AI Training on Copyrighted Data: Is It Fair Use?

The debate over whether training generative artificial intelligence on copyrighted books, art, and code constitutes fair use is the defining legal battle of the modern tech era. This article explores the core arguments of this conflict, examining how the four traditional pillars of copyright fair use apply to machine learning. While AI developers argue that ingesting vast datasets to create new content is a transformative and legal process, creators contend that this practice amounts to unauthorized exploitation that threatens their livelihoods. Ultimately, the resolution of this issue hinges on upcoming court rulings that will determine how traditional copyright laws adapt to artificial intelligence.

At the center of this dispute is the legal doctrine of fair use, primarily rooted in United States copyright law. Fair use permits the unlicensed use of copyright-protected works in specific circumstances. Courts evaluate fair use by weighing four main factors: the purpose and character of the use, the nature of the copyrighted work, the amount of the work used, and the effect of the use on the original work's market value.

The first factor, the purpose and character of the use, focuses heavily on whether the new work is "transformative." AI companies argue that training a model is the ultimate transformative use. Instead of merely repackaging books, art, or code, the AI analyzes patterns, structures, and styles to generate entirely new outputs. They often point to the landmark Authors Guild v. Google case, where the courts ruled that Google's scanning of millions of copyrighted books to create a searchable database was transformative and therefore fair use. AI developers claim their models similarly learn from data rather than copy it for redistribution.

Conversely, authors, artists, and software developers argue that the process is not transformative, particularly when look-alike art, derivative code, or text written "in the style of" a specific author is generated. They assert that the AI is only capable of functioning because it made direct, unauthorized copies of their work during the ingestion phase.

The second and third factors look at the nature of the copyrighted work and the amount used. Creative works like fiction and visual art normally receive stronger copyright protection than factual ones. Furthermore, AI training requires copying the entirety of a text, image, or codebase. While copying an entire work usually weighs against fair use, courts have historically allowed wholesale copying if it is necessary to achieve a transformative purpose, such as reverse engineering software for compatibility or indexing the internet for search engines.

The fourth factor—the effect of the use upon the potential market for the copyrighted work—is frequently considered the most critical. This is where creators have their strongest argument. Artists and writers claim that generative AI creates a direct market substitute for their labor. If an AI trained on an illustrator’s portfolio can instantly generate unlimited images in that exact style for free, the commercial market for the original illustrator is severely damaged. Similarly, if an AI coding assistant reproduces functional blocks of copyrighted code, it diminishes the value of the original software.

AI companies counter that copyright law does not protect an artist's general style, only specific expressions. They argue that competing in the same general market is not the same as usurping the market for a specific, original work, provided the AI output does not directly plagiarize the training data.

Currently, there is no definitive answer to whether AI training is fair use. Dozens of high-profile lawsuits—involving major players like OpenAI, Stability AI, GitHub, the New York Times, and various artists' guilds—are actively working their way through the courts. Because fair use is evaluated on a case-by-case basis, the final legal precedent will likely draw a distinction between how an AI is trained and what an AI outputs. Training a model might eventually be deemed fair use, while generating outputs that closely mimic copyrighted inputs could be ruled as infringement. Until the courts issue final rulings or lawmakers introduce new legislation, the legality of training AI on copyrighted data remains an unresolved gray area.