Where the case stands
Publishers moved for partial summary judgment on 23 March 2026 (ECF 594). They ask the court to hold that Anthropic committed prima facie direct infringement of all 499 works, in training and in Claude’s outputs, and that none of it is fair use. Anthropic cross-moved in April 2026 (public version re-filed 22 April as ECF 693). It asks for fair use on training as a matter of law, judgment on the output claims because users and not Anthropic cause any copying, and judgment on the DMCA claim because no removal of copyright management information (CMI) from any work in suit has been shown.
Judge Lee heard argument on 21 October 2026. The minute entry records 4 hours 43 minutes in court. Argument on the remaining Daubert motions was continued to 30 October 2026 by videoconference. On 23 October the clerk published the questions the court had sent counsel before the hearing (ECF 918), after the parties told the court of public interest in them. Those questions are the best available signal of how the court is approaching the motions.
The court intends to split fair use into four uses
Citing Andy Warhol Foundation v. Goldsmith, the court told the parties it is inclined to run the fair-use analysis separately for each use alleged in the second amended complaint (ECF 728):
- Training. Use of the works in suit during model training.
- Verbatim outputs. Claude responses containing verbatim copies of Publishers’ lyrics. Anthropic’s brief says it does not assert fair use for verbatim copies.
- Mash-ups. Outputs that combine or rework Publishers’ lyrics, which Publishers call derivatives.
- Synthetic lyrics. AI-generated lyrics produced in response to prompts for new songs, which Publishers say rely on their lyrics.
The split matters because it keeps evidence about outputs from automatically deciding the training question, which is what Anthropic argued for and what Publishers resisted. Publishers’ position is that the training purpose cannot be separated from a model trained to return lyrics on request.
The questions also probed the outer edges of each side’s theory. On market dilution, the court asked whether Claude would still produce an endless supply of competing lyrics in a world where Anthropic had licensed the 499 works, whether the theory reaches any lyrics in the same genre or style, and how it fits with the way Publishers described the scope of relevant market effects in their motion to exclude Anthropic’s economist. On Anthropic’s side, it asked whether the public benefits of Claude count if lyrics are interchangeable with other training text, and how its claim that nothing is uniquely valuable about Publishers’ lyrics should be weighed under factor three.
The output record, and who wrote the prompts
Publishers say their investigation drew verbatim or near-verbatim lyrics for each of the 499 works in response to direct requests, plus copying in response to dozens of indirect prompts. Their linguistics expert, Prof. Robert Leonard, compared those outputs against the lyrics and found substantially verbatim overlap. Publishers argue that copies made for an investigator are actionable infringement.
Anthropic’s answer is about authorship of the prompts. In a discovery sample of about 5 million prompts and outputs from a six-month period spanning the filing of the complaint, it says more than 83% of the prompts that produced lyrics to works in suit came from Publishers or their agents, many of them attempts to get around guardrails. It says Publishers identified third-party outputs for only 82 works, submitted close to 10,000 prompts of their own, and withheld the material that would let Anthropic check which outputs were theirs on work-product grounds. One output first described as an undisputed third-party example was later conceded to come from Publishers’ own expert. Anthropic describes outputs that reproduce portions of the lyrics as rare, about 0.01%. It says that in the sample Publishers’ expert analyzed, only two outputs generated after Anthropic strengthened its guardrails related to works in suit, and both were fragments.
Prompt authorship feeds two separate arguments. Rightsholders cannot infringe their own works, so outputs elicited by Publishers’ agents cannot support secondary liability. And under the volitional-conduct rule (VHT v. Zillow in the Ninth Circuit, and the Second Circuit’s Cartoon Network), Anthropic argues the person who prompts the model, not the company that built it, is the direct cause of any copy.
What text counts as the work
Anthropic argues that Publishers cannot win the copying element on summary judgment because they produced deposit copies for only 13 of the 499 works and compared Claude’s outputs to lyrics from the aggregator LyricFind instead. It points to discrepancies between LyricFind text and the deposit copies Publishers did produce, including a stanza that appears in one version of a song and not the other. The court asked directly what significance the missing deposit copies have for the verbatim-reproduction claim.
