Bartz et al. v. Anthropic PBC, No. C 24-05417 (N.D. Cal. 2025)
In Bartz et al. v. Anthropic, the U.S. District Court for the Northern District of California considered whether Anthropic’s use of copyrighted books—many sourced from pirated libraries, others destructively scanned from purchased print copies—to train its Claude AI models qualified as fair use. The court held that the latter was fair use; the former was not. The ruling raises significant new questions about the boundaries of fair use, particularly regarding AI-driven applications and the digital reproduction of copyrighted works. For example, does this holding implicitly allow the mass digitization of all fairly obtained copyrighted works so long as it is solely for “internal use” with no public access? At its core, the ruling asserts that the transformative nature of AI training outweighs traditional copyright concerns, at least where outputs have not reproduced the original works. Crucial to this ruling, the plaintiff authors did not allege, nor did the record show, that any output of the AI model infringed upon their copyrighted works.
The court’s characterization of Anthropic’s AI training as “exceedingly” and “spectacularly” transformative underscores a broad view of fair use. It analogized AI training as akin to a human reader internalizing texts to produce original writings—an unsettling comparison for authors wary of AI’s voracious ingestion of their works. This ruling implies that as long as AI-generated content does not replicate and distribute the original works, transformative fair use is likely to shield such practices. See also Authors Guild v. Google, 804 F.3d 202 (2d Cir. 2015).
In a less controversial turn, the court also explicitly rejected the retention of pirated books within Anthropic’s “generalized data area,” signaling that fair use has its limits: it cannot excuse a “library” of pirated loot. The court found that whether or not Anthropic intended a fair use for the copies – or no use at all – the initial act of piracy was determinative. That Anthropic later paid for copies of the same pirated works did not sanitize the initial act.
The court’s approval of converting legally obtained print books into digital copies for internal reference muddies traditional copyright boundaries. Citing precedents like Sony Betamax (Sony Corp. v. Universal City Studios, 464 U.S. 417 (1984)), the decision validates space-saving and searchability enhancements, such as converting physical copies to digital, so long as there is no external distribution. Importantly, the ruling emphasizes that the transformation in this case did not multiply the number of copies available or infringe distribution rights. Further, the transformation did not create a derivative work because no content was added or subtracted to the copyrighted material.
This decision suggests that courts may be willing to sacrifice a space in the marketplace traditionally reserved for copyright holders at the altar of technological innovation—provided that the use is genuinely transformative and not merely a fig leaf for traditional copying. Another California court has already ruled in similar fashion on a related fair use issue. See Kadrey et al v. Meta Platforms, 3-23-cv-03417 (NDCA Jun. 25, 2025). Regardless, significant uncertainties remain. How will courts navigate situations where the internal digital copies are inadvertently circulated or used in commercial outputs? The court accepted that Claude did not produce infringing outputs in this case, but the question of “model leakage” or memorization looms large. Plaintiffs seeking to challenge similar conduct should seek to show that specific outputs are substantially similar to their protected works in order to test the limits of what qualifies as a derivative work under 17 U.S.C. § 106(2). Ultimately, the decision outlines a porous but discernible boundary: While it robustly supports AI training as transformative, it also provides some limits on fair use protections, particularly regarding the source and retention of materials. Litigants and legislators alike would do well to heed the court’s subtle warning: fair use may be generous, but it remains “fair”—and piracy, no matter how transformative or well intentioned, is not.
