A federal judge in San Jose has dismissed the direct copyright claim against Anthropic chief executive Dario Amodei in the music publishers' case over Claude's training data, while leaving the far larger claims against the company itself undecided. Judge Eumi Lee also questioned whether the publishers had pleaded their central allegation — that Claude copied more than 20,000 songs from pirated sources — with enough detail, and ordered more briefing before ruling on the rest.
What the judge decided
The ruling came at a hearing on November 4 in Concord Music Group's second case against Anthropic, naming the company, Amodei and Benjamin Mann as defendants. Lee dismissed the direct infringement claim against Amodei with leave to amend, so the publishers can try to repair it. Amodei had sought dismissal with prejudice, which would have ended the claim.
The distinction matters. Direct infringement requires unlawful copying by Amodei himself, while the complaint attributes the downloads to Mann and other Anthropic employees. Approving, directing or knowing about that conduct describes contributory liability, not direct copying, and Amodei's motion did not challenge the contributory claim. That claim survives untouched.
The publishers' lawyer argued the point the other way. "Just because Amodei didn't click download on his own computer doesn't mean he isn't directly liable," Nicholas Hailey told the court. Lee was not convinced. "Every CEO for any small startup would be liable under the analysis here," she said.
The 20,517 problem
The bigger question is whether the publishers' case against Anthropic itself is pleaded with enough specificity. Lee had tentatively ruled from the bench to deny Anthropic's motion to dismiss, then changed course after hearing the company's lead counsel, Sonal Mehta.
The publishers' second amended complaint, filed on May 21, lists exactly 20,517 compositions and alleges Anthropic copied lyrics without permission for training, then generated verbatim and near-verbatim copies in Claude's outputs. Lee wanted to know where the number came from.
"It's oddly specific, yet oddly not because of the basis," she said, noting she could not proceed on "information and belief" alone. Mehta argued the complaint never explains how or why those 20,517 works were chosen, and that Anthropic's newer guardrails keep pirated-library material out of Claude's outputs.
Hailey countered that Claude memorized the material and can recall it when prompted, and that the publishers need not prove facts at the pleading stage. The records of users' prompts and Claude's corresponding outputs sit in Anthropic's possession, he said, not the publishers'. The judge asked for more briefing instead of deciding.
What stays open
Anthropic's separate partial motion remains unresolved. It seeks dismissal of the part of the direct-infringement claim based on Claude's outputs and the entire claim that Anthropic stripped copyright-management information from its training text. The company did not move against the torrenting-based direct claim or the training portion of the case.
The copyright-information dispute turns on a fine distinction. Anthropic argues the allegations concern lyrics rather than complete musical compositions, while the publishers say scanning and torrenting songbooks copied complete works and that extraction algorithms removed identifying information during training. Lee ordered the parties to meet and confer and file a joint status report by November 11.
Our Read
The dismissal is narrow, but the skepticism behind it is the story. Amodei's personal exposure is trimmed for now, and the publishers can amend, so nothing about the chief executive is settled. What matters more is the judge's doubt about the 20,517 figure, because that number is the spine of the publishers' case against the company.
If the publishers cannot plead specificity, the case narrows to whatever they can actually show — a much smaller fight than the one they filed. The guardrails argument cuts both ways: Anthropic says newer models exclude the pirated material, while the publishers say memorization makes the guardrails beside the point.
The ruling that will matter most is still to come. The fair-use question in the publishers' original 2024 case was argued on October 21 and Lee has not yet ruled on it. Fair use, not the CEO's personal liability, is the question that decides whether training on lyrics was lawful at all.
Why It Matters
Training-data copyright is the defining legal question for the AI industry, and this hearing shows courts testing the pleadings before they test the law. Requiring publishers to plead piracy claims with specificity raises the bar for the next wave of cases, while a CEO escaping personal liability narrows who can be sued for a company's training decisions.
The case also sits inside a wider pattern. Copyright lawsuits now touch nearly every frontier lab, and courts are drawing lines case by case: what counts as copying, what guardrails prove, and who is personally responsible. Anthropic's answers in this case will be read by every lab facing the same questions.
Background
The publishers — Concord, Capitol CMG, Universal and ABKCO — filed the second case this year after learning from Bartz v. Anthropic, the 2025 author class action, that Anthropic had used pirated materials to download full songbooks. Bartz settled for $1.5 billion. In January 2025, Lee ordered Anthropic to maintain its copyright guardrails for Claude.
Source: Courthouse News Service report and Unite.AI's account of the minute entry
0 Comments