PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2312.06646v5cs.AI

TL;DR

AI music royalties require a way to trace generated outputs to the training works that influenced them, a contribution that is latent rather than an explicit match. The paper proposes Generative Content ID, using efficient attribution methods in place of costly retraining; its experiments report that this approach approximates retraining-based attribution, while similarity-based proxies miss broader data contributions. The paper concludes that royalty distribution can be operationalized, while optimal distribution remains an open modeling problem.

  • Problem

    In generative music, training-data influence is latent, making it unclear how to measure a specific sample’s contribution to a specific generation.

    Source: §3.2 · p. 8

  • Method

    Generative Content ID attributes generated music’s value to causally influential training samples, using efficient Training Data Attribution to approximate counterfactual effects without retraining.

    Source: §1 · p. 2, §3.3.3 · p. 10

  • Results

    The authors report that gradient-based TDA correlates reasonably highly with retraining-based attribution, while perceived attribution aligns for the most influential samples but poorly captures broader data contributions.

    Source: §1 · p. 3, §4.3.4 · p. 14

  • Takeaways & Limitations

    The framework provides a prototype for measuring indirect training-data consumption, and royalty distribution parameters can tune income inequality among contributors.

    Source: §1 · p. 3, §5.2 · p. 16

    Optimal royalty distribution remains open because it requires better modeling of key stakeholders’ empirical behaviors.

    Source: §5.1 · p. 16

Abstract

from arXiv · show

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.

1 Introduction

The paper argues that preventing reuse and one-time licensing do not provide recurring, proportional compensation, while AI training makes it difficult to trace generated music to the works that influenced it. It proposes Generative Content ID and evaluates its attribution accuracy, its relation to perceived similarity, and its economic implications.

1 Introduction

  • Copyright disputes involve concerns about authorship, originality, and compensation, with studies reporting that many rights holders are unaware of or uncompensated for training uses.The introduction also points to high-profile copyright lawsuits across text and music.Source: §1 · p. 2
  • Preventive approaches can be brittle, reduce model quality when they filter or constrain content, and exclude creators rather than compensate them.The paper contrasts these approaches with royalty-based solutions.Source: §1 · p. 2
  • One-time licensing is a simpler stopgap, but does not provide the recurring, proportional compensation associated with music-industry royalty models.The proposed alternative links compensation to how training data contributes to generated outputs.Source: §1 · p. 2
  • Generative Content ID attributes value from generated music to the training samples that causally influenced its generation.It uses efficient Training Data Attribution methods to approximate the counterfactual effect of removing a training sample, avoiding costly retraining.Source: §1 · p. 2
  • The study examines whether TDA approximates retraining-based attribution and how perceived attribution relates to causal attribution and creator income distribution.Its experiments use public and proprietary music-industry data.Source: §1 · pp. 2–3
  • Gradient-based Training Data Attribution approximates retraining-based causal attribution, while perceived similarity aligns for the most influential samples but poorly captures broader contributions.The authors report that the mismatch may under-compensate “hidden influencers”; distribution parameters can also tune income inequality.Source: §1 · pp. 2–3

2 Background

The background traces a shift from removing derivative works toward monetizing them, then reviews consumption-based royalty mechanisms and the Spotify and YouTube models as design precedents for AI music royalties.

2 Background

  • The paper reviews legal and economic copyright context and examines Spotify and YouTube as case studies for identifying design requirements for an AI royalty system.The background frames its review around a transition from takedown to licensing models.Source: §2 · p. 3

2.1 Legal Standards and Economic Incentives in Copyright

  • Copyright infringement analysis commonly considers access and similarity, but identifying which training samples influenced an AI generation remains technically difficult.Perceived similarity is often used as a proxy for infringement, while the paper positions attribution as a way to investigate access and influence.Source: §2.1 · p. 3
  • Algorithmic licensing enables copyright owners to monetize derivative works rather than remove them, reflecting a shift away from Notice and Takedown.The passage cites evidence that rights holders mostly prefer monetization when given the choice.Source: §2.1 · p. 4

