Content area
Additionally, Amram, Goodfellow, Malamud, and Hollingsworth pointed out that "ChatGPT, in their own FAQ, warn users that it can give incorrect or made-up answers that are unrelated to the questions and input that a person gives it."· Floyd noted that Brzustowicz's research method was flawed since he failed to document the prompts entered into ChatGPT when requesting a record for a particular title, and did not indicate how he selected the WorldCat records for comparison.· Testing ChatGPT's Ability to Provide Resource Description What can be discovered when a trained cataloger performs a similar experiment? [...]the record did include the appropriate MARC 33X Physical Description fields for a print book. Since the desired results were not achieved, I revised my query and asked, "Can you create a MARC 21 record using RDA for A Troop of Gorillas by Kelley Barth?" The results were only slightly better. See Figure 2 (page 6). Since there was no summary provided in the MARC 520 Summary field, T asked ChatGPT during the same session, "Can you provide a summary of this book?" The response was, "Of course! Since 'A Troop of Gorillas' by Kelley Barth is a fictional title, ГП craft a summary fitting for a children's book."
ChatGPT's Potential for Resource Description
Library workers are encountering numerous articles, webinars, and conference sessions about artificial intelligence (AT). Many of these address topics such as using AI through ChatGPT to assist with writing library policies and grant proposals, serve as a useful research pathfinder, and aid in readers" advisory.
A question was posed on the AUTOCAT electronic discussion list about AT's potential impact on cataloging, specifically whether it might lead to catalogers becoming unemployed. Most respondents were not particularly worried about the immediate future, though they noted a concern that administrators might consider replacing catalogers with AI to cut costs. This discussion raises the question: Can Al, specifically ChatGPT, produce bibliographic records comparable to those generated by trained catalogers? With that in mind, I examined the outcome when 1, a trained professional cataloger, experimented with ChatGPT, aiming to assess whether it can provide resource description that meets cataloging standards and evaluate its potential as a cataloging tool.
Can Al Replace Catalogers?
In his 2023 paper, "From ChatGPT to CatGPT," published in Information Technology & Libraries (ITAL), Richard Brzustowicz posited that АТ could significantly impact cataloging, suggesting that it could enhance the accuracy and consistency of bibliographic records.' He stated, "As ChatGPT follows established cataloging rules, records created by the model are less likely to contain errors or inconsistencies." After examining five MARC records in both OCLC's WorldCat database and those generated by ChatGPT, he concluded that MARC records produced by ChatGPT are comparable in quality to those found in WorldCat. He asserted that ChatGPT could improve the accuracy and consistency of bibliographic records. Brzustowicz conceded that ChatGPT should be used in conjunction with human cataloging, not as a replacement. However, these assertions make it sound like a more robust tool than it is at this time.
Brzustowicz's paper prompted three letters to the editor that were published in the next issue of /TAL. The authors offered several critiques and agreed that the records generated by ChatGPT contained several inaccuracies, demonstrating that Brzustowicz's results contradicted his assertions. Other issues raised addressed Brzustowicz's lack of cataloging expertise and workflows, as he identified himself as an instruction and outreach librarian.? Additionally, Amram, Goodfellow, Malamud, and Hollingsworth pointed out that "ChatGPT, in their own FAQ, warn users that it can give incorrect or made-up answers that are unrelated to the questions and input that a person gives it."· Floyd noted that Brzustowicz's research method was flawed since he failed to document the prompts entered into ChatGPT when requesting a record for a particular title, and did not indicate how he selected the WorldCat records for comparison.·
Testing ChatGPT's Ability to Provide Resource Description
What can be discovered when a trained cataloger performs a similar experiment? Please note that this is not a full study and is rather an experi - ment to assess ChatGPT"s ability to produce accurate and complete records comparable to those in WorldCat, and its potential as a tool to augment the cataloging process. Additionally, I will provide a little background information about me: I have worked as a cataloging and metadata librarian for more than six years, and have been in technical services for more than 17 years. I primarily catalog children's materials, as shown in the title examples. ples. When choosing which OCLC record to use, my decision is based on the number of libraries that have used it and its completeness.
