What happened
Oxford announced that a partnership established in March 2025 will see OpenAI digitise and use historical texts from the Bodleian Library to populate its AI training set. By June 2025, 125,000 scanned images—including 19th‑century dissertations, 16th‑century broadside ballads, and other out‑of‑copyright works—had been shared with OpenAI. The university plans to publish the scans openly online within months, retaining rights to the material. Meeting minutes obtained via a freedom‑of‑information request reveal staff concerns about reputational risk and the energy‑intensive nature of the deal. Oxford is the sole UK participant in OpenAI’s “NextGenAI” project, which already includes US institutions such as MIT and the University of Michigan.
The University of Oxford’s Bodleian Library entered a data‑sharing agreement with OpenAI that builds on a 2025 partnership aimed at digitising rare and out‑of‑copyright texts. Internal documents show that the scanned material has already been incorporated into OpenAI’s training , a step confirmed by a spokesperson who said the company is "proud" to preserve historical knowledge for future AI models.
The digitisation effort has produced roughly 125,000 high‑resolution images of dissertations, ballads, and other historical documents. Staff minutes reveal concerns about reputational risk and the environmental impact of training large language models, which are known to be energy‑intensive. The university retains the rights to the scans and intends to publish them openly online, ensuring that the material remains accessible to scholars and the public.
Oxford is the only UK institution participating in OpenAI’s NextGenAI initiative, which already includes US libraries such as the Boston Public Library and MIT. The partnership also discusses the creation of an "Ask the Bod" chatbot, potentially allowing users to query the digitised collection directly via natural language.
Source details: theguardian.com ↗
Why it matters
The agreement marks a rare instance of a leading UK university directly supplying training data to a major AI developer, highlighting how AI firms are turning to physical, historical collections as web‑scraped data becomes saturated with AI‑generated content. By feeding large‑scale language models with culturally diverse, out‑of‑copyright texts, OpenAI aims to improve representation of different histories and perspectives in its products. However, the deal raises several open questions: the environmental footprint of the digitisation and training , the adequacy of consent and governance for using academic heritage, and the potential for future commercial exploitation of publicly funded collections. The partnership also sets a precedent for other libraries, which may face pressure to balance open access with the risk of their holdings becoming commodified training data.
The deal illustrates a shift in AI data acquisition strategies: as publicly available web content becomes increasingly polluted with AI‑generated text, developers are turning to curated, historical archives to enrich model training with authentic cultural artifacts. This could improve the factual grounding and diversity of AI outputs, but it also raises questions about who controls the curation and how benefits are shared.
Environmental concerns are salient because training large language models consumes significant electricity. Oxford staff highlighted the need to reconcile the university’s sustainability commitments with the energy demands of the partnership, a tension that may influence future collaborations between academic institutions and AI firms.
The agreement sets a for how libraries might negotiate data‑sharing terms that preserve open access while allowing commercial AI use. If other institutions follow suit, the landscape of AI training data could shift dramatically, potentially reshaping the economics of model development and the legal frameworks governing cultural heritage.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
In AI, what are a model's "parameters"?
What to watch next
Future scrutiny of the NextGenAI programme, especially any policy responses from UK research funders or cultural heritage bodies; whether Oxford will expand the scope beyond out‑of‑copyright works; the impact of the “Ask the Bod” chatbot prototype on public engagement with the library’s collections; and any legal or ethical challenges concerning the use of digitised academic material for commercial AI training. Monitoring OpenAI’s public disclosures about how the Bodleian data influences model performance will also be essential for assessing the tangible benefits of the partnership.
Policy responses from UK cultural heritage agencies or research funders that could impose licensing or transparency requirements on similar data‑sharing deals.
The rollout and public reception of the "Ask the Bod" chatbot, which could serve as a test case for AI‑mediated access to scholarly collections.
Any expansion of the partnership to include copyrighted or more recent works, which would raise additional legal and ethical considerations.
OpenAI’s future disclosures about the performance impact of the Bodleian data, which will indicate whether the partnership delivers measurable improvements in model behavior.