Archive — US — techcrunch
Researchers suggest OpenAI trained AI models on paywalled O’Reilly books
A new paper claims OpenAI trained its GPT-4o AI model using copyrighted, paywalled O'Reilly books without appropriate licensing. The paper suggests that GPT-4o shows significant recognition of this content compared to earlier models, prompting concerns over copyright practices in AI training. OpenAI has faced multiple accusations regarding its reliance on protected materials in its models.
Key points
- OpenAI is accused of training its AI models, including GPT-4o, on copyrighted content without permission.
- A study by the AI Disclosures Project suggests a higher recognition of paywalled O'Reilly content by GPT-4o than by earlier models like GPT-3.5 Turbo.
- The method used in the investigation, DE-COP, indicates potential prior knowledge of copyrighted texts by the AI models.
- OpenAI's practices contrast with broader industry trends where companies seek licensure for training data, yet they continue to rely on unlicensed materials.
- Legal challenges are ongoing against OpenAI concerning its training data practices.
Why it matters
- The debate over AI copyright usage echoes ongoing discussions about intellectual property in the age of digital content.
- OpenAI's practices reflect a trend where companies increasingly prioritize access to data for training AI models amidst copyright law uncertainties.
- Since 2025, regulatory scrutiny and calls for tighter copyright frameworks for AI-generated content are expected to intensify.
The allegations against OpenAI highlight the challenges and ethical dilemmas in the AI industry regarding data usage, copyright laws, and the balance between innovation and respecting intellectual property rights, prompting an urgent need for clearer regulations.