The Core Issue: Data Acquisition and Copyright

Large Language Models (LLMs) require vast amounts of text for effective training. The origin of this data, particularly for models like ChatGPT, is often unclear. While much of the data is sourced from the internet, its legal protection remains a notable concern. The central question revolves around whether AI developers are using copyrighted content without proper authorization.

Key Point: The legality of using publicly accessible data for AI training is under intense debate, especially when copyrighted material is involved.

Accusations Against OpenAI: The AI Disclosure Project

OpenAI has faced persistent accusations of training its AI models using copyrighted content without permission. The “AI Disclosure Project,” a non-governmental organization focused on the societal impacts of AI, has recently published a paper alleging that OpenAI is increasingly using non-public books, for which it lacks licenses, to train its advanced AI models.

The study specifically points to works from the International O’Reilly publishing house, with whom OpenAI reportedly has no licensing agreement.The authors state, GPT-4O, Openai’s newer and more powerful model, shows a strong detection of paid O’reilly book content. Thay further note that older ChatGPT versions only recognized publicly accessible reading samples.

Key Point: The AI Disclosure Project’s study suggests that OpenAI’s GPT-4O model has been trained on copyrighted material from O’Reilly publishing house without proper licensing.

Methodology and Limitations

The study employed a method to determine if the AI model coudl differentiate between human-written text and AI-paraphrased versions of the same text. The rationale is that if the model can distinguish between the two, it suggests prior knowledge of the original text from its training data.

Though, the study authors acknowledge limitations, emphasizing that thier findings do not constitute definitive proof. They concede that OpenAI could have obtained the book extracts from users who input them into ChatGPT. Furthermore, the study did not include the latest version, GPT-4.5, leaving open the possibility that this model was not trained on the same data or was trained with a different amount of copyrighted material.

Key Point: While the study raises concerns, it admits the possibility that the AI model learned the content through user inputs rather than direct access to copyrighted material.

Legal Challenges and Broader Implications

The allegations against OpenAI are not isolated. The company is currently embroiled in several lawsuits in the United States regarding its training data practices and copyright compliance, according to TechCrunch.

These legal challenges underscore the broader implications of using copyrighted material in AI training and the potential for significant legal and financial repercussions for AI developers.

Meta’s Alleged Use of Pirated Content

OpenAI is not alone in facing such accusations. The Atlantic reported that Meta allegedly used an illegal online library, LibGen, containing millions of pirated books and scientific works to train its AI model “Llama 3.”

LibGen is one of the largest databases of pirated content online,boasting over 7.5 million books and 81 million scientific works. Among the pirated copies were books from Austrian authors, including Ingrid Brodnig, Barbi Markovic, Stefanie Sargnagel, and Wolf Haas.

Key Point: Meta is accused of using the LibGen database, containing millions of pirated books and scientific works, to train its Llama 3 AI model.

Internal Discussions at Meta: Licensing Concerns

Court documents presented by The Atlantic revealed that Meta employees discussed the use of books and scientific works for training purposes with numerous publishing houses. Though, internal messages reportedly indicated that licensing texts was considered inappropriate and incredibly slow.

Furthermore, it was suggested that relying on the argument of “fair use” in court would become untenable once even a single book was licensed. Under U.S.”fair use” regulations, extracts from copyrighted works may be used under certain circumstances, such as for teaching purposes and in the context of critical discussion.

key Point: Internal communications at Meta suggest that the company found licensing copyrighted material for AI training to be too slow and cumbersome.

The “Fair Use” Debate and Legal justifications

Meta argues that training AI models with this data constitutes “fair use” because LLMs “convert” the original material into new works. OpenAI has similarly invoked this argument, having also used data from LibGen in the past. However, the legality of this justification is being challenged in several lawsuits against Meta in the U.S.

Court documents reported by Ars Technica in February indicate that Meta admitted to using data records with copyrighted books but claimed not to have disseminated them further.

Expert Viewpoint: Ingrid Brodnig’s Concerns

Tech expert Ingrid Brodnig views the situation as problematic, emphasizing that it involves a billion -dollar and listed group like Meta, who downloaded a database of millions of books on a large scale to improve its own commercial products rather than individual users accessing copyrighted material.

Brodnig suggests that interest representatives, such as unions and writers’ associations, should pursue legal clarification on the matter. She believes that large AI companies are currently operating according to the motto ‘Move Fast and Break Things’, making it crucial to establish clear legal boundaries.