Last year, when a “fake Drake” song created using artificial intelligence made major headlines, we questioned whether generative AI would have the biggest impact on the music industry since Napster. Well, stay tuned to find out.

Earlier this week, the “big three” record labels in the U.S. (Warner, Sony, and Universal) brought separate copyright infringement claims against two generative AI music startups, Suno and Udio (whose AI model was recently used to create the song sampled by Metro Boomin in the viral “BBL Drizzy” Drake diss).

According to the complaints, both defendants’ AI models allow users to generate “digital music files that sound like genuine human sound recordings in response to basic inputs” like text. The plaintiffs allege that both defendants copied “decades worth of the world’s most popular sound recordings” to train their AI models to generate similar outputs. As proof, they claim that both AI services “repeatedly generated outputs that closely matched” the copyrighted sound recordings referenced in certain inputs—including ABBA’s “Dancing Queen” and Jerry Lee Lewis’ “Great Balls of Fire”—when plaintiffs referenced "specific subject matter, genres, artists, instruments, vocal styles, and the like” in their written prompts.

In each case, the record companies are seeking an injunction to prevent the unauthorized use of their copyrighted sound recordings, including to train defendants’ AI models and generate infringing outputs. The plaintiffs also seek statutory damages under 17 U.S.C. § 504(c) in amounts up to $150,000 per work infringed, which could ultimately be staggering depending on the data used to train each model.

Anticipating a “fair use” defense from the AI startups, the labels proactively assert that each factor in the analysis weighs against a finding of fair use, focusing on the fact that the AI generated music files could serve as substitutes for and compete with the original recordings, including on streaming platforms, and potentially eliminate the existing market for licensing sound recordings. 

Other courts are currently addressing the issue of fair use as applied to various types of AI tools such as large language models (LLMs) and generative art tools. Any decisions in those cases may have a significant impact on the outcomes here, which we will continue to cover in our blogs.

In the meantime, these groundbreaking cases further demonstrate the importance of licensing training data and developing AI models in a way that does not infringe on copyrighted works such as art, text, or sound recordings. They also highlight the importance of setting up guardrails on AI models to minimize the risk of users generating infringing outputs.