Invited Talk: AI-Generated Rich Metadata

The talk will report on the lessons learned from our two experiments. The first experiment involves using large language models (LLMs) to generate cataloging records for highly visual library materials, such as posters. Traditional cataloging for these collections is time-intensive, and conventional records often fail to represent key visual elements. To address this challenge, we developed a custom AI-driven application that extracts text using OCR and generates structured visual descriptions—covering features such as color, layout, and artistic style. The system then encodes these descriptions into MARC 21 fields, including the 520 field. The second experiment evaluates whether enriched, AI-generated metadata can improve resource discovery beyond basic keyword searches. In experimental testing, when AI agents were provided with richer visual context rather than text alone, they were able to respond to more complex, thematic research questions and synthesize relevant visual evidence for users. Finally, we will discuss a human-in-the-loop workflow that redefines the role of the cataloger. In this model, catalogers move away from routine data entry toward higher-level responsibilities, including prompt design, validation, and domain and ethical review.