Towards A Royalty Model for Music Generative AI

View PDF of the article Computational Copyright: Toward a Property Model for music-Generating AI, by Junwei Deng and 7 other authors

View PDF HTML (beta)

a summary:The rapid rise of generative AI has exacerbated economic and copyright tensions in the creative industries, especially in music. Existing approaches to this challenge often focus on preventing infringement or creating a one-time license, which fails to provide the sustainable and recurring economic incentives needed to sustain creative ecosystems. To address this gap, we propose Generative Content ID, a framework for scalable and faithful property rights attribution in music-generative AI. Adapting the idea of ​​YouTube’s content ID, it attributes the value of AI-generated music to the specific training content that causally influenced its production, a process we call causal attribution. However, naively identifying a causal effect requires retraining the model on subsets of the training data, which is not possible. We address this challenge by using efficient training data attribution (TDA) methods to approximate causal attribution at scale.

We also perform an empirical analysis of the framework on public and private datasets. First, we show that scalable TDA methods provide a faithful approximation of the “gold standard” but expensive retraining-based causal attribution, demonstrating the feasibility of the proposed equity framework. Second, we investigate the relationship between perceived similarity used by legal practices and causal attributions that reflect real AI training mechanisms. We find that although observed similarity can capture the most influential samples, it fails to take into account the broader data contribution that drives the utility of the model, suggesting that similarity-based proxies are inappropriate for allocating equity.

Overall, this work provides a principled and operational basis for equity-based economic management of music-generating AI.