Anthropic Lyrics Case Resumes 30 Oct: Proving AI Market Harm

In Concord Music Group v. Anthropic (5:24-cv-03811-EKL) in the US District Court for the Northern District of California, the 21 October hearing ran for 4 hours 43 minutes and Judge Eumi K. Lee did not rule on fair use from the bench. The court has since set a continued hearing by Zoom for 30 October at 1 p.m. on the remaining Daubert motions, the disputes over whether each side's expert evidence can be admitted.

On 23 October the court made public the questions it had sent the parties before the hearing (ECF No. 918).

They show how the judge is approaching the case: she is inclined to analyse fair use separately for each use alleged in the complaint, of which there are four:

• use of the works in suit during training;
• Claude outputs that reproduce the publishers' lyrics verbatim;
• outputs that mash up the publishers' lyrics (what the publishers call derivatives);
• AI-generated new lyrics, which the parties call synthetic lyrics.

On the first factor she asks whether a secondary use must itself be a copyrightable work, and whether new expression must be copyrightable to be transformative. On the third factor she asks the parties to address Anthropic's position that there is nothing uniquely valuable about the publishers' lyrics relative to other training data.

The fourth-factor questions are the most pointed, and they target the publishers' market dilution theory. If Anthropic had to license the works, could Claude still generate an endless stream of competing lyrics? What makes synthetic output unfair, and does it extend to any lyrics in the same genre or style? How do the record's examples of dilution connect to the works in suit? And where is the line between harm from indirect substitution and harm from competition with a new work (citing Kadrey v. Meta)?

The 30 October session turns to the expert evidence behind those claims. Under Federal Rule of Evidence 702 the court must first check that an expert's method is reliable before admitting the opinion. The questions name five experts: Anthropic seeks to exclude Smith's opinions on harm to past and future licensing; the publishers seek to exclude Nagaraj's opinion on a market for all the data needed to train a general-purpose LLM; and the questions also cover Dennis, Davis and Stokes, on who created the sample sets, the fourth factor and music-theory analysis. The notice lists only the remaining Daubert motions and does not say which will be heard on the 30th.

The transcript of the 21 October hearing was filed on 26 October as ECF No. 922, but until 25 January 2027 it can only be viewed at the clerk's office or bought from the court reporter, and redaction requests are due by 16 November. Until it is public, the judge's words at the hearing cannot be quoted directly.

That is why the expert evidence matters. The judge's question about how the record's dilution examples connect to the works in suit means that usage figures or substitution examples cited by the publishers are not, by themselves, infringing outputs or proven economic loss; how much they prove depends on whether the expert methods tested on 30 October hold up.

J&C notes that the court is separating training from three kinds of output and asking the publishers to tie market dilution to the works in suit, so both sides' evidence must be specific to works and uses. AI developers and rights holders should organise training-data sources and licensing records, output monitoring and filtering logs, and data showing how particular works' markets are affected; until the court rules, both sides' positions remain claims.

Source: apps.cand.uscourts.gov | storage.courtlistener.com | www.courtlistener.com