Copley Connects Spring/Summer 2025

AI Comes for Digital USD By Amanda Makula , Digital Initiatives Librarian Generative artificial intelligence (Gen AI) is finding its way into many aspects of higher education and academic library work, and at Copley Library, the Archives, Special Collections, and Digital Initiatives department is no exception. In early 2025, we had the opportunity to participate in the beta program for Odyssey (now renamed JSTOR Seeklight ), a software platform made by the nonprofit organization ITHAKA, designed to employ Gen AI to speed the creation and processing of descriptive metadata for digital collections. For example, when it first premiered, ITHAKA teased that a task taking 250 hours of manual human work could be accomplished in under three hours with JSTOR Seeklight 1 . Hearing that, we were eager to put it to the test! The University of San Diego was one of eight colleges/ universities that participated in the beta program, alongside participants from places such as the University of Chicago, St. Lawrence University, and the University of Wisconsin-Milwaukee. The first stage of the program was the Qualitative Review. We provided 100 items – a mix of files from the USD Magazine collection: (https://digital.sandiego.edu/usdmagazine/) and the Southern Cross newspaper (https://digital. sandiego.edu/southern-cross/) — to feed through the platform so that the AI could generate metadata for

We compared the metadata assigned to the Southern Cross newspaper by JSTOR Seeklight with our own manual descriptions.

(“inches”). Some title records would include an issue number, and others would not. Moreover, some important fields such as Copyright Statement, Language, File Size, and File Type were missing altogether. Perhaps most concerning of all, we discovered some errors that were strikingly inaccurate, and rather baffling. For example, the Southern Cross , the official newspaper of the Diocese of San Diego, was at times tagged by JSTOR Seeklight as a “Latter Day Saint press,” a “Southern Africa newspaper,” a “Gay press publication” and “Women’s periodicals, Mexican.” Based on these results, we quickly learned that we could not rely on the tool without investing time and attention to closely scrutinize its outputs. Despite these shortcomings, at a closing round table discussion, participants noted strengths of the platform such as generating general descriptions of handwritten documents, creating base-level metadata that can be edited rather than manually starting from scratch, and helping to process large volumes of documents more efficiently. Overall, we felt that the tool was most useful when utilized for descriptive information — telling us what is included — rather than to evaluate, or make decisions or judgments about the content. If JSTOR Seeklight is any indication, AI-generated metadata platforms have a long way to go, but we look forward to following the developments and, hopefully, one day utilizing them in future workflows.

each one. Then we analyzed the results, noting where the tool performed well — and where it did not. One of the biggest issues that we noticed was a lack of consistency in outputs and formatting. Within a single field, such as the Format field, sometimes the tool would specify one type of measurement

(“centimeters”) and other times, something else

The Fall 2015 issue of USD Magazine was one of 100 items fed into JSTOR Seeklight for AI generated metadata.

1 https://about.jstor.org/blog/embarking-on-project-odyssey-a-journey-to-transform-digital-collection-stewardship/

UNIVERSITY OF SAN DIEGO | 7