OpenAI Accused of Training GPT-4o on Paywalled O'Reilly Books
Researchers affiliated with O'Reilly Media published findings reportedly showing that OpenAI's GPT-4o model was trained on paywalled O'Reilly books without a licensing agreement, using statistical analysis to detect copyrighted content in the model's training data. The research echoes similar accusations from The New York Times against OpenAI.
What happened
The AI Disclosures Project, co-founded by O'Reilly Media CEO Tim O'Reilly, published research claiming OpenAI trained GPT-4o on paywalled O'Reilly books without a licensing agreement. Using a method called DE-COP (membership inference attack), researchers analyzed 13,962 paragraph excerpts from 34 O'Reilly books to detect whether the model was trained on non-public copyrighted content.
Researchers found that GPT-4o showed strong recognition of paywalled content from O'Reilly books, performing significantly better on paywalled chapters than on public samples, suggesting the model was trained on content that was not publicly accessible. O'Reilly noted that OpenAI does enter into licensing arrangements with some publishers, making it unclear why the company did not seek permission for O'Reilly's books.
Reported impact
O'Reilly expressed concern about the broader implications for content creators, stating that if AI systems extract value from content without fair compensation, they risk depleting the resources upon which their systems depend, potentially creating an 'extractive dead end' for the internet's content ecosystem.
Index analysis
The research used a statistical technique to infer whether paywalled content was included in training data by testing the model's ability to distinguish human-authored texts from paraphrased versions. GPT-4o demonstrated significantly greater recognition of paywalled O'Reilly material compared to older models like GPT-3.5 Turbo.
Governance controls
O'Reilly described the issue as 'systemic,' noting that Meta allegedly trained models on LibGen (pirated books) and Anthropic used 'the Pile' dataset which also contains pirated books. O'Reilly called for changes in data collection and usage practices across AI model developers.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.