OpenAI Used Oxford’s Bodleian Texts for AI Training, Documents Show

Historical texts from Oxford’s Bodleian Library have entered OpenAI’s model-training set, according to internal documents reported by the Guardian. That use was absent from Oxford’s public announcement of the partnership, which focused on making the texts more accessible. Oxford disputes the suggestion that it hid the training element.

How the texts reached OpenAI

Oxford announced the partnership in March 2025. OpenAI would help digitise Bodleian texts, and Oxford said the work would make them more widely available to students and researchers. The newly reported documents establish a second use: material digitised through the project was used to populate OpenAI’s training set.

Digitising a library collection makes its contents available in a form that can be shared beyond the reading room. Model training serves a different purpose: feeding material into systems that learn patterns in words. The documents connect the Bodleian project to that training process, but they do not identify which particular scanned works entered which models.

What was scanned

The shared dissertation images came from theses written at European and American universities in the 19th and 20th centuries. Texts scanned through the project also include a collection of 10,000 16th-century broadside ballads, with song lyrics and musical notes. These details show the range of historical material involved; they do not show that each work was used to train a model.

Oxford is the only UK member of OpenAI’s NextGenAI project, which also includes US research libraries such as Boston Public Library, MIT and the University of Michigan. The Bodleian arrangement is therefore part of a wider effort involving library collections, though the newly reported documents concern Oxford’s own material.

The disclosure Oxford disputes

Oxford’s March 2025 announcement did not say the digitised material would be used to train OpenAI models. A university spokesperson rejected the suggestion that this use had been hidden from students or the public. Digitisation was Oxford’s primary interest, the spokesperson said, but staff had been open that the project would also provide training data.

Meeting minutes obtained through a freedom of information request record staff concerns, including from members of the Bodleian governance committee. They raised the reputational risk of partnering with OpenAI and questioned how a deal involving energy-intensive technology sat with Oxford’s environmental commitments. The minutes record concerns, not a measured environmental cost for this project.

The access Oxford promises

Oxford says the material being digitised is modest in scale, out of copyright, and not reserved for OpenAI’s exclusive use. The Bodleian retains rights to the scans, and Oxford says it will begin publishing them openly online within months. OpenAI used the material for training; the promised public release of the scans has yet to happen.

OpenAI says including historical knowledge in today’s models helps them reflect different cultures, histories and perspectives. Oxford’s case puts that aim beside a more immediate question for libraries: how clearly should they explain training access when they present a partnership chiefly as a way to open their collections?