Copyright debate intensifies over AI training data use

On October 30, 2023, venture capital firm Andreessen Horowitz (a16z) submitted formal comments to the U.S. Copyright Office regarding artificial intelligence and copyright law, marking a significant development in the ongoing debate about AI training data use. The submission came in response to the Office's August 30 Notice of Inquiry on Artificial Intelligence and Copyright.

According to the submitted document, a16z argues that using copyrighted content to train AI models constitutes fair use under existing law. The firm, which invests in AI startups, states that AI model training extracts statistical patterns and facts rather than storing copyrighted content, with research showing "extremely small rates of memorization."

The document reveals several key technical claims about AI training. According to a16z, large language models require training on "something approaching the entire corpus of the written word" to function effectively. The firm emphasizes that AI models "are not vast warehouses of copyrighted material," characterizing suggestions to the contrary as "a plain misunderstanding of the technology."

The submission addresses economic implications extensively. According to the document, "billions and billions of dollars" have been invested in AI development over the past decade, based on the understanding that extracting statistical facts from copyrighted works is permissible under current law. The firm warns that changing this interpretation could "significantly disrupt settled expectations" and potentially harm U.S. competitiveness.

National security concerns feature prominently in the argument. The document cites the Department of Defense's Third Offset Strategy, which identifies AI as a key component of national defense. According to a16z, China's aggressive integration of AI into military strategies and surveillance makes U.S. leadership in AI development crucial for national security.

The submission strongly opposes proposed licensing schemes for AI training data. The firm argues that such systems would prove "administratively impossible" to implement, citing the vast scale of content involved. For context, the document notes that while the Music Modernization Act deals with approximately 25 million musical works, AI training data encompasses "billions of pieces of text from millions of individual websites."

Critics have challenged these positions. A prominent figure in public relations and digital media, Kristen Ruby, responding on January 25, 2025, disputed a16z's interpretation, arguing that AI models do store copyrighted content during training. Ruby contests the characterization of training data use as "statistical facts extraction," maintaining that copyright infringement occurs regardless of how the data is processed.

The debate highlights fundamental questions about intellectual property rights in the AI era. The document reveals that even determining rightsholder identification for AI training data presents significant challenges, as identifying copyright owners for billions of internet-published works proves nearly impossible.

Financial implications for the AI industry remain contentious. According to the submission, any licensing framework providing more than "negligible payment" to rights holders would result in "tens or hundreds of billions of dollars a year in royalty payments." The firm argues this would advantage large technology companies while effectively barring smaller innovators from the field.

The submission concludes by emphasizing AI's potential societal benefits, citing applications in medical innovation, scientific research, and educational access. According to a16z, the technology could "improve the lives of everyone in a way that few other technologies—and maybe no other technologies—ever have."

This ongoing debate occurs against the backdrop of rapid AI development and deployment across industries. The U.S. Copyright Office's response to these arguments could significantly impact the future development of artificial intelligence technology and the broader digital economy.