Judge Inclined to Weigh Claude’s Training and Outputs Separately

SAN JOSE, Calif., Oct. 23 — JudgeJudge Eumi K. Lee, weighing music publishers’ copyright suit against Anthropic, told both sides she is “inclined to analyze the fair use defense separately as to each ‘use’”: training Claude on the publishers’ lyrics, and Claude’s outputs of verbatim lyrics, of “mash-ups” and of AI-generated lyrics. Her questions were made public on Friday.

In Manhattan, authors suing OpenAI asked for a split of their own in a brief made public Saturday: OpenAI’s downloads of books from LibGen and Books3, they argue, are a use separate from training, to be judged on its own. “Put simply, writing a book review does not mean you can steal the book,” they wrote.

How the uses are divided matters because statutory copyright damages are counted by the work, and a company that wins on one use can still face damages on another. In 2025, in Bartz v. Anthropic, Judge William Alsup ruled that Anthropic’s training on books was fair use but ordered a trial over the pirated copies in its library.Statutory DamagesA copyright owner may take statutory damages instead of proving its losses: $750 to $30,000 for each work, which a court may raise to as much as $150,000 if the infringement was willful. The case settled; in July the court approved a $1.5 billion fund, an estimated $3,000 a work before costs and fees.

The claims keep coming: on Oct. 21 Reservoir Media Management sued Anthropic in the same court over 853 musical compositions, seeking $127.95 million in statutory copyright damages, the $150,000 maximum for each, and about $151.4 million with its Digital Millennium Copyright Act claims. The complaint says Claude, asked recently to recreate songs “for a modern audience,” did so and “admitted” that the results were “edging past ‘inspired by’ into reproducing the copyrighted song.” An Anthropic spokesperson told Billboard: “We disagree with the allegations and intend to defend this case vigorously.” The same day Reservoir sued Suno, which makes an AI music generator, over 1,050 compositions; Suno declined to comment to Billboard. Reservoir’s lawyer, Richard S. Busch, also represents Round Hill Music, whose suit against Anthropic is before Judge Lee.

The Lyrics

Judge Lee sent the questions before Wednesday’s hearing “to focus argument”; the clerk filed them on the public docket, citing “public interest in these questions.” They show what she is weighing, not how she will rule. On the first fair-use factor, the purpose and character of the use, she asked: “For the fair use defense to apply, must the secondary use be a copyrightable work?” The publishers argue that AI generates “uncopyrightable digital facsimiles of human authorship.”

Six questions test the publishers’ market-dilution theory. If Anthropic had to license the publishers’ songs, she asked, would Claude “still have the ability to generate an endless stream of AI-generated lyrics that compete with Publishers, and also with artists unaffiliated with Publishers?” And if dilution does not count, she asked, may the court still weigh “the potential displacement of human creators”?

Anthropic answered the dilution theory in its April motion: “The fair use inquiry is anchored on ‘the problem of substitution’—not competition.” The judge also asked about the public benefits Anthropic claims: “If Claude was trained on lyrics and thus has certain capabilities, but could have been trained on other data, are the purported public benefits of Claude attributable to training on lyrics?”

Anthropic has also argued that whether Claude’s outputs are fair use “is a fact-heavy question best left for the jury.” The judge asked what questions a jury should get, and how a trial would proceed “given the number of Works in Suit and the many acts of alleged infringement.”

She pressed both sides on their experts.DaubertDaubert motions ask a judge to exclude an expert’s testimony. Judge Lee’s questions describe the court’s “gatekeeping responsibility to assess the reliability” of an expert’s methodology. She asked Anthropic, about its expert Nagaraj: “According to what legal authority may an expert opine that the ‘correct’ market is an impossible one?” She asked the publishers how the court could assess their expert Dennis’s method if “Publishers’ counsel created the sample sets that Dennis evaluated.” The remaining expert motions were put over to Oct. 30 at 1 p.m. Pacific, by Zoom.

The Books

The authors’ brief, opposing OpenAI’s motion for summary judgment on fair use, divides OpenAI’s conduct into three uses. First, it says, OpenAI downloaded the books “from pirate sites,” stored them “in an internal repository” and made them available to “all full-time employees and contractors.” Second came training. Third, it says, OpenAI distributed the books, to recipients the public version blacks out: “a use that OpenAI does not even try to justify in its motion.”

Put simply, writing a book review does not mean you can steal the book.

OpenAI, in a brief made public on Friday, says the law looks to purposes, not copies: “uses serving distinct purposes may warrant separate fair use analyses, but uses serving the same transformative purpose are evaluated together.” Of the transfers the authors call distribution, it says: “Each transfer was for the purpose of training.” It “did not ‘distribute’ books to contractors,” it says, but used “a custom-built web portal with restricted login-only access to display very short snippets to contractors evaluating model-generated summaries.” OpenAI calls it “undisputed” that it downloaded and copied the works “for the sole purpose of training LLMs”; the authors call that premise “genuinely disputed,” saying millions of downloaded books went instead to stock an internal library.

Both sides cite the same appeals ruling. The authors rely on the Third Circuit’s Sept. 29 decision in Thomson Reuters v. Ross Intelligence, while conceding that Ross “stated that it was not addressing generative AI models.”Thomson Reuters v. RossROSS trained a legal search engine on memos written with thousands of Westlaw headnotes, hoping to compete with Westlaw. Its AI “was not a generative AI,” the court wrote, and “ROSS’s use was not fair.” OpenAI quotes the ruling’s footnote: “Unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression.”

The dilution theory runs through both cases: the authors quote Judge Vince Chhabria’s 2025 ruling in Kadrey v. Meta on a tool “enabling the creation of a potentially endless stream of competing works.” Judge Chhabria nonetheless granted Meta summary judgment on its fair-use defense for training on that record, writing that “the plaintiffs presented no meaningful evidence on market dilution at all.” OpenAI calls Kadrey’s dilution discussion “dicta.”

Replies in New York are due Nov. 20.

Daily Cortex is written with AI models, including Anthropic’s Claude, and fact-checked with OpenAI’s GPT models.