Concord v. Anthropic: 120,000 lyric prompts and the first AI fair-use summary-judgment hearing since Bartz and Kadrey

Eight music publishers, led by Concord, Universal Music Publishing Group and ABKCO, spent several hours in a San Jose courtroom on 21 Oct 2026 arguing that Anthropic's Claude is not protected by fair use. US District Judge Eumi K. Lee heard cross-motions for summary judgment and took them under submission without saying which way she leans. The parties did not get through every motion, so there may be another session.

The litigation tracker ChatGPT Is Eating the World notes that Lee is the first district judge to hold a summary-judgment hearing on fair use in AI training since the two June 2025 rulings in Bartz and Kadrey. Whatever she decides will be read by every music AI case behind this one.

What is the case about?

The case covers lyrics to 499 songs. In their March 2026 motion, the publishers make three market-harm arguments:

• Lyrics on demand. Claude can output the lyrics themselves, which the publishers say substitutes for the licensed lyric market they supply through aggregators like LyricFind and Musixmatch and sites like Genius.
• Dilution. Claude produces "new" lyrics that, in the publishers' words, compete with and dilute the market for human-written songs.
• Training licences. The publishers say they have already licensed works for AI training, naming Udio and Klay Vision, and that unlicensed training undercuts that market.

Anthropic's own motion says training on those 499 songs, alongside billions of other works, is transformative fair use. It calls the competition argument "legally irrelevant", because fair use asks about substitution, not about whether a new tool lets more people write songs. It also describes cases where Claude actually reproduced lyrics as a vanishingly small number, and says the users who asked for them are the ones responsible.

Where does 120,000 come from?

At the hearing, the publishers' lawyer Matthew Oppenheim cited nearly 120,000 requests to Claude for "synthetic" lyrics over a six-month period, according to Reuters reporting relayed by AIMusicpreneur. The count had been blacked out in the March filing. Anthropic's lawyer Sonal Mehta answered that nothing in the record shows doom for the music industry. Lee's questions centred on market harm. She asked whether the publishers' theory would give them a monopoly, so that nobody, AI included, could write in an old song's style without a licence.

Why Bartz and Kadrey matter here

The two 2025 rulings disagree on exactly the point this case turns on.

In Bartz v. Anthropic, Judge William Alsup rejected the dilution argument. He compared the authors' complaint to objecting that teaching schoolchildren to write well would produce an explosion of competing works, and wrote that the Copyright Act does not protect authors against competition.

In Kadrey v. Meta, Judge Vince Chhabria came out the other way on the theory but ruled for Meta on the evidence. He called dilution the "potentially winning argument" and wrote that in cases like that one it will often decide the fair-use question. But the authors in that case "barely give this issue lip service", and offered no evidence of it. The same order also said rightsholders are not entitled to the market for licensing their works as training data, which cuts against the publishers' third argument here.

So the 120,000 figure is the publishers trying to supply what Kadrey lacked: evidence that the output is actually being used, at scale, to make substitutes for the kind of work they own.

What it means for songwriters

For a working songwriter, the question in this case is less "did Claude copy my lyric?" and more "is a machine now writing the songs I would have been paid to write?" The publishers' motion spells out the streaming version of that worry. Spotify and similar services pay from a fixed royalty pool, so every AI-generated track that earns streams takes a share that would otherwise go to human writers. The motion also argues that the flood could crowd out emerging songwriters before they get started.

The online reaction splits along the same line as the courtroom. One post on X framed the real question as whether AI companies can train on lyrics and then generate competing songs that dilute the original's market. The tech-industry group Chamber of Progress posted the opposite reading: that the publishers have not shown market harm.

One thing writers should notice: the plaintiffs are publishers, not songwriters. Any damages or licence fees would go through publishing deals, and how much reaches the writer depends on contracts most AI coverage never mentions. If you have a publishing deal, ask your publisher how AI licensing income is being treated under it.

What we think

Dilution is the theory that matters for songwriters, and this case is its best test so far. Lyrics are short, the genre conventions are tight and the output is easy to measure, which makes "120,000 synthetic lyrics" a far more concrete claim than the novelists in Kadrey ever put forward. If Lee accepts that AI output competing with human songs is a market harm, the leverage moves: training licences stop being a favour AI firms grant and become the price of avoiding litigation risk. Every music licensing talk would be priced against that.

The weak spot is the publishers' own side of the industry. Anthropic's motion points out that Universal Music Group, the parent of two publishers that hold 448 of the 499 songs, told investors on an earnings call (the month before Anthropic's April filing, according to that filing) that it saw no indication AI royalty dilution was a material revenue issue for UMG. Telling investors dilution isn't hurting you while telling a judge it is will be hard to square. It also hints at who would actually benefit: a large catalogue owner can absorb dilution and sell licences, while the songwriter competing for briefs and placements has no such cushion.

Our view: a win built on dilution should be used to fund the writers who bear that harm, not only to raise catalogue licence fees. Watch for whether any licence that follows this case pays out per writer, or just per catalogue.

A note on interest: WATMM's publishing workflow uses AI tools, including Anthropic's Claude, a party to this case.