The Bottom Line
- Two recent rulings found that the use of copyrighted works to train generative AI models was protected fair use. However, these rulings were very fact-specific and do not suggest that fair use will be found in other copyright disputes against AI companies.
- In cases where developers used copyrighted books to teach their AI algorithms how to use language to generate new outputs and did not reproduce any of the plaintiffs’ copyrighted works, such use was found to be highly transformative.
- Companies are more likely to be shielded from liability for infringement if they purchase or license data they seek to use for training AI models.
Two courts recently issued decisions of note regarding the use of copyrighted material for training generative artificial intelligence (AI) platforms. Both decisions — Bartz v. Anthropic PBC and Kadrey v. Meta Platforms — hail from the District Court for the Northern District of California and, while limited to the particular facts at issue, may offer insight as to how other courts could address similar issues.
In Anthropic and Meta, plaintiff authors alleged that AI companies infringed their copyrighted books by using them to train their proprietary large language models (LLMs). The judges in both cases found that the use of copyrighted books to develop and train these LLMs was highly transformative and constituted a fair use. Each decision, however, took different positions on key issues that courts are likely to continue debating, including how the use of pirated material to train AI platforms should be evaluated in conducting a fair use analysis.
How Did We Get Here?
Soon after the release of ChatGPT, Stable Diffusion, and other generative AI platforms, concerns arose regarding whether these companies had engaged in copyright infringement. Generative AI platforms develop and train their algorithms and LLMs on enormous sets of data scraped from the internet and other media sources, including copyrighted works.
In a series of lawsuits filed against several AI companies, plaintiffs from the news media, music publishing, film production, and other creative industries have alleged that unauthorized appropriation of their works for the purpose of developing and training for-profit tech products constitutes copyright infringement. In response, several AI companies have asserted that the fair use doctrine protects this practice and that such use is not infringing.
Why Were the Activities at Issue in Anthropic and Meta Considered a Fair Use?
Courts generally consider four factors in determining whether the use of particular copyrighted material is considered a fair use:
Fair Use Factor 1: Purpose and Character of the Use
Both decisions found that this first factor strongly favored the AI platform developers.
The Anthropic court described the use of copyrighted books to train Anthropic’s LLM chat tool, Claude, as “spectacularly” transformative, likening it to the way a person learns to read and write and noting that it allows Claude to freely converse in response to user prompts. Nonetheless, the court distinguished between Anthropic’s transformative use of books to train its AI model, on the one hand, and its non-transformative use and storage of pirated books to build a central research library, on the other hand. The ruling strongly disapproved of using pirated books which could have been lawfully purchased, stating that such “piracy of otherwise available copies is inherently, irredeemably infringing,” especially when they are used solely for the purpose of creating a library and not for training an LLM. The court ordered a trial to be held regarding Anthropic’s use of pirated books to create its central library, and the resulting damages it caused.
In Meta, the court also took the position that the use of copyrighted books to train Meta’s LLM, Llama, was highly transformative. The decision noted that the works were copied to develop a tool that can perform a wide range of functions, including translating text and conducting research. Like Anthropic, Meta also sourced some of its training materials from so-called “pirate libraries,” which the plaintiffs argued should disqualify Meta from asserting fair use. However, the court rejected this argument, noting that it “begs the question because the whole point of fair use analysis is to determine whether a given act of copying was unlawful.”
Fair Use Factor 2: Nature of the Copyrighted Work
Both courts found this factor weighed against a finding of fair use, as the works at issue in both cases are published fiction and non-fiction including expressive elements. As the Meta court noted, this second factor rarely plays a significant role in the determination of fair use disputes.
Fair Use Factor 3: Amount and Substantiality of Portion Used
In each case, the court acknowledged that Meta and Anthropic copied entirely or extensively from the copyrighted works at issue, but found that this factor nonetheless weighed in favor of fair use. The Anthropic court rationalized that extensive copying was reasonably necessary to train Anthropic’s LLM effectively and did not result in the works being a made available to the public. Likewise, the Meta decision concluded that copying full books was reasonable in light of the technical requirements of training Llama.
Fair Use Factor 4: Effect Upon the Market for the Copyrighted Work
While each court found that this fourth factor favored AI developers, the courts’ analyses were based on different conclusions.
The Anthropic decision concluded that training Claude on copyrighted works “did not and w[ould] not displace” the market for the original books. The court noted, however, that its analysis of this factor may have been different if the LLM produced infringing copies of the works.
On the other hand, the Meta decision faulted the plaintiffs for failing to make a compelling argument concerning market dilution, noting that generative AI training has enormous potential to flood the market with competing works. The court granted summary judgment to Meta, but stated that the ruling should not be interpreted as a blanket approval of the company’s conduct, and merely reflects the plaintiffs’ failure to present appropriate evidence and arguments relating to market harm.
What Do These Decisions Mean?
These decisions are limited in scope. They are not binding on other district courts, and the court in each case emphasized that its decision was limited to the specific facts on the record before it. But in similar cases, where AI developers use copyrighted books to teach an algorithm language in order to generate new outputs, as opposed to simply reproducing the copyrighted material, this use may be considered a transformative use.
An open question remains as to whether AI companies will be allowed to rely on content obtained unlawfully as part of the training process — or whether only lawfully acquired work (e.g., via purchase or license) will be considered permissible for transformative use. These decisions also do not address how transformative it would be for AI companies to train LLMs on different types of copyrighted material, such as artwork, sound recordings and photographs. And different facts and legal arguments presented in other cases could very well result in courts reaching a different conclusion as to the fourth factor, when examining the effect of the AI company’s use on the market for the copyrighted work.
In light of this evolving legal landscape, companies seeking to develop their own or even just use generative AI technologies should proceed with caution. Training AI technologies on data that has been lawfully purchased or licensed remains a safer course of action than using unlicensed material. Moreover, such companies are more likely to be shielded by fair use if their AI tools do not simply reproduce copyrighted training data. That said, as these cases illustrate, fair use decisions are highly dependent on the facts and arguments asserted by copyright owners and AI developers. As courts issue other fair use rulings in the months ahead, a clearer picture of best practices will likely emerge.