Anthropic Tells a Judge That Training Claude on Song Lyrics Is Fair Use. The Music Publishers Are About to Find Out Why That Matters to Everyone Else.

Anthropic filed a brief on April 20, 2026, telling a federal judge that training Claude on copyrighted song lyrics from Universal Music Group and the other major publishers is transformative fair use, that the publishers cannot seriously argue otherwise, and that the case should end. UMG sued back in 2023. The judge is now sitting on cross-motions that will produce one of the first real fair-use rulings in a music copyright case against an AI lab.

The stakes go well beyond publishing. If Anthropic wins, every foundation model in the market gets a template for defending its training data. If UMG wins, licensing becomes the cost of building a frontier model, and the companies that cannot afford to license get priced out of the conversation.

What Anthropic Actually Said

The brief makes three arguments, all of them calculated.

First, Claude ingests lyrics “alongside trillions of other words to understand the interrelationships between words and concepts in human language.” The point is not that Claude memorized “Purple Rain.” The point is that Claude learned what rhymes, what scans, what works emotionally, and it learned that from ingesting a universe of text in which song lyrics are a rounding error. That framing is designed to make the copying look incidental, statistical, and non-expressive.

Second, the use is transformative because “the vast majority” of what Claude does “is wholly unrelated to lyrics or music.” Users ask Claude to write code, summarize depositions, draft memos, and outline research. The argument is that a general-purpose model trained on everything cannot be reduced to a tool that competes with a song catalog.

Third, there is no market harm. And for this one, Anthropic did something that should make every in-house lawyer at UMG wince. It quoted UMG’s own chief digital officer, Michael Nash, from the company’s last earnings call, saying “thoughtful analysis will conclude that the impact AI will have on our business will be overwhelmingly net positive.” When your own executive tells investors AI is a win, you have made opposing counsel’s closing argument for them.

Why the Publishers’ Response Is Weaker Than It Looks

UMG’s spokesperson gave Billboard a fighting quote: “There is no excuse for Anthropic’s blatant infringement of Publishers’ copyrighted song lyrics.” That plays well in a press release. In a summary judgment brief, it is not an argument.

The publishers have a real case on one specific point, which is output. If Claude can be prompted to regurgitate substantial portions of copyrighted lyrics verbatim, that is a problem the fair use defense does not clean up. Courts have been consistent that the transformative-use analysis breaks when the output reproduces protected expression. Anthropic knows this. That is why its brief focuses on training, not generation.

The output question is where these cases will actually be decided, and it is where the music publishers should be putting their pressure. Training data is a defensible use when the model is learning patterns. It stops being defensible the moment the model spits back the thing it learned from.

The Precedent Problem

Every AI copyright case currently pending is watching this one. The New York Times against OpenAI. The authors’ guild cases. Getty against Stability. The Suno and Udio cases brought by the labels themselves. A clean fair-use ruling for Anthropic lowers the ceiling for all of those plaintiffs. A ruling against Anthropic raises the floor for every AI company still training on scraped data.

Federal judges are not supposed to write precedent based on what other courts will do with it, but they are aware. Judge William Alsup’s Google Books ruling in 2015 held that digitizing and indexing 20 million books was transformative fair use because the purpose was different from the original. That ruling became the backbone of every AI training defense that followed. The publishers know this. They are trying to distinguish lyrics from books by arguing lyrics are shorter, more expressive per word, and more easily memorized and regurgitated. The argument is not frivolous. It is also not winning yet.

What This Means for GCs and Operators

If you are licensing AI tools, running them internally, or building products on top of foundation models, the fair use fight is not abstract. It determines your indemnity risk, your vendor due diligence, and your exposure if a copyright owner decides you are an easier target than the model maker.

Three things matter right now:

Indemnification clauses are the whole ballgame. OpenAI, Anthropic, Google, and Microsoft all offer some form of copyright indemnification for enterprise customers. The terms vary. Read them. If your vendor’s indemnity excludes cases where the user “caused” the infringement through specific prompts, you are carrying more risk than you think.

Outputs are your problem, not just the vendor’s. If your team generates content with an AI tool and ships it without review, and that content reproduces copyrighted material, you are the defendant. The vendor’s training defense does not protect you. Your logs, your workflow, and your review process will be the evidence.

The licensing market is forming whether vendors want it or not. Suno and Udio have already signed partial licensing deals with labels at the end of 2025. Anthropic and OpenAI will eventually do the same for certain verticals, regardless of how this ruling comes out. The companies that build pricing and procurement around “no licensing needed” are planning for a world that is already disappearing.

What to Do Now

• Pull every AI vendor contract you have and find the IP indemnification section. Flag the carve-outs.
• Require output review for any AI-generated content that will be published, distributed, or filed. Log the review.
• If you are in a regulated industry or a creative industry, start mapping which of your vendors have licensing deals with rights holders and which are relying on fair use.
• Stop treating the training question as separate from the output question. Plaintiffs are not going to.

The Anthropic case will not be the last word. It will be one of the first words in an argument that is going to run for the next three years, and the opening brief tells you how the AI labs plan to fight. They plan to fight on the ground that training is transformative, that general-purpose models cannot be reduced to substitutes, and that their own customers’ executives have already conceded the market-harm question. The publishers have better facts on output than they do on training. Whether they use them is going to decide this case.