The following was submitted by AV Cataloging and Metadata Intern, Jenn Leishman.
Applying artificial intelligence to archives is a terrifying concept for many information professionals. After all, the reason I love archival work is how human of a task it is to pay attention to stories from the past that otherwise would be overlooked. Spending time as a GBH AV Cataloging and Metadata Intern, working under the guidance of Metadata Operations Manager Owen King, I was reminded that the human-to-human connections present in AV archival work are not lost with machine-assisted cataloging. Instead, we are able to catalog more efficiently and work towards higher patron discoverability through the American Archive of Public Broadcasting (AAPB). A vast majority of items in the AAPB lack structured contributor records (structured lists of the people who appeared in or helped produce a program), and this can begin to be amended through work towards advanced cataloging workflows.
My co-intern, Avery Schanbacher, and I, were first introduced to the world of AV cataloging through work on the New Jersey Nightly News (NJN) collection. Previous GBH Archives intern Kaycee Conover describes this experience perfectly: our time “… is spent with the people in the programs.” Through our interaction with NJN, while working on increasing the efficiency and accuracy of the cataloging workflow, we still found ourselves going back to watch segments with interesting chyrons, getting attached to recurring anchors, and following daily life in New Jersey. Because my introduction to AV archives was through cataloging this charismatic media, of course, it was not hard for me as an archives student to understand just why we want to prioritize the discoverability of local names and stories.
This is an example of a row from a cataloging aid, or a cataid for short. These are integrated with various open-source AI applications created via a partnership with Brandeis University’s CLAMS Project.2 The first box contains a frame from the program, identified via a scenes-with-text (SWT) detection app as a chyron and person. A “chyron” refers to the lower-third identifying text present in the broadcast. The second box contains text extracted via a vision-language model (VLM). For the beginning of the summer, we were using SmolVLM2, also packaged as a CLAMS app by the Brandeis team. In the third box, the extracted text is formatted into catalog data by prompting a GPT4.1-mini model. It is frames like this one above, frames of local community members, workers, activists, organizers, and family, that make the ability to search the AAPB so important. It’s the local figures that may not have any information publicly available online that I love cataloging the most, and that tend to appear most commonly in programs like NJN. With these catalog records, family members, friends, and anyone who knows of them now will have an online avenue to find them and listen to them. I think that’s beautiful.
The work I completed in this internship worked towards a solution to improving metadata across as many AAPB catalog records as possible. When we have an increasingly accurate AI workflow, there can be pressure to fully automate the process. However, this idea is hard to reconcile as an archivist, as interacting with programs, checking name spellings, and checking for sensitive cases is what feels like the heart of the work. Contextual integrity played a large role in conversations around chyron attributes this summer, and the consideration of these cases is just one area where human interaction with collections is incredibly important. Mistakes in contributor records can lead to a lack of trust, which is not something to take lightly.
That being said, I am very hopeful about the increase in speed of metadata creation. It’s important to remember that the purpose of this is not just for us, the archivists, but for the researchers, friends and family members of people, or even individuals themselves. Once these records are published on the AAPB, somebody may be able to find a new story about a parent or a funny memory of a friend. Some of this content is fifty years old, some of it is five. The metadata we are creating and cataloging participates in the protection of these memories, pulling them out from thousands and thousands of hours of online content into tangible collections of personhood.
Even just in my time at GBH, I saw the increase in speed of the workflow. We were constantly adjusting our approaches, and in July, the Brandeis team released the new VLM app, based on Qwen3.5. This approach allowed us to improve from our previous workflow that used the SmolVLM2 model. As you can see below, Qwen3.5 has less difficulty picking up accurate text, especially on top of patterned backgrounds.
With significantly more accuracy in extracted text using Qwen3.5, I was able to complete cataids at a much quicker rate, spending less time correcting errors in name spellings and more time focusing on scene selection and contributor records for appearances without chyrons. This increased speed has fantastic implications leading towards increased metadata access, and I loved getting to be the first to use new tools and see the differences in our workflows. With Qwen3.5, in mid-July, I performed a test on one of the collections I was cataloging, Charlie Rose, to see how many episodes could be cataloged in a 90-minute time frame. This led to a total of 26 episodes cataloged, averaging 3.4 minutes per episode. This is the fastest cataloging that has been recorded to date. With the help of improved technology, we are able to increase the speed of AI-assisted human cataloging, and still maintain human correction and care in the process. This provides hope that as the technology advances, we can adjust our workflows for smoother processes to hone in on human-necessary intervention.
The tradeoff between archivist interaction with material versus broader discoverability of material is something I have not quite figured out yet, and I can only assume shapes the fraught opinions of many archivists working with AI technologies. As discussed with Avery and Owen, there is a responsibility for the metadata we create to be accurate and reflective of the faces it is attached to.
Having graduated with my Bachelors in Computer Science not long before I started this internship, I never knew where exactly my technological understanding for software and machine learning would play a role in my professional library and archive work. However, it is at the heart of the work being done here at the GBH archives. How can we prompt machines to use text while maintaining accuracy, discoverability, and eliminating harmful circumstances? How can this output be stored and transformed to be ready for insertion into metadata? What kinds of tags can we create in the catalog records to best serve various organizations that this tool may be useful for? These are all questions related to applications of machine learning tools. I am excited to take these applications into future roles I hold, but also remember that as archival and cataloging work advances with AI usage, we have to keep asking ourselves questions. What do we owe the people appearing in the AAPB, or other cultural heritage institutions? I think we owe them our human care and caution at the very least.
References:
- “New Jersey Nightly News; New Jersey Nightly News Episode from 11/24/1981,” 1981-11-24, New Jersey Network, American Archive of Public Broadcasting (GBH and the Library of Congress), Boston, MA and Washington, DC, accessed August 7, 2026, http://americanarchive.org/catalog/cpb-aacip-259-ks6j3v1x.
↩︎ - CLAMS, “CLAMS Project,” Brandeis LLC, https://clams.ai/home/. ↩︎
- “Charlie Rose; #1016; 1991-10-22,” 1991-10-22, Thirteen WNET, American Archive of Public Broadcasting (GBH and the Library of Congress), Boston, MA and Washington, DC, accessed August 7, 2026, http://americanarchive.org/catalog/cpb-aacip-1ff50892968. ↩︎
