Fair use rulings favor Meta and Anthropic but are limited
While the judges agreed with the fair use argument, it doesn't mean other lawsuits will have the same ruling. It's also likely that the Supreme Court will be the ultimate decider.
Anthropic and Meta's limited wins in their AI lawsuit cases this week shed light on whether using copyrighted material to train generative AI systems is, in fact, fair use.
Fair use is a legal argument used in copyright infringement cases. Under fair use, a party can use a copyrighted work without the owner's permission for purposes such as news reporting, teaching or research -- generally if the use of copyrighted materials serves the public interest.
In legal terms, it includes a four-factor test: the effect of the use on the potential market; the amount of the material used; the nature of the copyrighted work; and the purpose and character of use, which also examines transformative use, instances in which the new work derived from the copyrighted content adds substantially new meaning or expression.
Two types of fair use
In both lawsuits, the judges ruled that the AI vendors can claim fair use, based on two distinct factors.
In Meta's case, authors -- including Sarah Silverman and Ta-Nehisi Coates -- sued the social media giant in 2023 for violating copyright rules and using their books without permission to train Meta's Llama model.
U.S. District Judge Vince Chhabria ruled on Thursday that fair use applies in this case because of the lack of potential market impact. This means the authors suffered no financial loss due to Meta's use of its material to train its system.
Meanwhile, the day before, U.S. District Judge William Alsup concluded that Anthropic's fair use argument against the suit brought against the vendor by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson is also valid. The authors had accused the AI vendor of using their copyrighted books to train its AI chatbot, Claude.
Anthropic and Meta are the first two decisions as to whether training a generative AI model is fair use.
Michael GraifIntellectual property partner, Brown Rudnick LLP
The judge ruled that, because of generative AI's (GenAI) transformative power, fair use can be considered in the Anthropic case. However, the judge also ruled that the case would go to trial for Anthropic's use of pirated books. Digital piracy is when copyrighted material is used, reproduced or distributed without permission.
"Anthropic and Meta are the first two decisions as to whether training a generative AI model is fair use," said Michael Graif, intellectual property partner at Brown Rudnick. "There's now a different but related precedent for other courts to follow: that training generative models on copyrighted works is fair use."
Not a complete victory
While these initial outcomes of the cases appear favorable to the GenAI vendors, the same fair use factors might not be used in other AI lawsuit cases. Therefore, it's not a total victory.
"They weren't resounding victories, especially with that Meta decision," said Tamlin Bason, an analyst at Bloomberg Intelligence. "There is the potential for this market harm factor to complicate future fair use analysis.
"What the judge found is that this is not fair use," Bason said. "This goes beyond that. He believes we don't want to incentivize that behavior using this pirated data to train your large language model [LLM]."
Moreover, while these two judges ruled in favor of fair use, it doesn't mean other judges will do the same.
"Fair use is a very fact-specific, case-by-case determination," Bason said. "It's important to note that both cases only dealt with inputs."
This means that the cases only looked at what went into the LLMs and not what came out of them.
"If you read both decisions, it was very clear that these would have gone the other way if infringing outputs were involved," Bason continued.
He added that a case dealing with LLM input and output that has yet to be ruled on is TheNew York Times' case against OpenAI and its main financial backer, Microsoft.
Moreover, he added that many authors and content creators will likely continue to sue GenAI companies.
"There's going to be quite a bit of trial and error on how to best position your copyright infringement claims to move them forward in the litigation," he said.
Ultimately, these two cases are just the beginning in a long line of decisions expected in numerous AI lawsuits that have been filed and could be filed soon.
"These are just district court rulings," Graif said. "It remains to be seen whether these fair use rulings will be upheld on appeal. Ultimately, this will be an issue that the Supreme Court decides, but right now, we have two decisions on training a generative AI model."
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.
