Transformative Use Analysis in Bartz v. Anthropic AI Case Marred by Fatal Flaws
In late June, summary judgment orders were issued in two Northern District of California cases—Bartz v. Anthropic and Kadrey v. Meta—that sent shockwaves through the copyright and AI communities. In both cases, the court ultimately concluded that the unauthorized use of plaintiffs’ copyrighted works for AI training qualified as fair use. On the surface, these cases may seem like big wins for AI companies, but, as explained in previous blogs (here, here, and here), the orders reach conflicting conclusions on many issues and are not the “wins” for AI companies some commentors have made them out to be.
Importantly, the Bartz decision applied a flawed and woefully superficial analysis of transformative use under the first fair use factor. Below are the five biggest errors made by the Bartz court in its analysis of transformative use.
Fatal Error #1: Untethering Transformative Use Analysis From Its Legal Standard
The Bartz v. Anthropic order states that Anthropic’s use of copyrighted works for training is “exceedingly” transformative and that generative AI technology is “among the most transformative many of us will see in our lifetime.” Those conclusions seem to be based solely on the judge’s opinion about generative AI technology generally and are completely unmoored from the actual legal standard for determining transformative use. The Supreme Court has made clear that the central consideration under a transformative use analysis is whether the use “merely supersedes the objects of the original creation . . . (supplanting the original), or instead adds something new, with a further purpose or different character.” Nowhere in the Bartz decision does the court reference or apply this legal standard established by the Supreme Court in Campbell v. Acuff-Rose and recently confirmed by the Court in Warhol v. Goldsmith. Instead, as this article explains, the judge in the case seems to think that because generative AI is an innovative new technology, that makes it a transformative use. He simply concludes that “the ‘purpose and character’ of using works to train LLMs was transformative—spectacularly so.” In reaching that conclusion, the judge fails to offer any real explanation or analysis as to why the use of copyrighted works for training is transformative and neglects to actually apply the legal standard to the facts at hand.
Fatal Error #2: Wrongly Focusing on the Similarity of Outputs
The Bartz court says that Anthropic’s use of copyrighted works was transformative because there was no evidence of public-facing infringing output. There are two problems with this analysis. First, whether there was any substantially similar infringing output is irrelevant to whether making wholesale, exact copies of works at the ingestion stage qualifies as fair use. It’s important to understand that copyright owners’ rights are implicated when their works are reproduced to create datasets or ingested by generative AI systems, regardless of whether the AI systems generate infringing output or distribute infringing copies to end users. The Ninth Circuit has addressed this issue, most notably in Sega v. Accolade, when it rejected the defendant’s argument that intermediate copying does not infringe unless the end product of the copying is substantially similar to the copyrighted work. The Sega decision explains that “the Copyright Act does not distinguish between unauthorized copies of a copyrighted work on the basis of what stage of the alleged infringer’s work the unauthorized copies represent.” The Bartz court, which is part of the Ninth Circuit, should have recognized this critical principle.
Second, the focus on infringing output is wrong in the context of a transformative use analysis. Whether a use is transformative depends, in part, on the “purpose” of the use. In other words, the judge in the Bartz case should have been comparing the ultimate purpose of the use of the copyrighted work to the ultimate purpose of the use of the output generated by the AI—not whether the copyrighted works and the AI output were similar to one another. The Supreme Court made clear in its seminal Warhol v. Goldsmith decision that when the ultimate purpose of use is the same as the purpose of the copyrighted work being used, defendants’ resulting work acts as a substitute and the use therefore cannot be transformative.
In many cases of generative AI training, the ultimate purpose of the use of copyrighted works for training is to generate material that serves the same purpose as the original works. To be sure, there may be instances when the ultimate purpose of training an AI model is not to generate material that serves the same ultimate purpose of the ingested works, but the court ignores Warhol’s instruction and never analyzes these purposes because it wrongly focuses on output instead.
