The cases involving Meta and Anthropic both tested the consequential legal question of whether tech companies can freely use copyrighted content to train their models.
Meta copied extensive amounts of copyrighted content, including a substantial number of books, from so-called “shadow libraries” – online databases of pirated books. Although the company reportedly employed “post-training” efforts to reduce verbatim reproduction, it did not seek or obtain permission from copyright holders.
In response, 13 authors – writers of novels, plays, memoirs, short stories, and nonfiction – filed a lawsuit alleging copyright infringement. The key question: did Meta’s use of copyrighted materials qualify as “fair use” under US law? The court dismissed the lawsuit, finding that the authors failed to prove that Meta’s models infringed on their works.
Authors similarly accused Anthropic of training its generative AI bot, Claude, with databases of copyrighted materials. Judge William Alsup ruled in favor of the tech giant, arguing that Anthropic’s use of copyrighted materials was “quintessentially transformative.”
Earlier this year, music publishers charged Anthropic with enabling the unauthorized reproduction of song lyrics. Here, too, the federal court ruled in favor of the AI giant, arguing that the music publishers failed to show that Anthropic caused them “irreparable harm.” Anthropic agreed to implement “guardrails” to prevent Claude from infringing on song lyrics.
These rulings go to the heart of a fundamental philosophical clash over “fair use,” the bedrock doctrine of US copyright law. Originally intended to balance the rights of creators with the broad public interest, fair use permits limited unlicensed used of snippets of copyrighted material for commentary, criticism, or research.
The AI revolution has turned this well-established principle into a legal battleground. Can ingesting vast troves of copyrighted books, music, and articles to generate new works, answer questions, and eventually compete with the original authors, be deemed “fair use”?
Get the Latest
Sign up to receive regular Bandwidth emails and stay informed about CEPA’s work.
Neither the Meta nor the Anthropic case offers a definitive answer. Both rest on answers to two key questions. To what extent can it be considered “fair use” for generative AI tools to utilize copyrighted materials to create new works? And to what extent are the AI-created materials “transformative” enough to be deemed distinct from the copyrighted materials they were trained on? AI models don’t merely use authors’ and artists’ works. They potentially replace or even compete with them.
The two rulings reveal a divergence in how courts are approaching the question. In Bartz v. Anthropic, Judge Alsup ruled that using copyrighted materials to train large language models and produce new content was transformative and “fair use.” But the judge ruled it illegal to use pirated copies. In the Meta case, Judge Vince Chhabria ruled that copying served a single, unified goal: to develop and train its AI models, which he deemed fair use.
While the outcomes are the same, the divergence in judicial reasoning leaves a key issue unresolved: should fair use be assessed for each discrete act in the AI pipeline, or only in terms of the end use? As more cases proceed through the courts, this question will become central.
Importantly, the rulings did not address potential economic harm. Although authors acknowledge that AI models may not reproduce their text word for word, they argue that derivative AI content diminishes the value of their original works. In trademark law, dilution refers to the erosion of a brand’s distinctiveness or value, even in the absence of direct competition or consumer confusion.
Judge Vince Chhabria focused narrowly on whether Meta’s models directly copied or replicated specific passages, and whether the purpose of that copying was transformative enough to qualify as fair use. In Bartz v. Anthropic, Judge Allsup stopped short of addressing the broad economic question: what happens to the creative economy when AI systems trained on copyrighted material begin to generate outputs that compete with, or replace, the originals?
This legal blind spot will become increasingly untenable. Without recognizing some form of “market dilution,” the law may fail to account for the systemic effects of generative AI on authors. Future rulings, and perhaps legislative reform, will need to grapple with whether fair use should continue to apply when the long-term result could be the commodification of human creativity.
These early rulings are far from the final word. Congress may yet step in. Appeals are inevitable. This is not just a conflict about intellectual property. It’s about whether the rules of the analog era can be applied to the digital age. Most boxing matches last 12 or 15 rounds. Tech has won Round One. But it has not delivered a knockout blow.
Elly Rostoum is a Google Public Policy Fellow with the Center for European Policy Analysis (CEPA). She is a Lecturer at Johns Hopkins University. You can find out more about her work here: www.EllyRostoum.com
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
Read More From Bandwidth
CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy.
