Matthew J. Oppenheim of Oppenheim & Zebrak LLP, who represents the publishers, informed U.S. Magistrate Judge Susan van Keulen of the issue over Anthropic data scientist Olivia Chen's declaration at the start of a two-hour virtual hearing on the latest discovery disputes in hotly contested copyright infringement litigation against Anthropic, a private company recently valued at $61.5 billion and backed by Amazon.com Inc., Google LLC and Lightspeed Venture Partners.
The music publishers, which include Universal Music Publishing Group, Concord Music Group Inc. and ABKCO Music Inc., contend in their October 2023 lawsuit that Anthropic's large language model, AI-chatbot Claude rips off their intellectual property rights by reproducing lyrics to thousands of copyrighted songs on command.
Anthropic has denied the allegations and counters that it has put in place measures that prevent this type of verbatim copying. The litigation has proceeded into discovery during which numerous fights have arisen, including the latest dispute over sampling Claude's output.
Anthropic submitted Chen's declaration to support its argument that it should only have to produce a sample size of 1 million Claude prompts and outputs — even if such a sample comes with a 25% relative margin of error — instead of the 24.4 million sample size sought by the publishers. Their expert claims the sample size would have a 5% relative margin of error.
But before the hearing on the dispute began, Oppenheim told the judge that plaintiffs' counsel discovered in the past day that at least one citation in Chen's declaration was fake.
The citation at issue supports Chen's statement that a 25% margin of error rate is "extensively supported by peer-reviewed research," and it points the court to a purported study titled, "A Statistician's Guide to Making Sound Inferences from Noisy Data," claiming it was written by Julien Dutant and Julia Staffel and published in the journal The American Statistician.
Oppenheim said the journal cited is real, and the purported authors are real academics, but Dutant is an epistemology lecturer at King's College London, and he confirmed in writing that he had never worked with Staffel, who is a philosophy professor at the University of Colorado, Boulder, and he has never published a study under that name. The journal also confirmed the study doesn't exist, Oppenheim said.
"This article is fictitious," he said. "It does not exist. It was never written. It was never published. This is not an error or a mis-citation. It is a complete fabrication."
Oppenheim added that there are two other citations in Chen's declaration that are also inaccurate, although not completely fictitious, and while the plaintiffs' lawyers do not think Chen "intentionally fabricated fake authority," the errors show she likely used Claude or another AI tool that hallucinates and creates fake citations to draft the declaration.
In light of the purported errors, Oppenheim asked the court to strike Chen's declaration in its entirety and to consider sanctions against Anthropic's lawyers at Latham & Watkins LLP as well as the broader impact of the expert's conduct.
To support his sanctions requests, Oppenheim cited a lawyer's ethical duty to check all legal filings for accuracy, and he noted that counsel alone bears all responsibility for legal filings. He also pointed to a Minnesota federal judge's January ruling in Christopher Kohls et al. v. Keith Ellison et al
., which threw out an erroneous expert declaration prepared by a Stanford University AI expert for citing fake, AI-generated sources in his declaration.Oppenheim additionally asked the judge to examine Chen during the hearing on her use of AI to draft the declaration before the defense's attorneys have a chance to talk to her, so she gives the court "candid, non-rehearsed" responses. But Judge van Keulen refused to interrogate Chen on her use of AI for now.
Sarang V. Damle and Ivana Dukanovic of Latham & Watkins, who represent Anthropic, appeared genuinely surprised by Oppenheim's accusations, and Damle asked the judge to give the defense's legal team time to investigate the matter.
"I wish Mr. Oppenheim had brought this to our attention before sandbagging us in this court," Damle said.
Damle quickly pulled up Chen's declaration and said it appears that the study might be legitimate, but the names of the authors were wrong, in which case the mistake would be merely a citation error.
Judge van Keulen ordered Anthropic to file a response to Oppenheim's allegations by May 15. She also ordered that Anthropic, Chen and Anthropic's attorneys to preserve "any and all documents and discussions" relating to the development of her declaration. The judge said after Anthropic responds to Oppenheim's allegations in writing, she will decide whether Chen's declaration should be stricken, and the music publishers can file a motion for sanctions if they want to.
The judge told the attorneys that despite the claims regarding Chen's declaration, she would still proceed with hearing arguments on the discovery disputes at hand. Even so, she noted that "no one should take away that this does not present a very serious and grave issue."
"There is a world of difference between a mis-citation and a hallucinated citation generated by AI, and everybody on this call knows that," the judge told the attorneys. "I look forward to [Anthropic's] explanation. I may strike this declaration in whole or in part."
Aside from the issue over Chen's declaration, the judge heard arguments on the sample sizes, and she didn't rule but did indicate she will likely reject both sides' sample size proposals for being too expansive or too narrow. Judge van Keulen also observed that Anthropic's 25% margin of error rate backed by Chen is likely "too large."
The judge also heard arguments on a dispute over Anthropic's discovery demands for information on the plaintiffs' Claude prompts. But the music publishers' lawyers argued Anthropic's requests are overbroad and unreasonable, and the information is subject to attorney work-product protections in light of a ruling in a similar discovery dispute in Tremblay v. OpenAI Inc.
Judge van Keulen agreed that Anthropic's requests are "quite vast and raise a number of concerns," but she didn't rule from the bench, and she observed that the music publisher's case isn't necessarily "on all fours" with the Tremblay case, although she didn't elaborate on the distinctions between the two IP cases.
Before the hearing wrapped, the judge said she was "dismayed" by Anthropic's efforts to designate so many documents as confidential and added that she's "concerned and frustrated" by the "over-designation" of confidentiality in the case.
"This is not a case where you can wave your wand and everything is confidential," she said.
But the judge added that the parties should have been able to work out their ongoing dispute over certain confidentiality designations before bringing their disagreements to her.
"It seems obvious that you could work this out," she said. "I expect more from each side in that regard."
The publishers are represented by Matthew J. Oppenheim and Nicholas C. Hailey of Oppenheim & Zebrak LLP.
Anthropic is represented by Sarang V. Damle and Ivana Dukanovic of Latham & Watkins LLP.
The case is Concord Music Group Inc. et al. v. Anthropic PBC, case number 5:24-cv-03811, in U.S. District Court for the Northern District of California.
--Editing by Kristen Becker.
For a reprint of this article, please contact reprints@law360.com.
