Historical material digitised through the University of Oxford’s partnership with OpenAI has been used as training data for artificial intelligence models, bringing new attention to how academic collections are licensed and incorporated into commercial AI systems.
Internal university documents obtained through freedom of information requests and reported by the Guardian describe material from the Bodleian Libraries as being used to populate OpenAI’s training set.
The disclosure adds detail to a collaboration announced publicly in 2025. Oxford said at the time that the project would test large-scale digitisation of public-domain holdings and make previously unavailable material searchable and accessible to students and researchers.
The Bodleian’s current project description confirms that OpenAI is funding the work as part of its five-year partnership with Oxford and the wider NextGenAI consortium.
The research examines whether the library can increase the throughput of its digitisation operation, whether generative AI can improve metadata and full-text records, and how collections could be prioritised for future digitisation.
Documents reported by the Guardian show around 125,000 images scanned from historical dissertations had been shared with OpenAI by June 2025. Material involved in the wider digitisation work has also included early dissertations and historical broadside ballads.
An Oxford spokesperson said the material was “modest in scale”, out of copyright, and supplied on a non-exclusive basis. The university also said the Bodleian retains rights to the digitised scans and intends to publish the material openly online.
Oxford has disputed suggestions that the machine-learning element was concealed, arguing that the project had been open about material contributing to AI training alongside its primary objective of expanding digitisation.
The distinction between ownership of original works, copyright status, rights in digital reproductions, and permission to use data for model training has become increasingly important as AI developers seek higher-quality datasets.
Early large language models drew heavily on enormous amounts of material accessible through the open web. As model development has expanded, libraries, publishers, archives, media companies, and other rights holders have become more active participants in decisions about licensing and access.
Historical collections have particular value because much of their content does not exist in machine-readable form online. Public-domain status can remove copyright restrictions around an underlying work while leaving institutions with control over physical access, digitised images, metadata, and contractual arrangements governing newly created datasets.
Universities also have considerations extending beyond copyright. AI partnerships can provide finance, computing capability, specialist technology, and resources for digitisation programmes that institutions might otherwise struggle to fund at scale.
They can simultaneously raise questions around transparency, commercial benefit, environmental impact, data governance, institutional reputation, and whether material collected for scholarship should be available for commercial model development.
Those questions are becoming more prominent as universities integrate generative AI into research, teaching, administration, and digital scholarship. Oxford has separately expanded institutional access to AI tools, making the OpenAI relationship broader than the Bodleian project alone.
For developers, research-library partnerships provide structured and potentially unique datasets while reducing reliance on uncontrolled web material. Historical archives can also improve representation of languages, subjects, cultures, and periods that are underrepresented in conventional online corpora.
The commercial value attached to high-quality training data is consequently changing. Rights holders increasingly regard archives as assets that may be licensed or used within negotiated partnerships, while developers face pressure to show that material has been acquired under clear legal and contractual terms.
The Bodleian project remains a pilot rather than an attempt to digitise the entire library. Its research questions are focused on testing workflows, costs, metadata, and scalability across selected collections.
That limited scale does not remove the governance questions. Similar partnerships will increasingly need explicit terms covering training use, access, exclusivity, retention, publication, attribution, and ownership of digital outputs if cultural and academic institutions are to work with AI developers without ambiguity over how collections will be used.
The Bodleian plans to publish research arising from its digitisation work. Those findings should provide further evidence on whether large-scale AI-supported digitisation can improve access to academic collections while preserving institutional control over the resulting digital assets and their subsequent use.
You must be logged in to post a comment.