Legal Battles Over Generative AI Training Data Highlight Copyright Fair Use Challenges

The rapid expansion of generative artificial intelligence (AI) technologies has sparked significant legal debate concerning the use of copyrighted materials in AI training datasets. Copyright holders assert they deserve fair compensation for the use of their works, which contribute to the capabilities and quality of AI models. Conversely, some technologists warn that imposing copyright liability costs on AI developers could severely hinder or even halt AI innovation.

This tension has led to multiple lawsuits against generative AI companies, alleging unauthorized use of copyrighted works in training data. Plaintiffs must prove two elements to establish infringement: ownership of valid copyrights and that the defendants copied their works by demonstrating likely access and substantial similarity.

Access claims rely heavily on the defendants’ failure to disclose detailed training data and the extensive web scraping practices commonly used to collect data. However, proving copying remains challenging because AI training often does not involve direct retention of original works. Instead, courts must assess whether AI outputs are substantially similar to copyrighted inputs, a complex and nuanced inquiry.

Notably, copyright law does not protect artistic style, and while AI models can generate works mimicking specific artists or authors, this alone does not constitute infringement. Courts have dismissed some claims on these grounds, indicating that individual output comparisons may be insufficient to resolve large-scale copyright concerns.

The central legal question is whether the mass use of copyrighted materials to train AI models amounts to direct infringement and, if so, whether such use qualifies as fair use. The 2023 U.S. District Court for the District of Delaware decision in Thomson Reuters Enterprise Centre et al. v. ROSS Intelligence was the first to address this issue directly. Reuters alleged that ROSS Intelligence’s AI model was trained on proprietary headnotes from Reuters’ Westlaw legal research service.

The court denied summary judgment, ruling that a jury must decide if the headnotes are protected by copyright and whether ROSS’s use falls under fair use. However, the court’s detailed fair use analysis offers critical insights into judicial approaches to AI training data.

The fair use doctrine requires courts to evaluate four factors collectively: (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of the portion used, and (4) the effect on the market for the original work.

Regarding the first factor, Reuters argued that ROSS’s commercial use weighs heavily against fair use, citing the Supreme Court’s recent decision in Warhol v. Goldsmith, which emphasized that transformative use is a matter of degree and not dispositive, especially in commercial contexts.

Historically, courts have found transformative uses in digital contexts, such as the Second Circuit’s ruling that Google’s scanning and snippet display of books augmented public knowledge without substituting the originals. However, generative AI outputs differ because they typically do not provide direct attribution or references to source materials.

The Supreme Court’s foundational fair use case, Campbell v. Acuff-Rose Music, recognized that transformative works generally further copyright’s goals but cautioned against rigid statutory applications that stifle creativity. Yet, the Court’s recent rulings underscore that transformative use alone does not guarantee fair use, particularly when commercial interests are involved.

Given the novelty and complexity of AI technologies, lower courts currently face the challenge of applying these principles to ongoing litigation involving generative AI models. The outcomes of these cases will shape the legal landscape for AI development and copyright enforcement.

As the debate continues, stakeholders await further judicial guidance on whether mass ingestion of copyrighted works for AI training constitutes infringement or qualifies as fair use. The resolution of these issues will have profound implications for the future of AI innovation and intellectual property rights.