School libraries and generative AI: a match made in hell

I’ve seen a lot of chatter lately about how generative AI is going to replace everyone’s jobs, including librarians. That generative AI can easily do the job of a reference librarian because all we do is look up a patron’s question and then give them the answer, right?

Wrong.

Let me walk you through all of the things I do in my job. For the record, I’m a Library Director. I’ve worked in libraries and archives since 2009, and had my MLIS since 2011. I’ve worked in public libraries (both in youth services and adult), special collections, academic libraries, and archives (ran an archives for nearly 4 years). I transitioned to independent school libraries in 2016 and am now in my sixth year. 

My current job is to manage the library and the archives, and the archives includes both institutional archival material (like yearbooks, photos, correspondence, ephemera, objects, etc) as well as historical student records. I’m going to set aside detailing all the work I do in the archives, as that’s a whole conversation in and of itself. Suffice it to say, managing an archives is a full-time, highly specialized field that comes with a variety of unique complications with regard to AI.

So, focusing on my work in the school library, here’s everything I do:

  • Manage the physical space – including maintenance, cleaning, furniture upkeep and arrangement, IT support for teachers using the space, monitoring student behavior and volume, managing reservations on library classroom and meeting space, organizing and managing supplies of arts and crafts room, etc.
  • Circulation – assisting patrons with check-out or self-check, processing holds, paging, etc.
  • Collection development – selection, acquisition, and removal of material including books (of all age ranges and genres), DVDs, ebooks, eaudiobooks, textbooks, games, special collections, databases, and maps; weeding collection, including pulling off the shelves, deleting from the LMS, marking physical material as discards, and disposal (either through donation or recycling); manage gifts/donations, etc.
  • Tech services – including cataloguing, labeling, and covering books, fixing old MARC records, book maintenance and repair, etc.
  • Teach – research skills, edtech, critical analysis, evaluation, note-taking, teaching parent classes, etc.
  • Director level – manage the budget, hiring/firing/supervision, develop/implement policies and procedures, liaise with administration and board, strategic planning, library onboarding of new faculty and staff, etc.
  • Wider school – participation in committees, department meetings, collaborating with administration, attending and hosting school events, running and creating content for the library’s social media account, attending and teaching professional development, etc.

If we were fully staffed, our library would have a full-time director, full-time instructional librarian, and part-time archivist. We have had other staff in the past, however, currently I do everything. Literally everything. No volunteers or student workers. I’m by myself. 

In theory, adding more AI, or specifically generative AI, into the library would actually make my job easier. I’d love to be able to offload digitization and metadata creation for archival material or copy-cataloguing. But given where generative AI is now and where I think it’s likely to go, that isn’t in the cards. For the key reason that artificial intelligence is exactly that: artificial. It’s not creating or learning but mimicking thought. It’s not doing research but generating a sentence based on words that are often associated together. It doesn’t do any of the things I actually need it to do.

Let’s take cataloguing. Generative AI models cannot create functional MARC records. Even if more data is added, it still won’t be able to catalogue anything. Because cataloguing is way more than a MARC record. Determining what information goes into that MARC record is what’s important. 

There is no one way to catalogue. Sure, most libraries in the US use DDC (Dewey Decimal System) or LCC (Library of Congress Classification). But there’s been a recent push in school libraries to de-dewey or genrefy nonfiction collections. And even just sticking with DDC, we don’t even have formal agreement on that. DDC is extremely problematic. Its foundation is racist, sexist, xenophobic, and white supremacist. For example, the KKK and the Black Panthers are sometimes catalogued together under 322.42; I put the Black Panthers under 973.0496, African American history. Christianity takes up 200-280, with 290 being “Other religions”. Don’t even get me started on the chaos that is 398 and J398. When I add new material to our LMS, I usually copy-catalogue using Z39.50. The LMS compiles a record based on the libraries I have added into my system, but no two MARC records are the same. I still have to edit what’s been created. For each copy-catalogued record I need to alter or remove LCSH, DDC, and book description, and add in tags.

There are official LCSH (Library of Congress Subject Headings) that go in a typical MARC record, but many of us have begun using our own versions of LCSH because the Library of Congress is too slow in updating their subject headings. The debate over “Illegal aliens” is a great example. Library staff had been clamoring for years to replace that subject heading. Several years ago, I got tired of waiting and created my own subject heading: “Undocumented people”. In 2021, Library of Congress formally announced replacements: “Noncitizen” and “Illegal immigration”, which, like, come on. Those are barely better than “Illegal aliens” and still aren’t reflective of what that community actually calls themselves. So I’m sticking with “Undocumented people”. I’ve replaced “Indians of North America” with “Indigenous people of North America”, and countless others. My new LMS, Access-It, offers tags, and ultimately I’d like to get rid of subject headings altogether and just use tags. 

How will generative AI manage all that? Whose cataloguing data will be added to the language model? How will generative AI determine what is “good” cataloguing or “bad” cataloguing? What about when changes are made to cataloguing best practices? Will generative AI rely on the Library of Congress as its authority? Or ALA? What about when there is grassroots change on best practices, how will that be incorporated? Again, no one catalogues the same way. Pick any nonfiction book and look it up in three different library systems and there will likely be several differences, up to and including different call numbers. Cataloguing is not a computation with one set answer. All those differences between systems are determined by a cataloguer using their knowledge of best practices, internal library procedures, patron usage trends, recommendations from the publisher and/or retailer, and cultural responsiveness. Generative AI is not built to handle that kind deliberation and evaluation. 

Or we can look at reference work. Some people think doing reference is a matter of a patron asking a question and a librarian typing that into a system and out pops an answer. But most of reference work is actually trying to figure out what the patron really wants. A patron asks for a mystery novel, but what they really want is an eaudiobook of a cozy mystery with a lesbian protagonist. A patron asks if we have a book on the Philippines but what they really want are primary sources on queer people in the Philippines before Spanish colonization. A patron asks for an article on gun control but what they really want are a variety of verified, scholarly sources discussing the debate so they can develop their thesis statement into a full-fledged essay. Google can’t get them those answers because the patron often doesn’t even know what question they really want to ask. My job is to first tease out the heart of their query and second teach them how to find that information. When I work through a reference query, I explain my process and try to let the student do it for themselves. I tell them why I’ve typed in this term or searched this site. It’s a learning opportunity as much as it is problem-solving. Generative AI can’t do that. Not even Google’s AI can do that. 

Generative AI, or AI as a whole, cannot replace library staff. To assume that it can is to wholly misunderstand what it is library workers do and what services libraries provide. Could generative AI be useful in libraries in the future? Sure, if we can address the myriad levels of unethical practices built into it and if we can find ways to tweak it so it’s useful even if not terribly profitable. But I can’t see that happening any time soon. Do you know why so many school libraries use Follett Destiny even though most of us despise it? Because there are only a handful of school LMSes out there and most of them are extremely outdated, difficult to use, and lacking a wide variety of important functions. Why are there so few products? Because school LMSes aren’t very profitable, so there’s no incentive to improve. You cannot convince me that some techbro looking at his stock prices is going to care about providing efficient, useful, ethically sourced LLMs for a school library. AI should not replace a librarian, but if we’re not careful, that’s exactly what’s going to happen. And you’re not going to like what comes after.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s