What do Elton John, Billie Eilish, Paul McCartney, Radiohead, Sting and Dua Lipa all have in common apart from great music? The answer is they have all objected to use of music for training AI large language models (LLMs).
Musicians are one of many types of content creators to complain about how AI LLMs are being trained on their content. LLMs typically ingest vast amounts of data from the internet. This gets sorted, processed (through something called Transformer Architecture), then fine tuned, so the AI tool can then create outputs. Musical content has to undergo additional processing into symbolic music representations to be understood by the Transformer Architecture. The concerns revolve around AI deepfakes, unauthorised vocal cloning, musical likenesses in AI-generated songs, personality rights abuses and most fundamentally loss of royalty opportunities.
Over 1,000 British musicians, including prominent artists like Kate Bush, Cat Stevens, and Annie Lennox, released a silent album titled "Is This What We Want?". The track list spelled out a message protesting the UK government's proposal regarding AI and copyright, which the artists fear would legalize music theft for AI companies' benefit. Hundreds of artists across the music industry have signed open letters and statements condemning the "predatory" and "irresponsible" use of their work to train AI LLMs. These initiatives emphasize that such practices devalue human creativity, infringe on artists' rights, and threaten the music ecosystem.
There are a number of fundamental copyright law issues are being raised by AI’s emergence and music and musicians are at the forefront of the debate. The issues include:
1. Whether training LLMs constitutes copyright infringement? Some laws specifically allow it e.g. Singapore, EU - in order to encourage AI companies to invest there.
2. Are copyright holder opt out systems workable? UK and HK are considering exceptions to LLM training called opt outs. The EU exception requires a machine readable opt out so AI LLMs can read and then reject such works. But when most works are already published without them?
3. Whether Fair Use applies? The US tends to take a wider view of fair use, but many US disputes have arisen. The Berne Convention three step test would seem to be the core consideration – whether AI LLM training conflicts with normal exploitation of the work and whether it unreasonably prejudices the legitimate interests of the author. That will likely need judges to determine it.
4. Whether outputs from AI tools can constitute copyright infringement? Many musicians express concern that AI-generated music, which often mimics existing styles and voices, will undermine the originality and hard work invested by human artists, potentially leading to a devaluation of their work and a reduction in royalties. Even complete deepfake songs are now commonplace – as this Taylor Swift AI version shows - That's So True (Piano Version) - Taylor Swift | AI Cover.
A number of legal actions are going on around the world. Major record labels, including Universal Music Group, Sony Music Entertainment, and Warner Records, have filed copyright infringement lawsuits against AI music startups like Suno and Udio. These lawsuits allege that the AI companies trained their models using copyrighted sound recordings without permission, seeking significant damages. The labels argue that this constitutes a massive scale of unlicensed copying and does not qualify as fair use. The AI companies typically argue this falls under fair use or as some kind of intermediate processing (US law permits this). They also accuse record labels of trying to stifle competition. Several record labels presented evidence of AI-generated tracks that are strikingly similar or even nearly identical to copyrighted songs.
Meanwhile other steps are being adopted. Some musicians are exploring technical methods to disrupt AI training. This involves introducing inaudible audio tweaks into their songs, which are imperceptible to humans but can disrupt AI models attempting to learn from the music. Musician Benn Jordan (The Flashbulb) has been a proponent of this approach with a tool called "Poisonify." Many artists and industry organizations are advocating for greater transparency from AI companies regarding their training data and for the establishment of licensing agreements that would ensure fair compensation for the use of copyrighted material. After all music licensing is perhaps the most advanced in the world with performance, mechanical, synchronisation and communication licenses from a variety of Collecting Societies around the world in operation. Could an AI training license be developed?
All of the above relates to the dispute between creators and the AI and tech world. A separate legal issue is the creation of AI content, and lack of human authorship, so absence of copyright protection. If the commercial and consumer world starts adopting AI created content produced at very low cost, what happens to musician artists ?
Over the 21st century musicians faced file sharing and the wars with Napster and Pirate Bay, then streaming piracy and fighting for fair royalties with online streaming services. Now the latest battle is with AI; they say with some degree of concern that the future of human musical creativity is at stake.
