Oxford partners with OpenAI to train AI on historical texts
Bodleian Library provides 125,000 thesis scans for AI model training
1Oxford's AI partnership details
Oxford University's Bodleian Library has provided OpenAI with 125,000 scans of historical PhD theses for AI training. The deal, made public in March 2025, initially focused on digitizing rare texts for academic access. The Guardian obtained documents showing these scans were subsequently used for AI training, despite initial public communications not mentioning this application. The texts include 19th and 20th century theses from European and US universities.
2Controversy over undisclosed use
Internal staff meeting notes reveal concerns about Oxford's reputation and AI's energy consumption. While the university claims the scanning project was small-scale and used non-exclusive, out-of-copyright materials, some staff expressed doubts. A spokesperson clarified that while digitization was the primary goal, the AI training aspect wasn't hidden from staff. The library retains rights to the materials and plans to make scans publicly available online in coming months.
3Cultural representation rationale
OpenAI justifies using historical texts by emphasizing the need for AI to reflect diverse cultures and perspectives. With over a billion users interacting with AI daily, the company argues this approach improves model quality. Oxford joins OpenAI's NextGenAI group as the sole UK member, alongside institutions like MIT and University of Michigan. This partnership reflects growing academic involvement in shaping AI development through curated historical content.
4AI's growing data challenges
The Oxford deal highlights AI firms' increasing difficulty in finding quality training data as the web fills with AI-generated content. Companies now seek printed materials, sometimes destructively scanning books - a practice that has drawn criticism from booksellers. Unlike some controversial methods, the Bodleian agreement preserves physical books while providing digital access. This approach offers a potential model for balancing preservation with AI development needs.
5Industry-wide data sourcing trends
404 Media recently exposed Amazon's book-scanning operation for AI training, where rare books were destroyed after digitization. In contrast, the Bodleian agreement maintains book integrity while providing content. The AI industry's data sourcing methods face increasing scrutiny as scale requirements grow. Oxford's partnership represents a more transparent approach compared to some covert commercial operations acquiring training materials.
6Academic-commercial collaboration
The Oxford-OpenAI deal illustrates growing ties between academia and tech firms in AI development. While offering access to rare materials, universities gain digitization capabilities and potential research opportunities. However, such partnerships raise questions about academic independence and commercial influence. The arrangement's public disclosure and retention of library rights attempt to address these concerns while enabling collaboration.
7Staff concerns and transparency
Documents obtained through freedom of information requests reveal internal staff worries about reputational damage and environmental impact. While Oxford maintains transparency about the AI training use, the initial public framing focused solely on digitization benefits. This gap between internal knowledge and external communication highlights challenges in managing complex institutional partnerships with commercial entities in the AI space.
8Future of cultural AI training
The Oxford-OpenAI collaboration may set precedents for how cultural institutions contribute to AI development. By providing curated historical content while maintaining material rights, it offers a potential model for ethical partnerships. As AI systems increasingly mediate access to knowledge, such collaborations could shape how cultural heritage informs technology while preserving institutional values and public trust.
Original article: The Next Web — Oxford partners with OpenAI to train AI on historical texts