data emerges as a key task in building sovereign AI systems for library applications.