Contribution
Attribution for Musical Generative AI: A New Perspective
* Presenting author
Abstract:
As musical generative AI systems have become more prevalent in recent years, questions have arisen about data provenance. Given the recent EU AI Act, revenue attribution methods—which distribute license revenues back to the original rights holders of the training data—have become an important part of the ecosystem. Exactly how to fairly perform such attribution is unclear. Some companies measure the similarity of generated outputs to each song in the training data. Other options include a uniform attribution, and attribution based on popularity. Further methods claim to calculate an attribution based on the model weights, but such claims have yet to be proven possible. We instead propose Creative Weight Attribution, which incorporates 25+ well-founded musicological features. For each, we weigh each training song in terms of complexity and uniqueness in the training set, allowing for a fairer measurement of what an AI actually learns while also being scientifically feasible and musicologically grounded. They are combined with meta-data analysis and measures of human resonance (popularity) of any training track. This method is thus applicable in cases where no output is available, or where money isn’t easily assigned directly to individual outputs, e.g. when income is through subscription fees.