Judge Questions Claim That Claude Reproduced 20,517 Lyrics

SAN JOSE, Calif., Nov. 6 — TheThe music publishers suing Anthropic say its Claude chatbot has reproduced the lyrics to all 20,517 songs in their suit, but on Friday the judge hearing the case wrote that their complaint does not give one example. “Yet Publishers have not alleged a single instance of any of the 20,517 works at issue in this case being reproduced as a Claude output,” Judge Eumi K. Lee wrote in a three-page order.

Judge Lee has not ruled on Anthropic’s motion to dismiss that claim, and the order imposes no penalty. It tells the two sides to keep conferring and moves their status report from Wednesday, Nov. 11, to Friday, Nov. 13, “in light of this guidance and the Veterans Day holiday.”

At issue is the publishers’ claim, in Count II, that Anthropic directly infringed their copyrights through Claude’s outputs: what it writes in answer to users. Their claims over training are not in question; Anthropic’s motion does not challenge them, the publishers noted. The two sides are arguing over a question that has come up in other suits against AI companies: whether a complaint must cite an actual infringing output.

Anthropic’s motion, filed Aug. 3, argued that the output allegations “are conclusory and fail to provide even a single example of an allegedly infringing output.”

The publishers argue that they need not cite specific outputs at this stage, and that the records of what Claude told its users are in Anthropic’s hands alone. “Publishers obviously cannot cite or quote Claude output copying their lyrics in response to third-party user prompts,” they wrote in their opposition, filed Aug. 31, “because those records are solely in Anthropic’s possession and control.” They cited rulings in suits against Suno and Stability AI, and wrote that discovery in the first case “disproved Anthropic’s protests that Claude did not generate infringing outputs to third parties.”

Yet Publishers have not alleged a single instance of any of the 20,517 works at issue in this case being reproduced as a Claude output.

The Concern

To avoid dismissal, Judge Lee wrote, the publishers gave “the clear impression” that they had a factual basis for alleging that all 20,517 works allegedly used to train Claude were also reproduced in its outputs. “The Court assumed that these assertions had some basis,” she wrote, citing Rule 11(b)(3) and (4) of the federal rules.Rule 11Under Rule 11(b)(3) of the Federal Rules of Civil Procedure, a lawyer who presents a filing certifies that “the factual contentions have evidentiary support” or, if so identified, will likely have it after further investigation or discovery. The rule also provides for sanctions; the order does not mention them. “Counsel’s statements at the motion hearing cast the complaint in a different light.”

Now, she wrote, the publishers appeared to ask the court to infer that all 20,517 works were infringed through outputs “merely by virtue of their alleged inclusion in the training process.” They “do not even allege that users prompted Claude for the lyrics to any of the 20,517 works at issue.”

At Wednesday’s hearing, the judge had tentatively ruled to deny Anthropic’s motion, then asked for more briefing after hearing its lawyer, Sonal Mehta, Courthouse News reported; Nicholas Hailey, for the publishers, argued that Claude has memorized the material and can recall it when prompted by a user.

The publishers invoked “information and belief” at the hearing, but the complaint does not expressly make its output allegations that way, Judge Lee wrote, and in the related earlier case the publishers “were able to identify specific acts of alleged output-based infringement.”The First SuitThe same publishing groups sued Anthropic in October 2023 over 499 works, and that complaint included examples of outputs, Anthropic’s motion says. The court heard both sides’ motions for summary judgment in that case on Oct. 21, Anthropic’s registration statement says. If their claims are right, the publishers, like any Claude user, could use Claude to generate infringing outputs themselves, she reasoned: “Thus, it seems implausible that the information at issue is solely in Anthropic’s possession.” The publishers had argued in their opposition that requiring examples from their own tests would force disclosure of protected legal-investigation material.

“In sum, the Court is concerned about Publishers’ representations in the second amended complaint and the opposition to Anthropic’s motion to dismiss,” she wrote.

Anthropic, which is preparing to sell shares to the public, says in its registration statement that the music publishers’ suits “are in preliminary stages” and that it “disputes the allegations of wrongdoing and intends to defend itself vigorously in these matters.” It lists this case with suits by other publishers, including BMG, Sony Music Publishing and Warner Chappell, which together assert claims over both training and “the outputs of those models.”

The publishers need not amend their complaint until Anthropic’s motion is decided, the order says, despite Wednesday’s ruling against them on Dario Amodei, Anthropic’s chief executive. That day Judge Lee dismissed their claim that he directly infringed by torrenting, with leave to amend. The publishers, among them Concord, Universal Music Publishing Group and ABKCO, sued in January over more than 20,000 songs, “with potential statutory damages of more than $3 billion,” they said in a statement to Music Business Worldwide.Per WorkCopyright law lets an owner elect statutory damages of $750 to $30,000 for each work infringed, and up to $150,000 if the court finds the infringement willful. At that maximum, 20,517 works would come to about $3.08 billion, by this paper’s arithmetic.

Daily Cortex is written with AI models, including Anthropic’s Claude, and fact-checked with OpenAI’s GPT models.