Five experts under challenge
The court directed argument on each Daubert motion to a specific issue. The briefs confirm Smith and Nagaraj; the offering party for the other three is inferred from how the court’s questions are addressed.
| Expert | Offered by | Subject | What the court asked |
|---|---|---|---|
| Smith | Publishers | Harm to past and future licensing | Methodology and bases for the licensing-harm opinions |
| Nagaraj | Anthropic | Whether a market to license training data exists or can form; public benefits | Authority for an expert to define the relevant market as one that cannot exist; whether his but-for world wrongly assumes Claude would be built the same way; where his reports tie benefits to Claude rather than AI generally |
| Dennis | Publishers | Review of sample sets of Claude records, Anthropic employee prompting, absence of CMI | How the court can assess reliability when Publishers’ counsel built the sample sets and the expert does not know counsel’s method, and how Anthropic could fairly cross-examine at trial |
| Davis | Anthropic | Fair-use factor four; Publishers’ financial records | What the testimony adds beyond fact witnesses, and what method supports opinions on how generative AI may help the lyric-licensing market |
| Stokes | Anthropic | Music theory applied to lyrics; benefits of AI | Whether music-theory analysis of the works in suit is relevant even if general opinions are not; where the report connects AI benefits to Claude |
Anthropic’s brief also certifies, under the court’s standing order on generative AI, that two of its experts used generative AI tools in preparing their reports and kept records of every prompt they submitted.
Claims that have narrowed
The evening before Anthropic’s brief was due, Publishers said they would not pursue their contributory and vicarious infringement claims. They filed a notice of voluntary dismissal (ECF 685) the night Anthropic filed. Anthropic asks for judgment on those claims, not just dismissal, so they cannot be raised again in Concord II or a later case. It relies on the Supreme Court’s 2026 decision in Cox Communications v. Sony Music Entertainment, which it reads as confining contributory liability to active encouragement of infringement or a service tailored to it.
The DMCA claim survived the pleadings in October 2025 on the allegation that Anthropic chose the Newspaper text-extraction tool over alternatives because it removed more CMI. At summary judgment Anthropic says discovery produced no evidence that Newspaper removed CMI from any work in suit, and that Publishers’ own experts recognized the tool works on page headers and footers, not the body text where lyrics and their credits appear.
Evidence notes
IPLYR’s reading of what this record means for anyone building model-side evidence. These are observations about method, not predictions about the ruling.
- The sampling frame is now a gatekeeping issue.
- The Dennis question goes to the foundation of any output study. If counsel chose the records and the expert cannot describe the selection rule, the court is asking whether the analysis can be tested at all. The selection rule, who applied it and when should be fixed and written down before anyone looks at results.
- Record who wrote each prompt.
- Investigator and adversarial prompts can show what a model is capable of producing. They do not show what ordinary users received, and here they are the basis of Anthropic’s volition and secondary-liability defenses. Capability and propensity results belong in separate, labeled arms with prompt authorship recorded for each.
- Measure against the registered text.
- A match rate computed against an aggregator transcription invites the deposit-copy objection. Where the deposit copy is available it should be the reference. Where it is not, the reference used and any known differences should be stated in the report.
- Log any AI assistance in the expert’s own work.
- In this case Anthropic certified that its experts kept records of every prompt they gave a generative AI tool. A report that cannot produce those records hands the other side an easy line of cross-examination.
Related matters
- Concord Music Group, Inc. v. Anthropic PBC (“Concord II”), No. 5:26-cv-00880-EKL (N.D. Cal., filed 28 Jan 2026), a separate suit by the same publishers focused on torrenting allegations.
- ABG EPE IP LLC v. Anthropic PBC, No. 5:26-cv-12473 (N.D. Cal., filed 23 Oct 2026), which leans on this record for its output allegations.
- BMG Rights Management (US) LLC v. Anthropic PBC, No. 5:26-cv-02334 (N.D. Cal.).
Documents
Filings link to the public RECAP copies on CourtListener except ECF 918, which links to the copy hosted by ChatGPT Is Eating the World. Several briefs were later re-filed under the court’s omnibus sealing order with new ECF numbers; the court’s questions cite those re-filed versions.