Music giants sue Cambridge AI startup over its alleged use of their songs

Several of the nation's largest record labels today sued Suno, a Cambridge firm started by AI gurus whose offering lets users create their own songs - based, the labels charge, on their copyrighted work.

In the suit, filed in US District Court in Boston, the music companies allege the company sucked their offerings into its computers to build a music-based "generative AI" system. In December, Microsoft announced it would integrate Suno's offering into its Copilot laptop and desktop AI-based operating system.

Upon information and belief, and consistent with the basic facts of how generative AI works, the content Suno used to "train" its AI model includes reams of Copyrighted Recordings that Suno reproduced without permission from Plaintiffs. Suno could not have built a model capable of producing audio so similar to the Copyrighted Recordings without the initial act of copying those recordings. This explains why one of Suno's investors has publicly recognized that Suno's service is likely to spawn litigation and that defending lawsuits from music labels is "the risk we had to underwrite when we invested in the company."

The suit continues:

When a user prompts Suno's service with a text input (e.g., make a jazz song about New York), the service generates an audio output by making generalizations about what the audio output should sound like based on the prompt and the corpus of sound recordings on which it was trained. Certain features of the outputs from Suno's model betray that it was trained on particular data—in this case, the Copyrighted Recordings. In particular, Suno's product frequently generates outputs with strong resemblance to the Copyrighted Recordings, a telltale sign that such recordings were included in its training data.

In technical terms, by generating outputs that mimic sound recordings in its training corpus, Suno's model reflects the machine-learning phenomenon known as "overfitting." An AI model is "overfitted" when it is too closely adapted to the data on which it was trained, making it difficult for the model to generalize to new data sets. One symptom of overfitting is a model that replicates portions of its training data. To take a simplified example, if a user inputs the prompt "a jazz song about New York" into an overfitted AI model, the model may output a file that closely resembles one of the jazz tracks on which it trained. As the myriad examples discussed below reflect, Suno's model obviously was trained on the Copyrighted Recordings. This infringement cannot be cured by simply loosening the model's fit or by implementing technical guardrails that make it less likely that outputs will match excerpts of the Copyrighted Recordings. In other words, modifying Suno's offering in a way that better conceals its training data would not alter the fact that Suno infringed the Copyrighted Recordings the moment it copied them to create its model.

The basic point is that Suno's model requires a vast corpus of sound recordings in order to output synthetic music files that are convincing imitations of human music. Suno's corpus includes the body of recorded music that people listen to in their everyday lives. Because of their sheer popularity and exposure, the Copyrighted Recordings had to be included within Suno's training data for Suno's model to be successful at creating the desired human-sounding outputs.

The companies are seeking a formal legal declaration that Suno is bad and should feel bad, should knock off what its doing immediately and pay at least $150,000 per copyrighted song sucked into its system, or damages to be determined by the court or Suno's profits, whichever is more, plus attorneys' fees and interest.

The plaintiffs are: UMG Recordings, Capitol Records, Sony Music Entertainment, Atlantic Recording, Atlantic Records Group, Rhino Entertainment, the All Blacks, Warner Music International Services and Warner Records.

Complete complaint (1.6M PDF).