AI Talk
This week, major record labels including Sony Music Entertainment, UMG Recordings, and Warner Music sued Suno and Udio.com for copyright infringement. What’s interesting about this case is the record labels allege infringement based on how closely the outputs of these AI music generation companies resemble their copyrighted sound recordings. For the official filings, Suno’s complaint is here, and Udio’s complaint is here.
Sunio and Udio are AI startups that take text prompts and generate songs in the style of the user prompts. They both launched in the last year and offer free and monthly subscriptions. The paid versions allow users to generate more songs per month and to sell them commercially.
The record labels allege direct copyright infringement of sound recordings they own or have exclusive rights to, via copying of those songs for purposes of training Suno and Udio’s AI models. While the record labels have no direct proof of copying, they managed to produce outputs from both companies’ platforms through targeted prompting, such as specifying lyrics from songs, decade, genre, topic, and even artist. The outputs purportedly contain virtually identical melodies, riffs, lyrics, and vocals are nearly indistinguishable from the original artists’ works. On top of it all, Udio co-founder and CEO’s supposed vision for the company is to create “music that sounds indistinguishable from music that’s created by professional human producers.” Note that the plaintiffs are NOT alleging infringement in the outputs themselves.
The record labels are seeking up to $150,000 per work infringed in statutory damages. Yikes.
While they haven’t yet responded, both companies are likely to respond with the fair use defense, which I covered in a previous post. Whether fair use is a viable defense for copying sound recordings for AI training is TBD with the Copyright Office, which should be issuing guidance this year. However, major AI companies’ stance so far has been that ingestion and use of copyrighted works for training data are highly transformative, the models’ purpose is to generate new content and not to copy the training data, and because of that, there is no impact on the market for the existing copyrighted works. This defense may have some legs here if the new content generated didn’t resemble the training data, either in lyrics, melodies, or vocals.
The interesting thing about this case is how closely the outputs supposedly resemble copyrighted works. The court will have to examine the fair use factors with that in mind. Here’s my hot take:
Based on these facts, here are some suggested best practices: