Is the US 'Fair Use' Doctrine a Sufficient Legal Framework for the Generative AI Revolution?

The generative artificial intelligence revolution is built on a foundational legal gamble: that ingesting billions of copyrighted works to train foundation models is protected by the United States' "fair use" doctrine. For technology companies, this mass ingestion is a highly transformative extraction of uncopyrightable patterns, akin to a human reading a vast library to learn how to write. For authors, artists, and publishers, it represents the largest unauthorized reproduction of intellectual property in history, designed to create automated machines that directly compete with human creators. The question is no longer just theoretical. With over 50 lawsuits currently moving through federal courts, the U.S. legal system is being forced to decide if a flexible, 19th-century copyright doctrine can adequately govern a 21st-century automated economy. The stakes are existential for both sides: a ruling against fair use could throttle a transformative technology, while a blanket endorsement could permanently sever the economic incentives for human creation.[2][6]

The U.S. Copyright Office recently weighed in on this tension, releasing Part 3 of its comprehensive report on artificial intelligence in May 2025. The Office concluded that creating and deploying a generative AI system using copyright-protected material involves multiple acts that, absent a license or other defense, constitute prima facie infringement. Crucially, the report noted that even if the training data is discarded after the model is trained, the initial infringement analysis remains unchanged. This finding strikes at the heart of the AI industry's primary defense, shifting the legal burden entirely onto the four-factor fair use test to excuse the initial copying. By confirming that the ingestion phase is an infringing act by default, the Copyright Office has signaled that AI developers cannot simply bypass copyright law by claiming their models do not store exact copies of the input data.[1][3]

To survive this prima facie infringement, AI developers rely heavily on the first factor of the fair use test: whether the use is "transformative." In June 2025, two U.S. District Court judges in the Northern District of California issued pivotal decisions that initially seemed to validate this defense. In cases involving major AI developers, the courts recognized that using copyrighted works to train large language models is "highly transformative." The judges reasoned that the models are not reproducing the works for their intrinsic expressive value, but rather analyzing them to build statistical relationships between linguistic tokens. This interpretation aligns with earlier technology-driven fair use precedents, such as the Google Books case, suggesting that extracting uncopyrightable facts and patterns from copyrighted text is a protected activity that serves the broader public interest in innovation.[2]

However, legal scholars argue this "transformative" narrative relies on an illusory anthropomorphism of how artificial intelligence actually functions. As detailed in a 2025 Vanderbilt Law Review article by Jacqueline C. Charlesworth, AI models do not "learn" or "reason" independently of their training data in any human sense. Instead, the copyrighted works are algorithmically encoded, token by token, into the model's vectors. Because the model's output is entirely a function of these copied materials, the argument that the training works are merely "jettisoned" after processing ignores the mechanical reality of how the copyrighted expression is stored and exploited. When a model generates output, it is relying directly on the encoded representations of the training works, undermining the claim that the original expression has been entirely left behind.[4]

The most significant hurdle for the fair use defense is the fourth statutory factor: the effect on the potential market for the original work. The recent Northern District of California decisions underscored that even if training is deemed transformative, market harm remains a pivotal and potentially fatal element. Generative AI systems are increasingly used to produce outputs that directly displace the market for the ingested works. Whether by generating competing software code, writing in the specific style of a syndicated author, or utilizing retrieval-augmented generation to serve up exact excerpts from news articles, these models are acting as market substitutes. If an AI system can fulfill the consumer demand that would otherwise go to the original creator, the fair use defense traditionally collapses, regardless of how technologically innovative the copying process might be.[2][4]

Courts are also beginning to draw sharp legal distinctions based on the provenance of the training data, adding another layer of complexity to the fair use analysis. In the recent Anthropic case, the court distinguished between lawfully acquired copies of works and pirated digital libraries. The judge denied summary judgment for the AI developer regarding the use of unauthorized, pirated copies, noting that such use displaces demand for the authors' books and threatens the entire publishing market. This introduces a critical vulnerability for foundation models trained on massive, unvetted internet scrapes. If fair use cannot excuse the ingestion of unlawfully acquired data, AI developers face massive liability for the contents of open-source datasets like Books3, forcing a retroactive reckoning over how the current generation of frontier models was built.[2][6]

If fair use is ultimately deemed an insufficient framework, the alternative is a mandatory licensing regime. Proponents of licensing argue it is the only way to ensure creators are compensated for the value they provide to commercial AI systems, proposing solutions like extended collective licensing. However, AI advocates warn that requiring explicit licenses for the sheer volume and diversity of content necessary to train frontier models—often trillions of tokens—is practically impossible. They caution that a strict licensing mandate would stifle innovation, entrench incumbent tech giants who possess the capital to buy data access, and severely damage U.S. competitiveness against nations with more permissive text-and-data-mining exceptions. Ultimately, the U.S. is attempting to solve a macroeconomic structural shift using a microeconomic legal tool, forcing courts to decide whether fair use can stretch to cover the wholesale transfer of value from the creative economy to the technology sector.[1][5]

As the legal landscape fractures, the technology industry is quietly preparing for a hybrid future where fair use and licensing coexist. Major AI developers have already begun signing multi-million-dollar licensing agreements with large publishers and stock image repositories, effectively hedging their bets against an adverse Supreme Court ruling. This dual-track approach suggests that while companies will continue to aggressively litigate their fair use rights for broad internet scraping, they acknowledge that premium, high-fidelity data requires a commercial transaction. The ultimate resolution will likely require congressional intervention, as the current strategy of litigating 50 separate cases through various appellate circuits promises years of uncertainty. Until a unified legal framework emerges, the generative AI revolution will continue to operate under a cloud of copyright liability, balancing unprecedented technological capability against the unresolved property rights of the human creators who fueled it.[3][7]