Bartz v. Anthropic: Settlement reached after landmark summary judgment and class certification
In Bartz v. Anthropic, 3:24-cv-05417 (N.D. Cal.), filed in August 2024 by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, the plaintiffs alleged that Anthropic relied on pirated e-books—downloaded from “shadow libraries” such as LibGen—to train its Claude large language models (LLMs).
From the outset, the case presented a test of how traditional copyright doctrines would be applied to LLM training.
In June 2025, Judge William Alsup of the Northern District of California issued a pivotal summary judgment ruling. He held that Anthropic’s use of lawfully acquired books for AI training was “quintessentially transformative” and thus protected by fair use. However, he simultaneously ruled that Anthropic’s creation and retention of a “central library” comprised in part of the pirated works was not transformative and constituted infringement. This partial win-and-loss positioned the case for trial on damages.
The decision stopped short of explicitly addressing whether training conducted with pirated source copies is fair use. However, Judge Alsup expressed skepticism that "any subsequent fair use” could ever justify pirated downloads. This suggests a doctrinal tension: while the fair‑use holding did not depend on lawful acquisition, the decision intimates that obtaining copies from piracy sites may erode the defense, even in a training context.
By mid-July, Judge Alsup certified a class of US copyright holders whose works Anthropic downloaded from the shadow libraries—potentially numbering in the millions. Because statutory copyright damages can reach US$150,000 per infringed work, the exposure theoretically extended into the hundreds of billions of dollars, raising existential risk for Anthropic. The company had sought an interlocutory appeal and a stay, but both were denied in early August, leaving trial scheduled for December.
On August 26, 2025, the parties filed notice of a proposed class-wide settlement, stating they had executed a binding term sheet and expected to enter a full settlement agreement by September 3. While the precise terms remain confidential, class counsel described the outcome as “historic” and beneficial to authors. The court set a deadline of September 5 for filing a motion for preliminary approval, with a hearing likely the week of September 8.
Key takeaways
The settlement halts a trial that might have tested the boundaries of copyright statutory damages. The class certification ruling also suggests that aggregation of claims is viable, magnifying risk exposure. The Bartz litigation could resonate beyond Anthropic. In fact, another group of plaintiffs—music publishers—already moved to amend their complaint against Anthropic last month to add piracy claims. We may see other AI platforms settle with creators as cases get closer to trial to avoid potentially astronomical damages.
Whether training LLMs with pirated data is defensible as fair use is hardly a decided issue. Bartz left this question open. Judge Alsup held that the training use was “exceedingly transformative” and therefore fair, but he cast doubt on whether downloading pirated source copies could ever be justified as reasonably necessary to such a use. By contrast, in Kadrey v. Meta, 3:23-cv-03417 (N.D. Cal.), Judge Vince Chhabria—also of the Northern District of California—squarely held that LLM training constitutes fair use regardless of whether the underlying materials were obtained from legitimate sources or shadow libraries. We may see higher courts resolve splits on this issue as cases develop.
While the settlement avoids a jury verdict, the case signaled that courts are willing to embrace transformative fair use in the AI training context, but not necessarily at the expense of condoning piracy. In addition to looking at whether outputs cross the line into infringement (not at issue in Bartz or in Kadrey), future disputes may hinge on whether training data was lawfully sourced.