Generative AI Debate Braces for Post-Warhol Fair Use Impact

The US Supreme Court’s recent guidance on how courts must evaluate copyright law’s fair use doctrine will have major implications on the debate over generative artificial intelligence models that are trained on billions of images, texts, and other copyrighted works.

But exactly how far the high court’s lengthy opinion in Andy Warhol Foundation v. Goldsmith will sway that legal battle isn’t entirely clear, experts say.

Generative AI art programs released by Stability AI Inc., Midjourney Inc., and DeviantArt Inc. were hit with separate lawsuits this year from a group of artists and the photo database company Getty Images Inc., which claimed copyright infringement over the ingestion of their images.

While the AI companies haven’t yet invoked a fair use defense, legal observers point to similar fair use cases that have favored companies like Google that extract data from copyrighted works for a “transformative” purpose. That argument could be upended—or at least complicated—by the justices’ Warhol decision, attorneys said.

In its 7-2 decision, the Supreme Court emphasized the need for judges to closely examine whether the unauthorized copying was done for a commercial purpose when evaluating fair use.

Justice Sonia Sotomayor’s majority opinion said that the Andy Warhol Foundation’s decision to commercially license Warhol’s silkscreen prints of a photograph of the musician Prince to Vanity Fair magazine didn’t support a finding of fair use.

The colorized prints may have transformed the “expression, meaning, and message” of Lynn Goldsmith’s original photos, but that “must be weighed against other considerations, like commericalism,” Sotomayor wrote.

But the court’s holding was a narrow examination of only one of four fair-use factors that courts must examine.

“Warhol certainly throws important shade on the question of fair use and the extent to which it applies to machine learning,” said Shyamkrishna Balganesh, a copyright law professor at Columbia University. “But it’s far, far from answering that question.”

Looking at ‘Use’

The high court’s decision affirmed the idea that what matters when examining fair use is how the unauthorized copy is being used, not just similarities and differences between it and the original, Balganesh said.

The court said that the Warhol Foundation was competing with Goldsmith by licensing the prints to a magazine, the same market for the original photo. But, if instead, the foundation had displayed the prints in a nonprofit museum, the fair use analysis might have come out differently.

The emphasis on market competition drew applause from creative industries worried about generative AI models training on copyrighted work.

“We hope those who have relied on distorted—and now discredited—claims of ‘transformative use,’ such as those who use copyrighted works to train artificial intelligence systems without authorization, will revisit their practices in light of this important ruling,” Mitch Glazier, CEO of the Recording Industry Association of America, said in a statement.

Courts will likely give less weight to whether a generative AI output looks similar to an ingested copyrighted work, said Schuyler Moore, an entertainment attorney at Greenberg Glusker LLP. Instead, the new test will likely focus on how the output is used to compete with the original work.

“It’s the question of what you do with it,” he said. “And if you are targeting the same market, you’ve got a problem, that’s for sure.”

‘Google Books’ Defense

Legal observers have pointed to the US Court of Appeals for the Second Circuit’s 2015 decision in Authors Guild v. Google as the fair use case most analogous to AI training.

In Authors Guild, the appeals court blessed Google’s scanning of millions of books into a database that displayed snippets of the books and allowed text search as fair use. The Google Books Project that resulted from the copying was sufficiently transformative, the court held.

Defenders of AI argue that training machine-learning models on copyrighted works is similar: the models use the training data for outputs that are transformative—and therefore wholly new creations.

“Most of the generative AI models we see today are designed solely to create new content,” former general counsel at the US Copyright Office and current Latham & Watkins LLP partner Sy Damle said at a House hearing in May on copyright law and AI. “The models are not designed to reproduce copyrightable expression from training data and, in nearly all circumstances, do not do so.”

But the Warhol decision could muddy that analysis.

“To the extent that the prior argument was, ‘I’ve used input data to generate a new output, and look, that’s transformative,’ that is in some ways off the table,” Balganesh said. “Because what the court is saying is we have to look at the use.”

While Warhol did reference the Authors Guild case, the citation “was not an endorsement of the decision,” he said.

While the courts may take years to decide whether generative AI models are fair use, the litigation underway is many steps ahead of Congress, which hasn’t yet enacted legislation to regulate the burgeoning technology. In the meantime, some observers see peril for the AI industry.

“Copyright law is the only law that’s already in existence that could bring generative AI systems to their knees,” Pamela Samuelson, a copyright law professor at the University of California Berkeley, said at a lecture on AI last month. “If the court says ingesting is infringement, the whole thing can be destroyed.”