New Way to Pay Artists When AI Generates Music

When AI generates music, it's often unclear which training songs influenced the output. This is important because artists whose work helped train the AI might deserve payment. Current methods for tracing influence, called Training Data Attribution (TDA), tend to give too much credit to a few songs for every output, which would mean those songs get paid for everything.

A new study proposes breaking down TDA scores into three levels: global (influence on the model overall), group (influence on a cluster of outputs), and specific (influence on a single output). They found that global and group influences are significant in all tested music AI models. This means that simply using raw TDA scores would unfairly distribute money.

The researchers introduce the term 'functional dependency' to describe how training data relates to the model and its outputs, distinguishing different scopes of influence. This allows for a more nuanced interpretation of contributions, which is crucial for legal, economic, and ethical discussions about remunerating artists. In practice, this could lead to a system where AI-generated music pays royalties more accurately, ensuring that artists are compensated fairly based on their actual influence. As AI music becomes more common, such methods will be essential to protect creators' livelihoods and maintain trust in the technology.