To train artificial intelligence models, developers need millions of pieces of “training data,” or material like books, webpages, articles and spreadsheets. When AI companies started to run out of free access materials online, some turned to library systems to access their archival collections.
This prompted Leo S. Lo, dean of libraries at the University of Virginia, to launch a Statement of Shared Practice regarding AI training requests with 11 other universities, including Tulane University, to “protect the integrity of unique cultural heritage materials as AI developers increasingly seek to access them.”
“What struck me was that the materials being requested are irreplaceable,” Lo said. “Unpublished letters, photographic archives, oral histories, manuscript drafts. In many cases they exist in only one place … That felt like something the archival community needed to address together, not one library at a time.”
The Shared Practice is a voluntary, 12-month commitment focusing on archival and special collections to establish standards for how institutions handle requests from AI companies to access their materials.
When considering AI training requests, the agreement outlines key priorities like transparency, clear definitions, heightened scrutiny to broad commercial training and a preference for retrieval-based approaches that keep source material under institutional control.
“After consideration and advice from across Tulane University Libraries and our service communities, we decided to become a founding signatory of the Statement of Shared Practice on AI & Archives,” Tulane Dean of Libraries Lindsay Cronk announced in a LinkedIn post. “No access without control.”
One of the signatories’ main concerns is that, as they are being trained, AI models “absorb” materials into their systems forever, according to the Shared Practice announcement. This means that it can be impossible to trace back the AI-generated response to the archival material it came from.
The shared agreement is not a legal contract and does not restrict any institution’s ability to approve a request for AI training.
“Most institutions are evaluating these requests in isolation, without shared language or a common framework. The Statement addresses that gap,” Lo said in an online post. “It is a floor, not a ceiling.”