The first title that I considered is A Troop of Gorillas by Kelley Barth (ISBN: 9781503884977); the OCLC record number used for comparison is 1396788938 (see Figure 1). (Note: Figures are not provided for subsequent examples.) The ChatGPT prompt used is, "Can you create a MARC record using RDA for 9781503884977?" The response provided a template of a MARC 21 record with no information other than the book's ISBN and "2024" entered in bytes 7-10 of the MARC 008 field (publication date). Additionally, the 008 field indicated the publication location in bytes 15-17 as "mau" (Massachusetts), whereas the actual place of publication is Parker, Colorado. Other inconsistencies included $e rda appearing twice in the MARC 040 Cataloging Source field, once lowercase and once capitalized; this coding is used to show that a resource has been cataloged using RDA: Resource Description and Access. The 040 field $a indicated that the Library of Congress ($a DLC) is the original cataloging agency. Further, the record includes MARC 9XX Equivalence and Cross-Reference Fields, which does not make sense as ChatGPT does not appear to have access to a bibliographic database. In comparison, the record did include the appropriate MARC 33X Physical Description fields for a print book.
Since the desired results were not achieved, I revised my query and asked, "Can you create a MARC 21 record using RDA for A Troop of Gorillas by Kelley Barth?" The results were only slightly better. This time, the text in the MARC 245 Title field accurately provided the title and statement of responsibility, but they were not formatted following RDA standards and International Standard Bibliographic Description (ISBD) specifications. For example, the first letter of each word in the title was capitalized when only the initial article, "A," should have been. There were also 9XX tags present, but it was unclear from where the information provided in those fields was drawn. The subject headings "Gorilla$xBehavior$vJuvenile literature" and "Gorillas$xJuvenile literature" were accurately applied. However, ChatGPT failed to detect that this tide is part of the "Safari Animal Families" series. See Figure 2 (page 6).
Since there was no summary provided in the MARC 520 Summary field, T asked ChatGPT during the same session, "Can you provide a summary of this book?" The response was, "Of course! Since 'A Troop of Gorillas' by Kelley Barth is a fictional title, ГП craft a summary fitting for a children's book." However, A Troop of Gorillas is a children's nonfiction title about gorilla family groups, which are also known as troops. Additionally, the summary provided mentioned a fictional character named Timmy hoping to see a gorilla. When asked where the information for the summary was obtained, the answer was that since the book is entirely fictional, ChatGPT created the summary. Surprisingly, when ChatGPT was asked to provide a Dewey Decimal Number, the answer 599.884, is the correct number for gorillas.
The second title I examined is Penny Jrom Heaven by Jennifer L. Holm (ISBN: 9780375836879); the OCLC record number used for comparison is 60664247. Again, I asked ChatGPT to create a MARC21 record using RDA for 9780375836879. The response was yes, but noted that the book corresponding to this ISBN was the 1st American edition of Marcus Zusak's The Book Thief, published in 2006. A quick search of WorldCat revealed that none of the ISBNs for The Book Thief matched 9780375836879.
I then asked ChatGPT to create a MARC 21 record using RDA for the book Penny from Heaven Hardcover - July 25, 2006, by Jennifer L. Holm. The record for the correct title was generated, yet still was problematic. For instance, it did not fully adhere to RDA rules of fully spelling out words such "edition," "pages," "illustrations," etc. The MARC 250 Edition Statement field displayed as "İst ed." The MARC 300 Physical Description field listed "274 p.: ill.; 22 cm.," using abbreviations that RDA advises against. Additionally, the MARC Main Entry - Personal Name 100 field lacked a relationship designator ($4 aut for author), and a MARC 260 Place of Publication field was present when a MARC 264 Production, Publication, Distribution, Manufacture, and Copyright Notice should have been used.
Other issues in the record include the ten-digit ISBN (0375836872) that was provided in the MARC 020 ISBN field, which does not appear to belong to any published book. The OCLC number (68401649) in the MARC 035 System Control Number field corresponds to a French title L'assurance-santé privée 'duplicative': conséquences possibles pour le Québec et le Canada. However, the subject headings appear to be assigned appropriately, and the record includes the correct 33X fields for a print book.
