Cooperate And Compete: Archives Can Gain More By Expecting Less From AI Training
Abstract AI companies have increasingly approached archival organizations seeking access to unpublished or undigitized collections for model training. Many archival organizations have responded cautiously to these requests and lean toward rejection due to various concerns. Two major concerns are about the loss of item-level provenance in AI outputs and the loss of control over archival knowledge absorbed into AI models. This paper examines these two concerns and challenges some of the assumptions underlying them. It argues that the loss of item-level provenance in AI outputs may in fact constitute a competitive advantage for archival organizations. It analyzes the potential costs of refusing to work with AI companies and suggests that archival organizations negotiate agreements that preserve their competitive advantage while indemnifying them against liability arising from the misuse of archival knowledge incorporated into AI models. The paper further suggests that archival organizations tailor policies according to the legal status of individual collections, rely on RAG technology for reliable item-level citation, and advocate for government policies that encourage large AI companies to contribute to public goods. Such policies could help improve the public image of AI companies and motivate more archival organizations to collaborate with them.
Authors
- Jinfang Niu (ORCID: https://orcid.org/0000-0002-4446-4849)
Institutions
- University of South Florida (US)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11392603/v1
- Primary Topic
- Digital and Traditional Archives Management
- Type
- preprint