What if we took "Use of Practical AI in Digital Libraries" seriously?

This presentation was delivered at the DNB conference “Rethinking AI in Libraries – Data Quality, Infrastructure and Applications for Small and Large Language Models” (Frankfurt, 30 Jan 2026). It asks what it would mean to take practical AI in digital libraries seriously, using automated subject indexing as the core use case. Focusing on the LLMs4Subjects shared tasks (SemEval 2025 and GermEval 2025), the talk introduces the task setting (top-k recommendation of GND subjects for library records), the dataset construction from TIBKAT open data and the GND subject vocabulary, and key real-world properties such as long-tail sparsity, distribution shift across splits, and signals of ambiguity in subject terminology. It then summarizes community participation and highlights representative system strategies—from retrieval and prompting to translation, synthetic data generation, and classical extreme multi-label methods—drawing implications for quality, robustness, and sustainable deployment in library workflows.