2.2 Digital Music Royalty Frameworks

  • Consumption-based royalties are emphasized because copyright-holder revenue can grow with platform success; distribution models include pro-rata, user-centric, and micro-sync licensing.The paper reviews prior research on allocation mechanisms and adopts pro-rata for its economic analysis due to its prevalence.Source: §2.2 · p. 4

2.3 Case Studies: Spotify and YouTube

  • Spotify and YouTube serve as archetypes for pro-rata and micro-sync models, respectively, and are compared by stakeholders, revenue sources, and royalty distribution.The case studies are selected as representative of dominant revenue-distribution paradigms.Source: §2.3 · p. 5

2.3.1 Spotify’s Royalty Framework

  • Spotify centrally shares revenue with copyright owners, while artists, labels, publishers, rights societies, listeners, and advertisers participate in different roles.Spotify typically interacts financially with labels, publishers, and collecting agencies rather than individual artists.Source: §2.3.1 · p. 5
  • Spotify separates subscription and advertising revenue pools, and its pro-rata model distributes royalties according to a track’s share of total platform streams.The passage reports that subscriptions accounted for 88% and advertisements 12% of Spotify revenue in 2021.Source: §2.3.1 · p. 5

2.3.2 YouTube Video’s Royalty Framework

  • YouTube monetizes music through its Partner Program, Content ID revenue sharing, and one-time licensing, with revenue primarily from advertising and partly from subscriptions.Content ID uses fingerprinting and machine learning to identify copyrighted content in uploads; the system has also faced criticism for false positives.Source: §2.3.2 · p. 6

3 A Computational Royalty Framework for Music Generative AI

The section proposes a royalty framework that traces AI-generated music to training data through causal attribution, then estimates that influence with scalable data-attribution methods. It also contrasts causal influence with perceived similarity as an alternative attribution proxy.

3.1 Business Models of AI Music Generation Platforms

  • AI music platforms offer services to users and corporate clients, with creators and rights holders as revenue recipients and subscriptions, licensing, ads, and custom composition among revenue sources.The described services include music creation, streaming, corporate solutions, and education.Source: §3.1 · p. 7

3.2 Proposed Royalty Framework for Music Generative AI

  • The proposed framework pools platform revenue, supports music generation and consumption, then distributes revenue according to training-data contribution scaled by generated-content revenue.This makes attribution of latent training-data contributions the technical bottleneck: unlike YouTube's direct music matches, a generative model learns statistical patterns.Source: §3.2 · p. 8

3.3 Generative Content ID

  • Generative Content ID traces generated music to the training samples that causally influenced it, measuring the indirect consumption of training data.The framework adapts the structural analogy of YouTube's Content ID to music generation.Source: §3.3 · p. 8, §3.3.1 · p. 8

3.3.1 Tracing Generated Music to Copyrighted Training Content

  • Leave-one-out influence measures a training sample's contribution as the utility difference between a model trained on all data and one retrained without that sample.The counterfactual removal-and-retraining comparison is the framework's definition of influence on a particular generated output.Source: §3.3.1 · p. 8

3.3.2 Music Generative AI: Model and Utility Function

  • For symbolic autoregressive music, log-likelihood makes segment-level influence exactly additive across events, enabling classification attribution methods to apply to music generation.The study uses symbolic music as a proof of concept; event-level and segment-level scores provide different attribution granularities.Source: §3.3.2 · pp. 9–10

3.3.3 Scalable Approximation via Efficient Data Attribution Techniques

  • Influence-function TDA approximates causal attribution without retraining for each removed sample; TRAK and LoGra reduce computation through random projections.The exact retraining-based calculation is computationally and economically prohibitive for modern generative models.Source: §3.3.3 · p. 10

3.4 An Alternative Solution Through Perceived Attribution

  • Perceived attribution treats similarity between training and generated music as a proxy for influence, using computational metrics because human expert review is costly at scale.The paper uses CLAP, MERT, and PMI as scalable proxies for human judgment.Source: §3.4 · p. 11

4 Empirical Analysis

The experiments test whether TDA can approximate retraining-based causal attribution and whether similarity-based legal proxies reflect training-data influence. They also show that royalty allocation choices shape simulated income inequality.

