Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Computational Copyright: Towards A Royalty Model for Music Generative AI

Junwei Deng, Xirui Jiang, Shiyuan Zhang, Shichang Zhang, Himabindu Lakkaraju, Ruijiang Gao, Chris Donahue, Jiaqi W. Ma

TL;DR

The paper tackles how to sustain creative incentives in music generated by AI by proposing Generative Content ID, a causal attribution framework that ties AI outputs to their training data via Training Data Attribution (TDA). It formalizes leave-one-out influence as a counterfactual utility difference, and provides scalable gradient-based TDA methods (TRAK, LoGra) to approximate true causality without retraining. Empirical analysis on MAESTRO and TheoryTab shows TDA closely tracks retraining-based attribution, while revealing that legal proxies based on similarity imperfectly capture data influence, particularly for less obvious contributors. The authors also simulate economic outcomes under different royalty schemes, demonstrating that distribution mechanisms can significantly shape income inequality and platform governance. Overall, the work offers a principled, scalable foundation for royalty-based governance of music generative AI and highlights regulatory implications for fair compensation of data contributors.

Abstract

The rapid rise of generative AI has intensified copyright and economic tensions in creative industries, particularly in music. Current approaches addressing this challenge often focus on preventing infringement or establishing one-time licensing, which fail to provide the sustainable, recurring economic incentives necessary to maintain creative ecosystems. To address this gap, we propose Generative Content ID, a framework for scalable and faithful royalty attribution in music generative AI. Adapting the idea of YouTube's Content ID, it attributes the value of AI-generated music back to the specific training content that causally influenced its generation, a process we term as causal attribution. However, naively quantifying the causal influence requires counterfactually retraining the model on subsets of training data, which is infeasible. We address this challenge using efficient Training Data Attribution (TDA) methods to approximate causal attribution at scale. We further conduct empirical analysis of the framework on public and proprietary datasets. First, we demonstrate that the scalable TDA methods provide a faithful approximation of the "gold-standard" but costly retraining-based causal attribution, showing the feasibility of the proposed royalty framework. Second, we investigate the relationship between the perceived similarity employed by legal practices and our causal attribution reflecting the true AI training mechanics. We find that while perceived similarity can capture the most influential samples, it fails to account for the broader data contribution that drives model utility, suggesting similarity-based legal proxies are ill-suited for royalty distribution. Overall, this work provides a principled and operational foundation for royalty-based economic governance of music generative AI.

Computational Copyright: Towards A Royalty Model for Music Generative AI

TL;DR

The paper tackles how to sustain creative incentives in music generated by AI by proposing Generative Content ID, a causal attribution framework that ties AI outputs to their training data via Training Data Attribution (TDA). It formalizes leave-one-out influence as a counterfactual utility difference, and provides scalable gradient-based TDA methods (TRAK, LoGra) to approximate true causality without retraining. Empirical analysis on MAESTRO and TheoryTab shows TDA closely tracks retraining-based attribution, while revealing that legal proxies based on similarity imperfectly capture data influence, particularly for less obvious contributors. The authors also simulate economic outcomes under different royalty schemes, demonstrating that distribution mechanisms can significantly shape income inequality and platform governance. Overall, the work offers a principled, scalable foundation for royalty-based governance of music generative AI and highlights regulatory implications for fair compensation of data contributors.

Abstract

The rapid rise of generative AI has intensified copyright and economic tensions in creative industries, particularly in music. Current approaches addressing this challenge often focus on preventing infringement or establishing one-time licensing, which fail to provide the sustainable, recurring economic incentives necessary to maintain creative ecosystems. To address this gap, we propose Generative Content ID, a framework for scalable and faithful royalty attribution in music generative AI. Adapting the idea of YouTube's Content ID, it attributes the value of AI-generated music back to the specific training content that causally influenced its generation, a process we term as causal attribution. However, naively quantifying the causal influence requires counterfactually retraining the model on subsets of training data, which is infeasible. We address this challenge using efficient Training Data Attribution (TDA) methods to approximate causal attribution at scale. We further conduct empirical analysis of the framework on public and proprietary datasets. First, we demonstrate that the scalable TDA methods provide a faithful approximation of the "gold-standard" but costly retraining-based causal attribution, showing the feasibility of the proposed royalty framework. Second, we investigate the relationship between the perceived similarity employed by legal practices and our causal attribution reflecting the true AI training mechanics. We find that while perceived similarity can capture the most influential samples, it fails to account for the broader data contribution that drives model utility, suggesting similarity-based legal proxies are ill-suited for royalty distribution. Overall, this work provides a principled and operational foundation for royalty-based economic governance of music generative AI.
Paper Structure (76 sections, 1 theorem, 12 equations, 7 figures, 5 tables)

This paper contains 76 sections, 1 theorem, 12 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Legal Standards and Economic Incentives in Copyright
  4. Legal Standards.
  5. Economic Incentives.
  6. Digital Music Royalty Frameworks
  7. Related Literature on Existing Royalty Allocation Mechanisms:
  8. Case Studies: Spotify and YouTube
  9. Spotify's Royalty Framework
  10. Stakeholders.
  11. Revenue Sources.
  12. Royalty Distribution.
  13. YouTube Video's Royalty Framework
  14. Stakeholders.
  15. Revenue Sources.
  16. ...and 61 more sections

Key Result

Lemma 3.1

Let $h_S$ be an autoregressive generative model trained on dataset $S$, and let $m_i \in S$ be a specific training sample. Let $\hat{m} = (e_k, \dots, e_l)$ denote a generated sequence of events. If the utility function $f(\hat{m}, h)$ is defined as the log-likelihood, i.e., $f(\hat{m}, h) = \log P(

Figures (7)

  • Figure 1: The royalty framework of Spotify.
  • Figure 2: The royalty framework of Youtube Video.
  • Figure 3: The computational royalty framework designed for music generative AI.
  • Figure 4: Average similarity score vs. rank by average TDA scores: MERT (left), CLAP (middle), and PMI (right). The x-axis represents the rank of each training sample based on the average TDA scores across generated samples, and the y-axis represents the average similarity scores between each training sample and all the generated samples.
  • Figure 5: Box plots comparing human evaluation with computational similarity or TDA. The x-axis represents groups of training samples divided by increasing rank according to computational similarity scores (MERT) or TDA scores (LoGra) calculated on the generated samples. The y-axis represents the similarity scores between the pair of training and generated samples rated by human participants. The red line indicates the median, and the blue dotted line indicates the mean.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Lemma 3.1: Additivity of Influence under Log-Likelihood
  • proof