Judgments on AI and Fair Use Favor Anthropic and Meta—for Now
In the publishing world, technology companies’ use of copyrighted content to train generative AI tools has been a matter of growing concern. Indeed, in a recent report on outputs from NISO workshops on AI and publishing, standardized licensing models and legal terminology for AI usage emerged as one of several priorities for the information community: “There is a need for shared, flexible licensing templates or consistent contractual terms that define rights for training, grounding, and inference. These model licenses would facilitate consistent interpretation, reduce negotiation costs, and help ensure fair compensation.” Two recent rulings in US district court however, cast doubt on when and (in one case) whether licensing is necessarily required, in the eyes of the law, to train AI large language models (LLMs).
Anthropic, sued by three authors for using copyrighted material to train its LLM, Claude, asked for a summary judgment in the case. On June 23, Judge William Alsup granted that request in part, ruling that Anthropic's use of the material fell under the law’s definition of fair use, because the resulting AI-generated content represented something new and altogether different. However, the judge allowed the case against Anthropic for downloading the pirated works used for training Claude to move forward.
In a separate case, Meta was also sued by a group of authors for using copyrighted material to train its AI model, Llama. Just two days after the Anthropic ruling, Judge Vince Chhabria dismissed the case against Meta. The ruling did not state that Meta's ingestion of the copyrighted content fell under the definition of fair use. Rather, it stated that the plaintiffs failed to demonstrate that the AI-generated content based on their works would dilute the market for them or affect their value. The judge went on to note that he could envision cases in which the AI-generated content would affect the market for the copyrighted material and thus violate the terms of fair use.
In some ways, the rulings, although both arguably victories (or partial victories) for the technology companies, are contradictory, and if you’re unclear now on whether using copyrighted content to train AI is—or isn’t—fair use, you’re not alone. The issue seems likely to see many more days in court.