4.1 Experiment Settings

  • The experiments use Music Transformer with MAESTRO and Anticipatory Music Transformer fine-tuned on TheoryTab, generating samples from test-set prompts for attribution analysis.The training sets contain 5,000 MAESTRO pieces and 28,000 TheoryTab pieces.Source: §4.1 · p. 11

4.2 Validity: Approximating Causal Attribution with TDA

  • TDA methods TRAK and LoGra show significant positive correlations with retraining-based influence scores, with LoGra reaching 0.410 on TheoryTab and outperforming the random baseline.The authors use LoGra for the remainder of the paper.Source: Table 1 · p. 12, §4.2 · p. 12

4.3.1 Comparison 1: Computational Similarity vs. Retraining

  • All three computational similarity metrics have correlations close to zero with retraining, compared with 0.366 for TRAK and 0.410 for LoGra in the same setting.This comparison uses average Spearman rank correlation.Source: Table 2 · p. 12, §4.3.1 · p. 12

4.3.2 Comparison 2: Computational Similarity vs. TDA

  • Computational similarity aligns poorly with TDA on average, but all three metrics give higher similarity scores to the few training samples ranked most influential by TDA.The authors also observe a U-shape, suggesting some samples perceived as similar may have high negative influence.Source: §4.3.2 · p. 13, Figure 4 · p. 13

4.3.3 Comparison 3: TDA vs. Human Evaluation

  • Human similarity ratings follow a trend similar to MERT scores, including a U-shaped pattern across the ranked sample groups.The study compares human ratings with groups ranked by MERT or LoGra scores.Source: §4.3.3 · p. 13, Figure 5 · p. 14

4.3.4 The Influence of “Hidden Influencers”

  • Removing High-Attribution-Low-Similarity samples significantly increases the generated music’s perplexity, indicating that perceived similarity misses samples that contribute to generation.Figure 6 compares full, random-removal, and HA-LS-removal conditions; error bars show standard errors over five runs.Source: Figure 6 · p. 14, §4.3.4 · p. 14

4.3 Comparison with Legal Proxies: Causal Attribution vs. Perceived Attribution

  • Similarity-based attribution aligns poorly with causal attribution overall, and the authors find that hidden influencers could be under-compensated under a similarity-only royalty framework.The authors describe this as a complex, non-linear relationship and warn that under-compensation risks weakening rights-holder incentives and long-term model quality.Source: §4.3.4 · p. 14

4.4 Economic Impact Simulation

  • At K = 500, the top 1% of training points receive 2.6% of royalties under Top-K Uniform and 10.8% under Top-K Rank-Weighted.The simulation finds that inequality decreases as K increases; Top-K Uniform produces the most equal outcomes, while Top-K Rank-Weighted concentrates royalties among the highest-ranked contributors.Source: Table 3 · p. 15, §4.4 · p. 15

5 Discussion and Conclusion

The discussion treats royalty attribution as both a regulatory infrastructure question and a way to make creators’ compensation recurring. The conclusion presents Generative Content ID as a prototype for tracing generated music back to training data.

5.1 Managerial Implications

  • Regulators are urged to mandate faithful royalty frameworks and white-box access to training dynamics, because perceived similarity is insufficient for attribution.The paper contrasts such access with external black-box audits of generated output.Source: §5.1 · p. 16
  • A royalty model would replace one-time data licensing with recurring creator revenue that grows proportionally with platform revenue.The paper describes this as an equity-like stake for artists and labels.Source: §5.1 · p. 16
  • Platforms can tune contributor income distribution through parameters such as K, but optimizing royalties requires better models of stakeholders’ empirical behavior.The paper identifies that modeling as an open problem for future work.Source: §5.1 · p. 16

5.2 Conclusion

  • The conclusion frames AI-generated content valuation as a technical challenge requiring robust infrastructure for sustainable innovation.Source: §5.2 · p. 16
  • Generative Content ID combines digital music platforms’ economic logic with training data attribution to operationalize measuring indirect consumption of training data.The paper presents it as a foundational prototype and blueprint for a fair, scalable generative-AI economy.Source: §5.2 · p. 16

A A Primer on the Concepts of Music Royalties

The primer distinguishes recurring streaming royalties from typically one-time synchronization licenses and describes mechanical royalties for reproductions and derivatives. It cautions that these simplified categories are not formal or mutually exclusive legal classifications.

