Concord Music Group, Inc. v. Anthropic PBC
- Van Keulen
- 5:24-cv-03811
- U.S. District Court · Northern District of California
- 8
In Concord Music Group v. Anthropic, Judge Van Keulen denied Anthropic’s discovery request, ordered a five-million-record sample, and partly changed confidentiality designations.
Anthropic must produce the ordered five-million-record sample and revise some confidentiality designations. The Publishers’ request for undisclosed prompts, outputs, and settings was denied without prejudice, and their challenge to Anthropic’s designations was sustained in part and overruled in part.
What happened
Concord Music Group, Inc. v. Anthropic PBC involved three discovery disputes in the parties’ copyright litigation. Anthropic sought prompts, outputs, and settings that the Publishers had not disclosed from their investigation; the parties disputed how to sample Claude records; and the Publishers challenged Anthropic’s confidentiality labels.
The court denied Anthropic’s request without prejudice because the requested material was protected attorney work product and the requested waiver was broader than necessary. The court ordered Anthropic to produce five million randomly selected prompt-output pairs, split evenly between specified periods before and after the lawsuit began. It also allowed some confidentiality labels to remain but required Anthropic to reduce others.
Judge Susan Van Keulen partly sustained and partly overruled the Publishers’ challenge to Anthropic’s confidentiality designations, and struck paragraph 9 of an expert’s declaration because it contained an inaccurate citation generated with artificial intelligence.
The detailed version
- Concord Music Group, Inc. v. Anthropic PBC · No. 5:24-cv-03811
- Van Keulen
- May 23, 2025
Background
The court addressed three joint discovery submissions in the parties’ case. Anthropic asked the Publishers to produce undisclosed prompts submitted to Claude, Claude’s corresponding outputs, and related account or system settings from the Publishers’ pre-lawsuit investigation. The Publishers said they had already produced all prompts and outputs on which they relied, totaling approximately 4,659 records or nearly 5,000 prompt-output pairs.
The second dispute concerned the sampling protocol for the Publishers’ requests for Claude prompts and outputs related to song lyrics. The court had previously ordered a statistically significant sample. The parties proposed different margins of error and submitted expert declarations.
The third dispute concerned the Publishers’ challenge to Anthropic’s use of the “Highly Confidential—Attorneys’ Eyes Only” designation for three categories of material: 9,418 Claude prompt-output records, Claude use statistics and figures, and two training datasets.
Anthropic’s request for undisclosed prompts and outputs
The court treated the prompts, outputs, and settings created during the Publishers’ legal investigation as attorney work product. Work product is material prepared for litigation, including an attorney’s mental impressions, opinions, and legal theories. The court agreed that the Publishers had waived that protection to a limited extent by relying on certain prompts and outputs in their complaint and other filings and by producing the records on which they relied.
The court nevertheless held that the waiver had to be limited to what was necessary to address any unfair advantage. Anthropic’s requests sought all prompts and outputs, including those not relied upon by the Publishers, and requested related account information, system prompts, temperature settings, and dates and times. The court found those requests overbroad because it remained unclear what evidence the Publishers would use to prove their allegations about the ease or scale of allegedly infringing outputs.
The court therefore DENIED without prejudice Anthropic’s request to compel production of the Publishers’ undisclosed prompts, outputs, and settings. The court left Anthropic free to seek specific facts supporting specific contentions or opinions disclosed during the litigation, as necessary.
Sampling protocol
The court adopted a 95% confidence level, a corresponding statistical score of 1.96, and an expected prevalence of 0.00006. It selected an approximately 11.3% margin of error as a balance between the burden of production and the needs of the case.
The court ORDERED Anthropic to produce five million prompt-output pairs. The sample must contain 2.5 million randomly selected records from September 22 through October 18, 2023, and 2.5 million randomly selected records from October 19, 2023, through March 22, 2024. Anthropic was required to produce the sample as soon as practicable and no later than July 14, 2025, the then-current deadline for substantial completion of remaining document production.
The court also addressed an issue involving Anthropic expert Ms. Chen’s declaration. Claude.ai had been used to format citations, and it generated a fictitious article title and inaccurate authors for one citation. The court found that this was an artificial-intelligence-generated error and STRUCK-IN-PART the declaration by striking paragraph 9. The court stated that the issue undermined the overall credibility of the written declaration and affected its conclusion on the sampling dispute.
Confidentiality designations
The court held that a blanket “Highly Confidential—Attorneys’ Eyes Only” designation for Claude prompts and outputs was not appropriate. Anthropic could initially designate the prompts and outputs as “Confidential,” recognizing the privacy interests of Claude users and the practical difficulty of reviewing and redacting millions of records individually. The court cautioned that the standard for discovery confidentiality was separate from the standards governing whether material could be filed under seal.
For Claude use statistics, the court found Anthropic’s designation overbroad in part. Detailed query counts could remain “Highly Confidential—Attorneys’ Eyes Only,” but Anthropic had to reduce other statistics, including less detailed and less comprehensive use counts, to “Confidential.”
For the two training datasets, the court agreed that the datasets were generally competitively sensitive and could remain “Highly Confidential—Attorneys’ Eyes Only,” unless Anthropic had publicly disclosed that it used them to train Claude.
The court therefore SUSTAINS-IN-PART and OVERRULES-IN-PART the Publishers’ challenge to Anthropic’s confidentiality designations.
Read the full 8-page opinion on CourtListener, the free public archive maintained by the Free Law Project.