What happened
According to court filings, music publishers are suing Anthropic in the US District Court for the Northern District of California, alleging that its Claude models reproduce copyrighted song lyrics. In spring 2025 the two sides disagreed over how large a sample of Claude prompts and outputs Anthropic should hand over, and each brought in a statistics expert to make its case. Anthropic’s expert was one of its own data scientists.
The problem surfaced at a hearing on 13 May 2025. The publishers’ counsel told the court that a citation in the expert’s declaration seemed to point to an article that did not exist, by authors who had never worked together. The court gave Anthropic time to look into it.
Two days later, a Latham & Watkins associate acting for Anthropic filed a declaration explaining what had gone wrong. According to that declaration, a lawyer at the firm had found a genuine article in The American Statistician through a Google search and identified it as further support for the expert’s opinion. The associate then asked Claude.ai to turn it into a properly formatted legal citation, giving it the link to the correct article.
What came back, the declaration says, had the right journal, year and link, but the wrong title and the wrong authors. The declaration says the same formatting step introduced two more errors: one source was given the wrong title, and another was credited to “Lower Windward Environmental LLC” rather than “Windward Environmental LLC”.
The firm said the article was real and did support the expert’s opinion. It described the mistake as honest and embarrassing, apologised, and said it had since added multiple levels of review.
The magistrate judge ruled on 23 May 2025. The court found the issue serious, though not as grave as it had first appeared, because a person had found the real article and linked to it correctly. It described the invented title and authors as “a plain and simple AI hallucination”.
What an auditable version would have shown
According to the firm’s declaration, there was a check meant to catch this kind of error. The team confirmed that the article supported what the expert had said and corrected the volume and page numbers. The declaration says the team clicked the link and read the article, yet still did not notice that the title and authors were wrong.
The court said it was not clear how a person doing a manual citation check could have missed an error that changed the article’s title completely.
The declaration describes a check of the volume and page numbers. It does not describe any check of the title or the authors.
The record that would have helped is a simple one: which parts of the citation came from the original source, which came from Claude.ai, and which a person had checked against the article.
Where the gap was
The court’s order points to a written rule. A standing order of the district judge presiding over the case requires a certification that lead trial counsel has personally verified the accuracy of the content. The court found that neither the certification nor the verification had happened.
The court struck paragraph 9 of the expert’s declaration, which relied on the citation, and said the episode undermined the overall credibility of the written declaration, a factor in its conclusion. On the sampling question itself, Anthropic’s expert had proposed a 25% margin of error and the publishers’ expert 5%. The court settled on about 11.3% and ordered Anthropic to produce a sample of 5 million prompt-output pairs. The order did not impose a monetary sanction.
A VerificationGate is designed to check an output against a trusted reference before it is used. In this case that reference was already to hand: the article the citation linked to.
A ConductRecord is designed to keep what the tool produced, what a person checked and what was changed, so a reviewer can see which parts of a filing were actually verified.
VerificationGate and ConductRecord are Headlights designs.
What governance should have looked like
According to the firm’s declaration, additional procedures had been introduced, including multiple levels of review, to prevent the same mistake from happening again.
When an AI tool is used only to format a citation, best practice would be to treat everything it returns as unverified, including details nobody asked it to change. Before filing, each element, from the title and authors to the journal, volume and pages, should be checked against the original source. And where a court requires counsel to certify that the content is accurate, that certification should be given, and it should be true.
Failure Pattern: an AI tool formatted a citation but returned an invented title and the wrong authors. The firm’s reviewers opened the correct article and corrected some details, yet the title and author errors went unnoticed.
Governance Principle: when an AI tool handles a citation, every element it returns should be checked against the original source, and the filing should record who verified it.
The reference implementation of VerificationGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Order on joint discovery submissions, Concord Music Group, Inc. v. Anthropic PBC, No. 5:24-cv-03811 (N.D. Cal., 23 May 2025), Dkt. 377
- Declaration of counsel related to ECF No. 365, Concord Music Group, Inc. v. Anthropic PBC (N.D. Cal., 15 May 2025), Dkt. 371