Judge weighs whether Claude’s lyric training and synthetic songs are fair use

The story

A federal judge is considering whether Anthropic’s use of copyrighted song lyrics to train Claude is protected by fair use, after hearing competing requests to resolve a major part of the case without a trial. U.S. District Judge Eumi Lee heard arguments in San Jose on October 21 but did not rule. That restraint is important: the hearing tested the parties’ evidence and legal theories; it did not decide that AI training is lawful or that Anthropic infringed copyright.

The case was brought by music publishers including Concord, Universal-affiliated companies, Polygram and ABKCO. Their March motion for partial summary judgment focuses on 499 works and argues that Anthropic copied lyrics both into its training process and in Claude’s outputs. Anthropic filed its own summary-judgment motion, contending that training transforms a large corpus into a system that learns patterns of language and supports uses extending far beyond reproducing the original songs.

That creates two related but separable questions. The first is whether copying lyrics during model development serves a sufficiently transformative purpose under U.S. copyright law. The second is whether the deployed model reproduces lyrics, generates close derivatives or supplies synthetic songs that substitute for licensed or human-created material. A court could analyze those stages differently rather than announcing one rule for every act described as AI training.

At the hearing, the publishers emphasized the output market. Reuters reported their attorney said Claude received nearly 120,000 requests to produce synthetic lyrics over six months and warned that model-generated songs could crowd a market served by human writers. Anthropic responded that the publishers had not identified concrete market damage and that demand for the underlying lyrics had not fallen. The prompt count is evidence in a disputed record, not a judicial finding that 120,000 outputs infringed copyright.

The parties also disagree over responsibility for outputs. Courthouse News reported that Anthropic argued users initiate requests and that the company should not be treated as the direct actor each time Claude responds. The publishers counter that Anthropic built and trained the system to produce the disputed material. Lee questioned whether their position would effectively give rights holders control over new works inspired by an older style—an area where copyright protects original expression but generally does not grant ownership over style alone.

Fair use requires a contextual assessment of purpose and character, the nature of the copyrighted work, the amount used and the effect on actual or potential markets. Lyrics are highly expressive works, which strengthens the publishers’ position on the second factor. Anthropic’s transformative-purpose argument is directed mainly at the first. The most contested evidence may sit under the fourth factor: whether Claude’s behavior substitutes for authorized lyric services, weakens a viable training-license market or competes with derivative works that publishers are entitled to license.

The case arrives after early federal rulings produced different routes through the AI-training question. One Northern California judge found Anthropic’s use of books for training transformative while separating that issue from the acquisition of pirated copies. Another ruled for Meta on the record before him but warned that AI training would not automatically qualify as fair use when plaintiffs can show meaningful market harm. Lee is now confronting a record centered on short, expressive works and alleged output behavior.

The publishers’ filing says authorized lyric markets already include search services, streaming platforms, social networks, sheet music, karaoke, remixes and other derivative uses. Anthropic argues that Claude’s broader language capabilities are not a replacement for those products and that synthetic outputs are new material. The dispute therefore tests whether a claimed new purpose is enough when the same system can also return protected expression or generate material aimed at an adjacent creative market.

INNOVOX analysis: the likely significance reaches beyond music. Developers have often framed training as one technical act, while rights holders describe a chain from data acquisition to model output and commercial substitution. The stronger legal and engineering approach is to examine that chain component by component. Provenance controls, memorization testing, prompt-level safeguards and auditable response policies may become commercially important even if courts accept that some training uses are transformative.

What to watch next is whether Lee divides the dispute cleanly. She could rule on training while leaving output liability or market harm for trial, grant one party broader summary judgment, or find factual conflicts that a jury must resolve. The wording will matter as much as the result: a narrow decision tied to lyrics and Claude’s documented behavior would guide model developers without settling every AI copyright case, while a broad treatment of licensing markets could reshape how training data is acquired across the industry.

INNOVOX analysis

The hearing exposes why a single slogan about AI training is inadequate. A court may view statistical learning as transformative while treating memorized or substitutive outputs differently. The durable business question is therefore not only whether training is fair use, but whether developers can prove that acquisition, model behavior and deployment controls keep protected expression from becoming an unlicensed product feature.

What to watch

Watch for Judge Lee’s ruling on the cross-motions, whether she separates training from output liability, how she treats lyric-licensing and synthetic-song markets, and whether disputed facts require a trial. Any appeal could give the Ninth Circuit its first major opportunity to shape generative-AI fair use.