The University of Oxford has allowed OpenAI to train its AI models on historical texts and dissertations from its Bodleian Library, according to internal documents obtained via freedom of information requests. While the university publicly announced a digitization partnership with OpenAI in March 2025, public statements omitted that digitized materials would be fed into OpenAI’s training sets.
Meeting records indicate that university staff and governance committee members expressed concerns regarding potential reputational risks and the environmental impact of supporting energy-intensive AI training. By June 2025, 125,000 scanned images of historical PhD dissertations had been shared with OpenAI under its broader “NextGenAI” project with academic libraries.
An Oxford spokesperson stated that the digitized material is modest in scale and restricted to out-of-copyright works, with the library retaining ownership rights to publish the scans publicly online. The deal comes as AI developers increasingly turn to physical library archives to acquire uncorrupted, pre-digital training data.
Why it matters
AI developers are shifting toward offline academic archives as public web data becomes increasingly saturated with synthetic content.
Institutions face reputational and governance pressure when licensing proprietary library data to commercial AI developers without explicit public disclosures.
Source: theguardian.com



