OpenAI has used historical texts digitised from Oxford University’s Bodleian Library to help populate its AI training set, according to The Guardian. The important wrinkle is that Oxford’s original partnership announcement did not clearly say the material would be used for model training, prompting staff concerns about transparency, reputation and energy use.
OpenAI Watch analysis
What happened
Internal documents obtained through a freedom of information request reportedly record that material digitised by OpenAI was used to “populate the OpenAI training set”. By June 2025, 125,000 images from historical dissertations had been shared, alongside a collection of 10,000 sixteenth-century broadside ballads. Further digitisation under discussion includes eighteenth-century Irish state papers, Marie Edgeworth’s private letters and Dorothy Hodgkin’s penicillin notebooks.
The partnership began in March 2025, when Oxford said OpenAI software would digitise texts from the Bodleian and make them more accessible to students and researchers. Oxford’s spokesperson said the material was modest in scale, out of copyright and non-exclusive. The Bodleian retains rights to the scans and plans to publish them openly online in the coming months. OpenAI said it was proud to help ensure that today’s AI models reflect historical knowledge, cultures and perspectives.
Why it matters
This is not simply a story about an old library meeting a new model. It is about whether digitisation, public access and AI training are being presented as one tidy bargain when they may involve different permissions, risks and public expectations.
The material described is out of copyright, so the central issue is not a straightforward claim of copyright infringement. It is consent and disclosure: what did the university agree to, what did staff and the public understand, and how much access should a partner receive before the paperwork becomes a footnote?
Our read
The partnership has a respectable public benefit: rare material is being digitised and may become easier to access. But “the scans will be more available” does not automatically answer “how will they be used to train commercial models?” Libraries are trusted partly because they make such distinctions carefully. The eyebrow rises when the training-data consequence appears in internal minutes rather than the headline announcement.
What to watch
- Publication of the scans:
Whether the promised open release arrives on schedule and covers the material described. - Contract terms:
What the agreement says about training, reuse, exclusivity and future digitisation. - Collection scope:
Whether the project expands towards the Bodleian’s wider 23 million-item collection. - Institutional policy:
Whether Oxford or other universities change how they disclose AI-training partnerships. Sources and evidence: The Guardian reports the internal documents, staff concerns and statements from Oxford and OpenAI. The Guardian also reports that Oxford says the material is out of copyright and that the digitisation project’s primary purpose was access.
Discussion spark: Should universities be required to disclose AI training as a separate, explicit use whenever a digitisation partnership is announced, even when the material is out of copyright?
Sources and evidence
- Oxford lets OpenAI train its AI models on Bodleian library – The Guardian (26 September 2026, 11:00 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.