‘Clear Evidence’ Shows Udio, Suno Trained Their AI Models on Stolen Music, Allege Labels

The labels filed their complaint against Udio in U.S. District Court for Southern New York in Manhattan (docket 1:24-cv-04777) and against Suno in U.S. District Court for Massachusetts in Boston (docket 1:24-cv-11611). Hueston Hennigan is the labels' counsel in both actions.

It’s critical that AI technology “be implemented responsibly, ethically, and legally,” said the labels. AI companies “must abide by the laws that protect human creativity and ingenuity,” they said. Nothing exempts AI technology from copyright law or excuses AI companies “from playing by the rules,” they said. The twin lawsuits seek “to enforce these basic principles,” they said.

The Udio and Suno generative AI services allow users “to generate digital music files that sound like genuine human sound recordings in response to basic inputs,” said the labels. “The capacity for a generative AI service to produce convincing imitations of genuine sound recordings starts with copying a vast range of sound recordings,” they said. When those that develop such services steal copyrighted sound recordings, the services’ “synthetic musical outputs could saturate the market with machine-generated content that will directly compete with, cheapen, and ultimately drown out the genuine sound recordings” on which the services are built, they said.

Foundational principles of copyright law “dictate that copying protected sound recordings for the purpose of developing an AI product requires permission from rightsholders,” said the labels. Such AI offerings otherwise “will erode the value of the artistic works that comprise the essential raw materials that allow them to function in the first place,” they said. “If left unmoored from existing and longstanding legal constraints, such products could supplant, rather than support, genuine human creativity,” they said.

Building and operating AI services like Udio and Sono requires at the outset “copying and ingesting massive amounts of data” to train an AI software model to generate outputs, said the labels. For Udio and Suno “specifically,” that process “involved copying decades worth of the world’s most popular sound recordings,” and then “ingesting” those copies into the Udio and Sono AI models so they can generate outputs “that imitate the qualities of genuine human sound recordings,” they said.

Udio and Suno charge many of their users monthly fees to access their products and produce digital music files, “which are designed to entertain, evoke emotion, and stoke passion just like the genuine sound recordings” that those services copied, said the labels. In light of the foundation of the Udio and Suno businesses “to exploit copyrighted sound recordings without permission,” the services have been “deliberately evasive” about what exactly they have copied, they said.

If there were any doubt about Udio’s and Suno’s unauthorized copying, the services dispelled it “by effectively conceding in pre-litigation correspondence” that they copied copyrighted sound recordings, said the labels. When the labels “directly accused” Udio of copying the labels’ sound recordings to train their models, Udio and Suno didn’t “deny or proffer any facts to undermine those allegations,” they said.

It would have been simple for Udio and Suno to say that they used “other, legally acquired recordings,” if that were the case, said the labels. The services instead “deflected and disingenuously asserted” that their training data is competitively sensitive and constitutes trade secrets, despite being based on publicly available music for music fans, they said. Udio and Suno also claimed that their large-scale copying of sound recordings is fair use, “which was telling because fair use only arises as a defense to an otherwise unauthorized use of a copyrighted work,” it said.

The labels could have proceeded with their litigation “based solely” on eliciting those reasonable inferences of copying, said the complaints. But the labels’ claims “are based on much more,” they said. That's because the labels tested Udio’s and Suno’s products, they said. The testing “generated outputs using a series of prompts that pinpoint a particular sound recording by referencing specific subject matter, genre, artist, instruments, vocal style, and the like,” they said.

The Udio and Suno services “repeatedly generated outputs that closely matched the targeted copyrighted sound recording,” said the labels. That means that Udio and Suno copied those copyrighted sound recordings to include in their training data, they said. The public also has observed, and the labels have confirmed, that “even less targeted prompts” can cause the Udio and Suno services “to generate outputs that resemble specific recording artists and specific copyrighted sound recordings,” they said. Such outputs are “clear evidence” that Udio and Suno trained their models on the labels’ copyrighted sound recordings, they said.

At their core, both cases are about “ensuring that copyright continues to incentivize human invention and imagination, as it has for centuries,” said the labels. Achieving this end doesn’t require “stunting technological innovation,” they said. But it does require that Udio and Suno “adhere to copyright law and respect the creators” whose works allow them to function in the first place, they said.

The recording industry “has embraced AI and we are already partnering and collaborating with responsible developers to build sustainable AI tools centered on human creativity that put artists and songwriters in charge,” said Mitch Glazier, chairman-CEO of the Recording Industry Association, in a Monday statement trumpeting the "landmark" lawsuits. But the recording industry can only succeed if AI developers “are willing to work together with us,” he said. Unlicensed services like Udio and Suno that claim it's fair use “to copy an artist's life's work and exploit it for their own profit without consent or pay set back the promise of genuinely innovative AI for us all,” he said.