This critical nature of considering the ultimate purposes of the uses was also recognized by the U.S. Copyright Office in its report on AI training and fair use. In that report, the Copyright Office explains that when the ultimate goal is to generate works that occupy the same market space as the works used for training and “satisfy the same consumer desire,” the purpose is much less likely to be transformative. Ultimately, the report concludes that “making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries.” In sum, the Bartz decision is fatally flawed due to its complete failure to heed the Supreme Court’s Warhol instruction and the Copyright Office’s expertise in its transformative use analysis.
Fatal Error #3: Falsely Treating AI Training and Human Learning As the Same
In the Bartz case, the conclusion that the use was transformative seems to be based in large part on the judge’s specious comparison of LLM training to human learning and creation. The decision states that Anthropic’s purpose of using copyrighted works is “[l]ike any reader aspiring to be a writer” and therefore is not to “replicate or supplant them.” This comparison could not be more misplaced. There are myriad ways that AI ingestion and generation of output differs from the human experience, most significantly the speed, scale, and super-human ability of generative AI to memorize and regurgitate works that supplant those used for training makes such an analogy absurd. The ultimate purpose of a human reading a book is not invariably to create a new literary work that competes with what the human has read, whereas that is the exact purpose of many generative AI models. In the Kadrey v. Meta case, which was issued just two days later, the ridiculousness of AI training-human learning comparison, was characterized as “inapt” and “not remotely” the same.
Fatal Error #4: Failing to Consider Whether the Use Was Justified
The Supreme Court’s Warhol decision requires that for a use to be transformative, there must be a justification for using the work—namely, copying of the work at issue must be reasonably necessary to achieve the user’s new purpose. For example, in Warhol, the Supreme Court discusses the justification for using a work in the context of creating a parody because of the inherent necessity to mimic the underlying work. Conversely, Warhol explains that other uses, like satire, need not copy a specific work and thereby require a higher justification. Rather than discussing whether Anthropic’s use is reasonably necessary, the Bartz court fails to address justification at all in its transformative use analysis. Maybe Anthropic would satisfy the justification requirement in this case, but we simply do not know because the Bartz court never even attempts to discuss it in deciding whether the use was transformative, rendering the court’s conclusion even more suspect.
Fatal Error #5: Failing to Consider Commerciality
Another fatal flaw in the Bartz case relates primarily to the first factor more generally, as opposed to the transformative use analysis. While many courts in the past elevated transformative use over commerciality, the Supreme Court in Warhol disabused courts of that approach, explaining that commerciality can offset to some extent a finding of transformativeness. Warhol makes clear that even when a use has a further purpose or different character, “the degree of difference must be weighed against other considerations, like commercialism.”
Once again, the Bartz court drops the ball by omitting any discussion of commerciality, as required under the first factor. While the court does acknowledge that Anthropic is a “commercial outfit,” it fails to discuss the highly commercial nature of Anthropic’s use and certainly does not weigh such commerciality against transformative use to determine whether and to what extent the first factor would weigh in favor or against a finding of fair use. The Bartz court’s failure to engage in this analysis is a clear flaw that directly impacts its conclusion that the first fair use factor weighs in favor of Anthropic.
Conclusion
The court in Bartz v. Anthropic arrived at the conclusion that the unauthorized use of the plaintiffs’ works for training a generative AI model qualifies as fair use, but the order’s analyses of transformative use has many fatal flaws. The disregard for Ninth Circuit precedent and misapplication of the Supreme Court’s Warhol decision is nothing short of alarming, so much so that it’s hard to see how this decision will not be corrected on appeal. Judges in the many other ongoing AI infringement cases would be wise to recognize that the Bartz decision fails to correctly analyze transformative use and to use the decision as a roadmap for how not to analyze transformative use in an AI Infringement case.
If you aren’t already a member of the Copyright Alliance, you can join today by completing our Individual Creator Members membership form! Members gain access to monthly newsletters, educational webinars, and so much more — all for free!