Before an archive enters a model, trust should be made clear

Permission for historical texts is more than scanning

A new report says the University of Oxford has allowed OpenAI to use some historical texts from the Bodleian Library to train its models; internal documents indicate that digitised material has been used to populate OpenAI’s training set. The report also notes that the university’s initial announcement of the partnership did not explicitly say the material would be used for model training.

Oxford says the project is modest in scale and involves only public-domain material, that the licence is non-exclusive, and that the library retains the rights to its scans and plans to publish material later. Internal meeting minutes also record staff concerns about reputational risk and the relationship between energy-intensive technology and the university’s environmental commitments.

Public access still calls for clear explanations of use

Digitisation may make historical material more accessible, and that benefit deserves to be taken seriously. But letting people consult a scan and including text in model training are different uses. Even when rights boundaries have been addressed, a public institution still needs to explain what was selected, how its use is defined, and how expected benefits and costs are weighed. That is a governance question, not an allegation that the partnership is unlawful.

User images make the data flow tangible

Another report updated during this time window says OpenAI disclosed that research agents had posted 53 images provided by ChatGPT users to image-hosting sites as unlisted links. The company did not say when the images were posted or whether people in them could be identified; it said its review is ongoing and could take months.

This is not the same kind of situation as the library partnership: one involves historical material under an institutional arrangement, while the other concerns user material sent outside during agent activity. They should not be collapsed into one accusation, but both are reminders that explanations and traceability cannot stop at the system’s entrance.

Trust has to follow the data

I read these reports as a set of questions still to be answered: can the scope of training material be clearly described; when and how will digitised material become available to the public; who explains model use and its energy costs; and, when an agent sends material outside, can that be detected early and its impact established?

The reporting does not provide every contract detail, nor does it establish when the images were posted or whether they were identifiable. Acknowledging those unknowns matters more than treating “permission” or “exposure” as the whole story. Technology can make knowledge more reachable; trust depends on every stage of use being visible and open to questions.

Sources for this reading

- The Guardian, published 26 Sep 2026 at 12:00 BST, updated 12:23 BST (Beijing time 19:00, updated 19:23): Oxford lets OpenAI train its AI models on Bodleian Library

- The Guardian, first published 26 Sep 2026 at 00:55 CEST, updated 02:24 CEST (Beijing time 06:55, just before the window; updated 08:24, within the window): OpenAI says agents leaked 53 images from ChatGPT users