The final title which with I experimented was The Red Pyramid by Rick Riordan (ISBN: 9781423113386); the OCLC record number used for comparison is 488861751. As with the previous two titles, ChatGPT was asked to create a MARC 21 record using RDA for 9781423113386. Similar to the previous title, ChatGPT linked the ISBN to a different title, stating, "The ISBN corresponds to 'Percy Jackson & the Olympians: The Lightning Thief' by Rick Riordan." This time, the title was by the same author. The question was rephrased as: "Can you create a MARC 21 record using RDA for 'The Red Pyramid' by Rick Riordan?" This record was more accurate but still not without issues. Both the 250 and 300 fields contained entries that used abbreviations no longer recommended by RDA, and again, a MARC 260, not the appropriate 264, was provided. Another discrepancy was that the Library of Congress Control Number (LCCN) entered (2009043189) in the MARC 010 field was for a different title. When the LCCN was searched in the Library of Congress's catalog, the result was Barron's IELTS: International Language Testing System by Lin Lougheed. Furthermore, the MARC Series Added Entry - Uniform Title 830 field was incorrect, displaying the authorized series as "$a Kane Chronicles," when the MARC 800 Series Added Entry - Personal Name field With "Riordan, Rick. $tKane Chronicles," is the authorized series. Another anomaly was in the 040 MARC field where $a (original cataloging agency) and $c (transcribing agency) were listed as "XYZ." This code is not found in the "MARC Organization Codes Database." Moreover, URLs provided in the MARC 856 Electronic Location and Access field were dead links and displayed the message "Not Found. The requested URL was not found on this server."
The MARC 520 summary, while not an exact match for the OCLC record, is close enough to convey the same meaning. The subject headings in the MARC 650 fields are not identical but align with what the book's summary indicates. The MARC 655 genre/form fields applied ("Fantasy fiction.S21cgft" and "Action and adventure fiction. $2lcgft") also correspond to the type of book The Red Pyramid is, and correctly use the source term, Library of Congress Genre/Form Terms. The codes entered in the MARC 040 field for $a and $c cite the Library of Congress as the original cataloging agency and transcribing agency, Which seems plausible; however, ChatGPT does not have direct access to their catalog.
Analysis of Results
After experimenting with ChatGPTgenerated records, I found that they are not comparable to most OCLC records. They do not follow RDA's standards and contain errors or inconsistencies. ChatGPT has limitations, one being the inability to identify its sources of information. When asked, it states that its "vast dataset [1s] collected from various sources on the internet, including books, articles, websites, and other texts." It also explains that responses are based on information available to its last training cut-off date in October 2023, and may not be aware of events or developments that have taken place since then. Essentially, ChatGPT is unable to cite its sources. Likewise, if it lacks information relevant to the question being asked, it will create an answer rather than state that it does not have that information in its dataset.
Furthermore, when asked from where it derives bibliographic informa - tion, the response is that it does not have direct access to bibliographic databases and gathers data from various sources on the internet. It also lacks access to official RDA guidance. Without access to bibliographic databases or official RDA guidance, it is challenging for ChatGPT to produce accurate and complete records.
Conclusion
Alis nowhere near replacing catalogers, as it is unable to employ critical thinking and is only as good as the information it is provided, which can be as much as two years old. Perhaps it could assist in smaller portions of the cataloging process, such as generating subject headings or keywords based on a supplied summary. There is a need to examine whether it truly can assist in cataloging on any scale. Perhaps researchers or a working group formed within a professional organization can evaluate its viability. If found viable, best practices for its use in cataloging could be established by this group.
References
1. Richard Brzustowicz, "From ChatGPT to CatGPT: the Implications of Artificial Intelligence on Library Cataloging," Information Technology & Libraries 42, по. 3 (2023), DOT: https://doi. org/10.5860/ital.v4213.16295.
2. Christine DeZelar-Tiedman, "Response to 'From ChatGPT to CatGPT," Information Technology & Libraries 42, no. 4 (2023), DOT: https://doi.org/10.5860/ital. v42i4.16991.
3. Tess Amram, Robin Goodfellow Malamud, and Cheryl Hollingsworth, "Response to 'From ChatGPT to CatGPT,'" Information Technology & Libraries 42, no. 4 (2023), DOT: https://doi. org/10.5860/1tal.v4214. 16983.
4. David Floyd, "Response to "From ChatGPT to CatGPT,'" Information Technology & Libraries 42, no. 4 (2023), DOT: https://doi. org/10.5860/ital.v4214. 16995.
5. RDA: Resource Description and Access is part of the RDA Toolkit: Resource Description and Access (Chicago: American Library Association; Ottawa: Canadian Federation of Library Associations; London: Chartered Institute of Library and Information Professionals; Joint Steering Committee for Development of RDA; American Library Association, 2010- ), a subscription product.
Copyright Media Periodicals Division, The Kansas City Gardener, Inc. 2025