Judge reins in music publishers: without concrete evidence on the 20,517 works, the lawsuit against Anthropic wobbles

By Courthouse News · November 4, 2026.

A federal judge in San José, California, voiced doubts on Wednesday about a lawsuit filed by several music publishers against Anthropic. Eumi Lee had indicated that she would deny Anthropic's motion to dismiss the case, but after hearing oral arguments she questioned whether the plaintiffs had laid out their claim in enough detail and asked for additional briefs, according to Courthouse News.

The core of the dispute: the publishers, among them Concord Music Group, Capitol CMG, Universal Music entities and ABKCO, accuse Anthropic of copying more than 20,000 works via torrents, meaning downloads over file-sharing networks. Specifically, they cite 20,517 works. According to the report, the publishers filed the complaint this year, alongside a 2024 case over the training of Claude on copyrighted material. They say they learned of the sheet-music book downloads through Bartz v. Anthropic, the 2025 class action brought by book authors, and so could not have included them earlier.

The defense, led by Sonal Mehta, argues that the second amended complaint does not explain how or why those 20,517 works were selected, and that it fails to move from "possible" to "plausible." It adds that Anthropic has built "enhanced guardrails" that prevent Claude from reproducing copyrighted lyrics, and that the guardrails on newer models contain no material from the pirate libraries. The publishers' attorney, Nicholas Hailey, countered that they do not have to prove every fact at this stage, nor know the internal details of the infringement. He also argued that Claude has memorized the works and can retrieve them when a user asks.

The judge was blunt about the figure: it is "curiously specific, but at the same time not," she said, noting that she cannot rely solely on "information and belief." Lee also partially dismissed, with leave to amend, the direct infringement claim against Dario Amodei, Anthropic's co-founder and CEO. Her reasoning: under that standard, "any CEO of a small startup would be liable." The secondary infringement claim, which Anthropic has not challenged, remains open. Lee is presiding over both cases, the 2024 one and the 2026 one, and has yet to rule on summary judgment in the first. In January 2025 she ordered the existing guardrails to be maintained and applied to future models, without broadly halting training on lyrics.

Our take. This is a procedural setback, not an acquittal. Having to sharpen a complaint's allegations is routine; it says nothing about whether the torrenting happened. And that is the underlying issue: discovery (the compelled production of documents) is where the truth would come out, and the publishers say they have no access to the details of how the infringement occurred. It is the classic dilemma of lawsuits against opaque labs: the accuser can't see the data, and the judge won't let the case in without concrete evidence.

The distinction the judge appears to be drawing between acquisition and output also matters. For industry context, the line left by the Bartz case (settled for $1.5 billion, according to the report, the largest copyright settlement in U.S. history) is that the serious risk lies less in training than in how the material was obtained. If the publishers cannot tie their works to specific downloads, their argument weakens; if they can, the Bartz precedent carries weight.

Short term: more litigation, higher legal costs and more uncertainty for every lab, which will have to document the provenance of their data better. Long term, and this is where our qualified optimism lies: this kind of friction is the path toward a clear licensing system that gives creators fair compensation and gives labs legal certainty. An AI capable of accelerating science and medicine should not be built on a permanent gray area. The outcome of this case, still open, will help define where the line is.