FOI documents show Oxford's Bodleian digitisation deal fed OpenAI's training set

What happened: Internal University of Oxford documents obtained through a freedom of information request show that the university's March 2025 digitisation partnership with OpenAI was used to "populate the OpenAI training set" -- a use the public announcement did not spell out. The announcement framed the project as an access effort, saying OpenAI software would digitise Bodleian Library holdings for students and researchers. By June 2025, 125,000 images scanned from historical dissertations (19th- and 20th-century PhD theses from European and US universities) had been shared with OpenAI. Other material includes a rare collection of 10,000 16th-century broadside ballads (song lyrics and musical notes); staff have discussed digitising 18th-century Irish state papers, novelist Marie Edgeworth's private letters, and Dorothy Hodgkin's penicillin notebooks. Meeting minutes reference the prospect of mass digitisation of the Bodleian's 23 million items under the contract and plans for an "Ask the Bod" chatbot. Bodleian governance committee members raised concerns about reputational risk and the environmental cost of partnering with an energy-intensive AI company. Oxford is the only UK member of OpenAI's NextGenAI consortium (others include the Boston Public Library, Caltech, MIT and the University of Michigan). Oxford says the material is modest in scale, out of copyright, non-exclusive, and that the Bodleian retains rights and will publish the scans openly online; OpenAI said the work preserves historical knowledge and cultural perspective.

Why it matters to creators: institutional library collections are becoming commercial AI training data, sometimes without the training use being disclosed when the partnership is announced. The 19th- and 20th-century theses sit near the copyright boundary, so the "out-of-copyright" assertion and the consent and transparency questions around digitised cultural heritage matter to writers, researchers and rights holders. The deal also reflects the industry-wide shift toward physical archives as fresh training data as scraped web content becomes saturated with AI-generated text.

What is not yet known: whether any shared material was in fact still in copyright (the university asserts it is not); the full future scope of digitisation under the OpenAI contract; and how the Bodleian material was incorporated into model training.