Cataloguing, Metadata, and Generative AI. Early Experiences and Future Perspectives

Abstract

The article deals with the intersection of generative artificial intelligence (AI) and bibliographic/metadata practices, assessing how large language models (LLMs) can support cataloguing and metadata creation while navigating the constraints of formal knowledge architectures. In the first section, the article discusses the evolution of cataloguing paradigms from MARC to Linked Open Data (LOD), emphasizing the shift from rigid records to semantic, entity-based models like FRBR, RDA, and BIBFRAME. The second section deals with the epistemological clash between deterministic, rule-based metadata standards (the "architect") and probabilistic, generative AI systems (the "oracle").
Three strategies are proposed for integrating AI into bibliographic workflows:
1) Specialized AI systems trained exclusively on controlled, high-quality datasets.
2) Retrieval-Augmented Generation (RAG), blending LLMs with authoritative knowledge bases.
3) Next-generation LLMs enhanced via reasoning models, multimodal inputs, expanded context windows, and small/medium-scale local models to align generative outputs with metadata standards.
Key challenges include hallucinations, data sparsity in bibliographic corpora, and the obsolescence of MARC-centric experiments. The article argues for caution against retrofitting AI onto outdated data models, urging alignment with LOD and IFLA’s Library Reference Model (LRM). Ethical considerations (bias, transparency, AI literacy) and the potential of local SLMs/MSLMs for privacy-sensitive applications are highlighted.