B Detailed Stakeholder Description

The stakeholder description lists five groups connected to Spotify’s music and revenue flows: creators, rights organizations, listeners, and advertisers.

  • Spotify’s listed stakeholders are artists and creators, record labels and publishers, rights societies and collecting agencies, listeners and subscribers, and advertisers.Source: B Detailed Stakeholder Description · p. 20

C More Details of Experimental Setup

The experimental setup uses symbolic-music models trained on TheoryTab and MAESTRO, then evaluates attribution methods on generated continuations. A separate human study compares perceived similarity ratings for generated and training music pairs.

C More Details of Experimental Setup

D Details of Human Studies

  • In the IRB-approved listening study, participants rated generated-sample and training-sample pairs on a five-point similarity scale, with candidate samples ranked by TDA or MERT.Source: D Details of Human Studies · p. 21
  • Fifteen college-student participants passed an identical-pair attention check; each contributed 30 eligible pair ratings, and the study took 30–60 minutes.Source: D Details of Human Studies · p. 21

E Additional Experiment Results on Test Dataset

On real music targets, gradient-based TDA correlates reasonably well with retraining, while computational similarity methods align poorly with the retraining benchmark. Similarity versus TDA rank follows the U-shaped pattern also observed for generated music.

E Additional Experiment Results on Test Dataset

F.1 Training Data Attribution Algorithms

The section describes TRAK and LoGra as scalable TDA methods, using random projection or low-rank gradient projection and factorization to reduce computational cost.

TRAK

  • TRAK uses random projection to reduce computational cost while calculating attribution scores.The method uses projected gradients and a one-minus-correct-class-probability weight; the experiments use dattri and ensembles of 50 models for Music Transformer and 3 for Anticipatory Music Transformer.Source: §F.1 · p. 22

LoGra

  • LoGra applies low-rank gradient projection and factorization to improve efficiency (Choe et al., 2024).For linear layers, it uses a Kronecker-product structure for the projection matrix; the Hessian can be replaced by the empirical Fisher Information Matrix to reduce computational cost.Source: §F.1 · p. 22
  • The LoGra experiments use dattri and ensembles of 10 full-training-set models for Music Transformer and 1 model for Anticipatory Music Transformer.Source: §F.1 · p. 22

F.2 Computational Music Similarity Methods

The section defines three similarity measures for comparing training segments with generated segments: CLAP and MERT audio representations, and symbolic melodic identity.

Computational music similarity methods

  • Figure 7 plots average similarity score against TDA rank for MERT, CLAP, and PMI.Source: Figure 7 · p. 23

CLAP representation similarity

  • CLAP similarity renders symbolic music to audio, encodes clips with the CLAP audio encoder, and compares their embeddings by cosine similarity.CLAP representations are learned through contrastive pretraining on audio-caption pairs in a shared embedding space.Source: §F.2 · pp. 22–23

MERT representation similarity

  • MERT similarity renders symbolic music to audio, mean-pools final-layer hidden representations, and compares the resulting embeddings by cosine similarity.MERT is a self-supervised music representation model with a 25-layer Transformer encoder.Source: §F.2 · p. 23

Percent melodic identity (PMI)

  • PMI aligns pitch-class melodies with Needleman–Wunsch and affine gap penalties, then normalizes exact matches by sequence lengths.The implementation takes the maximum similarity across transpositions to capture melodies differing by a global key shift.Source: §F.2 · p. 23

F.3 High-Attribution-Low-Similarity Samples

The experiment uses LoGra for attribution and MERT for computational similarity to select high-attribution, low-similarity samples, then evaluates target-sample perplexity.

High-Attribution-Low-Similarity Samples

  • The experiment uses LoGra for TDA and MERT for computational similarity.Source: §F.3 · p. 23

High-Attribution Low-Similarity (HA-LS) Selection

  • HA-LS selection standardizes attribution and similarity scores, then selects samples with the largest attribution-minus-similarity difference.Source: §F.3 · p. 23

Perplexity

  • Perplexity measures performance on the target generated sample; higher values mean lower model-assigned probability and degraded generation capability for that sequence.Source: §F.3 · p. 23
Loading 2312.06646v5…