The recent ruling in a lawsuit against Anthropic highlights a growing complexity in how courts are approaching fair use in the context of AI training. Judge William Alsup held that developing Anthropic’s Claude model was “spectacularly” transformative. But he declined to grant summary judgment, pointing to the fact that the training materials were acquired by downloading millions of books from piracy websites.
Days later, in a similar case, another court granted summary judgment to Meta, holding that the level of transformation rendered the source and method of acquiring the materials inconsequential.
Anthropic is now seeking interlocutory review of Judge Alsup's decision, arguing that the entire AI training process should be viewed as a unified transformative use, regardless of where or how the data was sourced. Anthropic argues that the appeal should be accepted because not only does the case present novel questions, but the fact that the Meta judge recently found similar data acquisition to be fair use reflects a divergence that could create uncertainty unless addressed.
In undertaking the fair use analysis, courts have traditionally considered whether the defendant’s use harms the market for the original work. In mass-copying cases like this one, where authors allege that entire books were taken from unauthorized sources, judges may begin to look at whether AI companies made any effort to work with the rights holders, even though the use does not directly compete with or undermine the authors' existing markets.
This raises another set of questions. While no court has held that a fair use defense requires prior efforts to license and courts have consistently found that a transformative purpose can outweigh market harm, as litigation continues and more facts come to light about how training data is gathered, these questions may become more central. If developers can show that licensing was impractical or unavailable, that may strengthen their position. If they ignore existing paths to permission, some judges may view that as relevant to the fairness of the use.
Moreover, it creates a strategic consideration for rights holders themselves. If courts increasingly weigh the availability of licensing in the AI-training context within the fair use analysis, the absence of a clear market could weaken claims of market harm. By creating viable pathways to permission through opt-in registries, collective licenses, standardized terms, or other mechanisms, authors and publishers can strengthen the argument that licensing is both possible and reasonable.
The emergence of such frameworks would not only bolster the market harm factor, it could also shift the tone of future litigation. Developers would have a clearer roadmap for risk mitigation, and courts would be better positioned to distinguish between parties who made efforts to respect copyright and those who did not. In that sense, a functioning licensing market is both a business opportunity for rights holders and a potential element of how courts define fair use in the AI era.
The Anthropic appeal may not resolve these questions, but here and otherwise, the courts and litigants will need to grapple with how transformative the AI output is, as well as what steps, if any, developers are expected to take regarding the source of materials before relying on massive volumes of copyrighted work.
These cases disagree on whether AI model development should be viewed as a single process or broken into separate steps, and whether the source of training material matters for fair use analysis.
www.law360.com/...
[View source.]
