In the fall I wrote about how AI can help, rather than replace, libraries and their collections. Let me summarize these positive attributes before I begin to address a new, but related, topic.

• Unlike most library systems, LLMs are based on vectors rather than indexes, making them comfortable with:

  • language as it is commonly spoken

  • uncontrolled vocabularies and ambiguity

  • keyword cousins, not just keyword twins

• Transformers allow LLMs to excel at parallelism, meaning they can:

  • move between forms fluidly, including text, images, and other media

  • translate well between languages

• Chat versions of LLMs are not just conversational, but iterative and longitudinal, so they can:

  • ask for clarifications or refinement

  • digest some initial documents to further refine or sequence a query

  • synthesize information from multiple sources

  • remember context over time (if you let them)

• Multimodal LLMs and computer vision are swiftly getting better at transcribing:

  • unprocessed and handwritten documents

  • unstructured data

So far, so good, at least from the perspective of the librarian trying to provide better access to a library’s collections. (For more on the above, including a Q&A with other librarians, watch the CNI panel I was on in December, “Discovery and Use Reimagined: Connecting Scholarly Collections and Artificial Intelligence Workflows.”)

I will also note here briefly — and I’ll return to this subject in a future piece in greater depth — that most of the advantageous uses of AI I have listed hold true even if we use more minimal or open-weight LLMs rather than frontier models like Claude, ChatGPT, or Gemini, which means that for those concerned about the environmental impact of AI, there very well might be a path to using just enough AI to improve research while not using so much it has measurable detrimental effects.

* * *

Let me now take off my dean of the library hat and put on my professor of history hat: From a researcher’s point of view, if AI-assisted interfaces are fully implemented, have the librarians gone far enough in using this new technology to improve research substantially? Or, conversely, have they already gone too far, thus impairing the work of dedicated researchers?

Anyone who has spent considerable time in academia, journalism, the literary and creative arts, and libraries, archives, and museums — I’m looking at you, the fine audience for this newsletter — knows of colleagues who believe deeply in the sanctity of the arduous, comprehensive research process — and maybe you do too. Intellectual sweat precedes lucid sense-making; serendipity befriends those who are thorough. (Bless us, Robert Caro.) From this viewpoint, any mediation of the search and discovery process by AI hijacks human agency, elides details, and perhaps even removes opportune luck from the scholarly process.

Having spent considerable time myself turning pages in the stacks and archives, I get the criticism. But it is worth noting that search and discovery within large-scale collections have always been mediated, whether that mediation occurred through physical means, like a card catalog, or digital means, like today’s library website. Choices have been made about metadata, descriptions, and the various information architectures of sorting and surfacing. AI merely presents the opportunity for a new ordering based on different principles. In early testing, I have come to believe that for most students and researchers AI can, if implemented thoughtfully and within certain parameters, provide a better search and discovery experience. And just as important, if these interfaces are done well there should be nothing preventing a scholar from proceeding the old way by sequentially scanning a digital collection or shelf, or doing a translation or data extraction themselves.

* * *

Modesty has not been an attribute of AI, however. At the opposite end of the spectrum from the traditionalists, and to their great horror, the invasion of LLM-assisted scholarship threatens to erode not only the Puritan ethic of serious research, but human understanding itself, just as it has upended methods of teaching and learning. In fields related to AI, such as computer science, journals and conferences are already flooded with chatbot-assisted articles and presentations. The practice is, inevitably, spreading to other fields. The extreme end point of this AI trend could be called “one-shot scholarship,” in which a researcher uses a single prompt not only for search and discovery of sources, but also to have the AI read all of those sources, analyze the data, and — heck, why not, we’ve come this far — compose the entire academic paper, from abstract to bibliography.

There are, of course, many latitudes between the two poles of Zero AI and AI For Everything. But in the hype and anxiety of these early AI days, we are having trouble talking about these middle climes, focusing too much on the icy extremes. Finding the sure ground between automation and human agency is not simple, but charting some basic principles can help us determine whether specific uses of AI cross a line. This semester, I’d like to begin that mapping. Looking carefully at what happens at each stage of what librarians call the “research lifecycle,” from initial ideas, through source and data acquisition, to analysis and publication, seems like a good place to start.