There has been a lot of discussion lately about artificial intelligence (AI) and the law and legal education. My colleague Sarah Gotschall has written extensively about various facets of the intersection of law and AI, asking if AI can make legal practice less stressful; what does AI think a law librarian looks like; and even getting AI to wax poetical about the state of the profession.
Most of the AI discussion I've read has centered on its impact on education and the practice of law. Those areas are significant but outside of the kind of law library work I and my department does. I decided to see how AI would fare at doing traditional Technical Services work. To that end, I asked ChatGPT to do a fairly simple task: catalog a recently purchased book for our collection.
There seemed to be some confusion about the difference between a bibliographic record and a citation. To ChatGPT's credit, it did give me two examples when I only asked for one, albeit two wrong examples. I clarified the question and asked for a bibliographic record in MARC format.
To my surprise, ChatGPT began constructing a line by line MARC record before my eyes. I was incredibly impressed and my mind began reeling with questions about the implications for human catalogers. MARC was developed back in the day to be machine readable so it made sense that something as seemingly sophisticated at ChatGPT would have already assimilated it and become proficient at using it. But then I began to wonder if ChatGPT had access to OCLC and was just copying the OCLC record. Still, pretty impressive since copy cataloging is the vast bulk of what our human cataloger does in this library.
But then I took a closer look at the record. I noticed the date of publication seemed off. Then I realized the publisher was wrong. The record includes ISBN numbers that don't seem to exist. It also includes an OCLC number that does not exist. To ChatGPT's credit, it did warn me that the record might need some editing. On the plus side, some of the subject headings are accurate and the call number is in the ballpark.
My library does require correct publication information. Doing a little research on this book, it seems to have grown out the author's previous article on the topic. Maybe ChatGPT was confused by the publication history but I couldn't find any connection between this work and New York University Press. Upon further questioning, ChatGPT said it pulled the information about this book from the LC Catalog. However, LC has the correct publication data. So while at first look, this looks like a good record for this title, it seems to be a fabricated mix of true and untrue elements, with AI filling in gaps. Imagine if human catalogers did the same.
Speaking of AI making something out of nothing, I also experimented with
DallE. My colleague had previously asked it to generate images of law librarians. I got more specific and asked for an image of "Technical Services librarians celebrating a successful reorganization of their department."
This AI seems to have a very specific idea of what a Technical Services librarian looks like. Biases in AI generated art have been well-documented and I think we can see a little of that in these examples. After these two experiments, I reached the same conclusion as my colleague: AI is still very much a work in progress. I'm not quite ready to worry about AI taking Technical Services librarian jobs.
