The music publishers suing Anthropic say its Claude chatbot reproduced lyrics from all 20,517 songs in their case. On 6 November the judge pointed out that their complaint does not show one such output.
“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,” U.S. District Judge Eumi K. Lee wrote in a three-page order in Concord Music Group v. Anthropic.
The judge has not yet ruled on Anthropic’s motion to dismiss that claim. But the order shows what a court may expect before an output claim goes further: an actual output.
What the order says
The dispute is over Count II, the claim that Anthropic infringed through what Claude writes in answer to users. Anthropic argued the claim fails “to provide even a single example of an allegedly infringing output,” the order notes.
In their opposition brief, the publishers wrote that “Anthropic generated AI output copying the lyrics to their 20,517 works verbatim and near-verbatim.” The judge assumed that had some basis. At the hearing, she wrote, counsel instead seemed to ask the court to infer output infringement “merely by virtue of their alleged inclusion in the training process.”
She also doubted the publishers could rely on “information and belief,” a way of pleading facts a party cannot yet confirm, typically because the other side holds them. The publishers, “like a third-party Claude user,” could prompt Claude themselves. So “it seems implausible that the information at issue is solely in Anthropic’s possession.”
The parties must keep talking and file a status report by 13 November 2026. The publishers need not amend their complaint until the court rules on Anthropic’s motion.
The hearing
On 4 November the court granted a partial motion to dismiss by chief executive Dario Amodei, with leave to amend, according to the case docket. Courthouse News reported it covered the direct infringement claim against him, and that Judge Lee said of the publishers’ theory: “Every CEO for any small startup would be liable under the analysis here.”
Follow the money
Under 17 U.S.C. § 504(c), an owner may elect statutory damages of “not less than $750 or more than $30,000 as the court considers just” per work. For willful infringement the cap rises to “not more than $150,000.”
By our arithmetic, 20,517 works span about $15.4 million at the floor to about $3.08 billion at the willful ceiling. But the statute sets one award per work “for all infringements involved in the action.” Our reading: the output claim is a second route to liability for the same songs, not a second bill. The court has not addressed damages.
Some Claude prompts are already preserved
In July a magistrate judge ordered Anthropic to preserve a random sample of 500 million prompt and output records. It must also keep “all records that contain the terms ‘song!,’ ‘lyric!,’ ‘verse!,’ or ‘chorus!'” from the same periods, which run to 24 March 2026. The order does not say which Claude products the records come from.
For a small team or agency, this cuts both ways. A claim about AI output rests on the output, so your own record of what your tools produced is evidence too. Content owners now contest both ends of the pipeline: collection, as with Meta’s paused scraping terms, and output, which in the EU carries a marking duty with a 2 December deadline for older tools.
What to do this week
Do not ask a chatbot for full song lyrics or other published text for client work. Check drafts for long verbatim passages before they ship, and keep a dated log of prompts and outputs for paid content. Then watch the 13 November status report to see whether the publishers add real output examples.
Sources
Get the weekly email
The week's most useful AI news, a tutorial and an interview.
Martin Scott
Martin Scott is the editor for business and policy at AI Observer, covering startups and deals, regulation, health and science, and AI agency profiles. He follows the money and the legal text, and separates what a rule says from what people claim it says. His style is analytical and plain-spoken.
Spotted a mistake? Tell us on the corrections page.