Training AI music models is about to get very expensive

🌈 Abstract

The article discusses the legal challenges facing AI music startups Suno and Udio, who are being sued by major record labels for allegedly using copyrighted music in their training data. It explores the implications of this case for the future of AI music generation and the potential need for licensing deals between AI companies and music rights holders.

🙋 Q&A

[01] AI Music Startups Sued by Record Labels

1. What are the key allegations made by the record labels against Suno and Udio?

• The record labels claim that Suno and Udio used copyrighted music in their training data "at an almost unimaginable scale," allowing their AI models to generate songs that "imitate the qualities of genuine human sound recordings."
• The labels allege that the AI music tools are more imitative than generative, with examples of songs generated that closely resemble copyrighted works by artists like the Temptations and ABBA.

2. How do the AI companies respond to the allegations?

• Both Suno and Udio have released statements mentioning efforts to ensure their models don't imitate copyrighted works, but they have not specified whether their training sets contain such works.
• Suno CEO Mikey Shulman says their training set is "both industry standard and legal" but the exact recipe is proprietary.

3. What are the potential outcomes of this legal case?

• The case could go in three ways:
  1. The AI startups win, and the court determines they did not violate fair use or imitate copyrighted works too closely.
  2. The court finds the AI companies did not violate fair use in their training but must better control their models' output to avoid imitating copyrighted works.
  3. The court finds fault on both the training and output sides, forcing the companies to restructure through costly licensing deals or potentially go bankrupt.

[02] Implications for the Future of AI Music

1. How do the stakes differ for AI music compared to other generative AI applications?

• Music in the public domain is much more limited, and it's more difficult to create music worth listening to compared to generating readable text or passable illustrations.
• The rights in music are more concentrated, with the three major record labels collectively owning more than 10 million songs.

2. What are the potential paths forward for AI music companies if they are required to pay for licensing deals?

• The company with the deepest pockets may end up on top, as expensive licensing deals with record labels could be the only way forward.
• AI companies could try to bypass licensing by building models exclusively on public domain music, but the resulting models would be far inferior to their current offerings.

3. What are the challenges around music licensing for AI companies?

• Music licensing is complicated by the fact that two different copyrights are involved: the song copyright and the master recording copyright.
• Some artists own the masters of their catalogs, while others have the record labels retaining the masters, which could complicate licensing negotiations.

[03] Musician Perspectives on AI Music

1. How have musician groups responded to the rise of AI music?

• There is a divide, with some groups like SAG-AFTRA allowing AI clones of member voices with minimum compensation, while others like the Indie Musicians Caucus express concerns about AI replacing musicians without consent or compensation.
• The American Federation of Musicians (AFM) appears hesitant to facilitate deals between musicians and AI companies, seeing it as a difficult choice between "a swarm of fire ants crawling all over you, or roll around in a bed of broken glass."