da-kuk/E+/Getty Images
This summer, 404 Media followed a tracking device hidden inside a shipment of rare books to an Amazon facility in Las Vegas. Workers there said the bindings were being cut off so the pages could be scanned for AI training. The story attracted attention because the books were being cut apart, but that was not the part that interested me. Libraries get rid of books all the time. We call it weeding. Sometimes a book is damaged, outdated or simply no longer useful to the collection. Nor is the legality of what Amazon did my concern here.
What stayed with me was what happened after the book became data. Once scanned, its words can travel much farther than the physical object ever could. They can be searched, analyzed and used by AI systems that answer questions for millions of people. There is enormous potential in that. Material that once took days or weeks to find might become discoverable in seconds. But an AI system might also give you the right fact from a book without telling you which book it came from. It might “remember” the argument without telling you which edition it read, or give you the conclusion without showing you the disagreement that surrounded it.
The answer can be right while the path back to the evidence is gone.
Librarians and archivists have a word for that path: provenance. It tells us where something came from and what happened to it along the way. That connection is essential to scholarship. Research libraries hold much more than published books. We hold letters, photographs, oral histories, manuscript drafts and records that exist nowhere else. Often the words are only part of what matters. A crossed-out sentence matters. The next letter in the folder matters. Who wrote something matters. So do the promises made when a family or community entrusted those materials to a university.
In my work, questions about AI and these collections rarely arrive as grand questions about the future of knowledge. They arrive as proposals, contract language and offers to fund work libraries have wanted to complete for years. A company may offer to digitize a collection in return for permission to use the material for AI training. That can be genuinely attractive. Digitization is expensive, and a collection that only a handful of people can see in a reading room could become available to a student across the country, a scholar on another continent or a family trying to recover part of its history. AI could also make those collections easier to find and use. I want those things to happen.
The difficult part is what happens next. There are different ways to use AI with a collection. A retrieval-based system can be designed to search materials while keeping the source documents under institutional control and pointing users back to them. Training a general purpose model is different. We can keep a record of what material went into the training process, but current systems cannot reliably tell us that a particular statement in an answer came from a particular item in the training data. That difference sounds technical, but it matters if a donor restriction surfaces later, a community objects to a use no one anticipated or the AI system misrepresents the collection. The question then becomes: What can the university still do?
We spend a great deal of time asking what happens when AI gets the answer wrong. I am increasingly interested in what happens when it gets the answer right. Imagine asking an AI system about an unpublished letter in a university archive. It accurately tells you when the letter was written, who wrote it and what happened. But you cannot see the sentence the writer crossed out, read another letter in the same folder that complicates the story or determine what conditions came with the collection. The answer may be correct, yet something important has been lost.
I call this memory without origin. The knowledge survives and circulates, but the connection between what we know and how we know it becomes harder to recover. This is part of a larger change I think of as the end of search. Of course, search is not literally ending. I still search every day. But increasingly, we ask a question and receive an answer before we encounter any of the sources behind it. For many everyday questions, that may be perfectly fine. Universities should care about the cases where it isn’t.
For higher education, this is more than an archival problem. Universities are built around practices that connect claims to evidence. We ask students to show their work, scholars to cite their sources, researchers to document their methods and institutions to preserve the records that make knowledge contestable. If AI becomes a primary way students and researchers encounter information, then maintaining that connection becomes part of the university’s teaching and research mission.
At the University of Virginia, we developed an Archival AI Protocol to help us think through these decisions. It asks fairly basic questions: Can someone get from an AI answer back to the source? Are we still honoring promises made to donors and communities? If something goes wrong later, what can the university still do? We have also worked with other libraries and archives on a Statement of Shared Practice so institutions can compare what they are being asked to do rather than negotiate every AI training proposal alone.
I do not think universities should keep their collections away from AI. AI could help a student discover someone who has disappeared from the historical record, or allow a scholar to work across thousands of documents that no individual could read in a lifetime. But access is not the only thing universities are responsible for. We are also responsible for helping people understand where knowledge comes from, how claims were produced and whether they should trust them.
That responsibility becomes more important when the answer arrives first. The tracking device in that rare book shipment answered one question: Where did the book go? Future readers will face a harder one: When its words return to us as an answer, will we still know where they came from?
