Today, Tim Hwang, with help from Joshua Levine and Sam Roland, filed an amicus brief regarding Bartz et al. v. Anthropic PBC.
Introduction and Summary of Argument
As the district court sees it, this case “exemplifies the classic litigation” for Rule 23: every author is “aggrieved” by Anthropic’s downloads from “pirate libraries,” so proof of the “class-wide wrong” will be “straightforward.” Order on Class Certification, Bartz et al. v. Anthropic PBC, No. C 24-05417 WHA, at 1 (N.D. Cal. July 17, 2025). On that view, seven million titles proceed in lockstep as questions of ownership wait until judgment.
To establish this logic, the district court took a snapshot of the pirated works in one moment of time, implied that Anthropic had become a digital book storage company, sua sponte created a class off this claim, ignored or carved out the true upstream and downstream uses of the storage, then certified said class by making a merits adjudication on fair use—while simultaneously claiming that any fair-use defense would be manageable class-wide. Facts point the other way on each step. As the district court noted, Tom Turvey testified that Anthropic did not intend its general research library—the class’s predicate—merely to store files; rather, it designed the library to “use [them] for research” and to “inform … products.” Order on Fair Use, Bartz et al. v. Anthropic PBC, No. C 24-05417 WHA, at 4-5 (N.D. Cal. June 23, 2025). Two distinct end uses.
As for the class itself, Rule 23(b)(3) demands that “the questions of law or fact common to class members predominate” (Fed. R. Civ. P. 23(b)(3)); yet as seen in the district court’s analysis of LLM training—the “one further use” recognized—the books varied in the extent they contained the “well-curated facts, well-organized analyses, and captivating fictional narratives” required to train a model. Fair Use Order at 5–6. Even so, the district court claims that such variance in book type and market impact—the core of the second and fourth prongs of fair use—give it “no reason to believe that it would not be susceptible to treatment by common evidence and common methods” in further analysis. Class Cert. Order at 26. By folding every book into one pirated-library class and postponing ownership and value, the order misses Rule 23(b)(3)’s demand for predominance. This claim of unvaried fair use can stand only so far as one book is as good as another—a claim which the district court itself has shown above to be untrue.
The district court may be correct that Anthropic should have purged its collection after deciding to stop training on the material, but in its haste to certify a class, it has created a dynamic that will preclude proper analysis of that question. It has drawn a qualitative distinction among the LLM training copies, the pirated library copies, and any further downstream copies, but no quantitative one. Training is declared fair use, downstream LLM research may also be, but the prerequisite storage for both is not—that Anthropic is “liable” for. Id. at 24. For how long may a company retain such information when training a model? Is retention acceptable for downstream research purposes? Does the extent of acceptability vary with the content retained? The class papers over all these critical questions.
Normally, these issues should be handled on appeal. But the same over-reaching order leaves a threat of up to $1 trillion in potential statutory damages, creating “death-knell” settlement pressure. Chamberlan v. Ford Motor Co., 402 F.3d 952 (9th Cir. 2005). Worse still, the blessing of this class would influence future litigation, in turn shrinking domestic innovation and weakening national security.
Because the order ignores predominance, and uses statutory damages as leverage, the Court should grant Rule 23(f) review, vacate certification, and remand with instructions to deny class treatment.
If the Court decides against decertification, Rule 23 supplies narrower tools: it may certify only the liability issue under Rule 23(c)(4) and defer ownership, willfulness, and damages; or it may order a bellwether trial on fair-use market harm and revisit class lines once the record is complete. Either of those steps would cure the predominance defect while keeping real claims alive.
