PaperlayerView exact revision on arXiv

Source-linked AI summary

What's a Credit Worth? A Market Framework for Attribution-Aware Compensation in Generative Music

Luyang Zhang, Xirui Jiang, Junwei Deng, Beibei Li, Jiaqi W. Ma, Chris Donahue

arXiv:2607.00641v1cs.CYcs.LG

TL;DR

Generative music raises the question of how to compensate creators whose catalogs train commercially used models. This paper develops a catalog-level payment framework that incorporates attribution informativeness and benchmarks attribution methods against leave-one-catalog-out ground truth. It finds that noisy current attribution favors fixed fees, while more precise attribution can benefit creators and platforms.

  • Problem

    How should creators be compensated when their catalogs train generative music models used commercially?

    Source: §1 · p. 2

  • Method

    The paper derives a catalog-level payment framework incorporating attribution informativeness and measures attribution methods against leave-one-catalog-out ground truth.

    Source: §1 · p. 2

  • Results

    Current attribution methods remain below the royalty threshold across audio and symbolic settings; in audio, they close about 10% of the welfare gap.

    Source: §1 · p. 2, §4.1 · p. 8

  • Takeaways & Limitations

    Better attribution can benefit creators and platforms, and platforms should benchmark multiple methods rather than commit to one.

    Source: §1 · p. 2

    The framework analyzes both symmetric information and the more practical case where creators cannot reproduce proprietary platform attribution without model access.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

Advances in generative AI are rapidly increasing the quality and commercial value of generated music, and this progress depends on large catalogs of creators' recordings. This raises a central question for platform design: how should creators be compensated when their work is used to train generative AI models that in turn produce commercial outputs? We develop a framework for fairly compensating creators in generative-music markets, where each creator's payment depends on a data-attribution score estimating their contribution to model outputs. Compared to past compensation frameworks, our framework has two unique considerations: (1) attribution is traced to entire creator catalogs, not individual songs, and (2) the informativeness (signal-to-noise ratio) of the attribution score is an input to the payment mechanism. The framework yields a closed-form payment rule per creator and measures the welfare cost of inaccurate attribution for both creators and the platform. Whether the welfare-optimal contract is royalty-based or takes the form of fixed-fee licensing depends on how informative attribution is for that creator's catalog. We show that better attribution translates directly into welfare gains for both creators and the platform, yet under multi-platform competition a platform only captures gains from attribution improvements when its signal becomes the most precise in the market. To ground our framework in empirical behavior, we train acoustic and symbolic music generation models and measure the informativeness of scalable attribution techniques against a leave-one-catalog-out ground truth. Our experiments reveal that noisy attribution signals push payment toward fixed-fee licensing and diminish welfare for both creators and the platform, providing an economic motivation for further research on improved attribution.

1 Introduction

The paper links noisy data attribution to payment design, developing a framework in which catalog-level attribution informativeness shapes creator contracts. Its theoretical results and empirical benchmarks connect attribution precision to contract choice and welfare.

1 Introduction

  • The paper unifies data-attribution methods, which produce noisy contribution estimates, with contract design for payments under noisy signals.Prior music-attribution work addressed measurement noise without incorporating it into payment mechanisms.Source: §1 · p. 2
  • The framework derives a closed-form creator payment rule balancing a fixed fee with royalties and quantifies welfare loss from imperfect attribution.It treats whole creator catalogs, rather than individual songs, as the units of attribution and compensation.Source: §1 · p. 2
  • More precise attribution can benefit creators and platforms, while noisy signals favor fixed-fee licensing over royalties.Under multi-platform competition, a platform captures gains from improved attribution only when its signal becomes the market’s most precise.Source: §1 · p. 2
  • Across 22 audio catalogs and a 100-catalog symbolic setting, current attribution methods sit below the royalty threshold; audio attribution noise makes fixed-fee licensing welfare-optimal for nearly all creators.Source: §1 · p. 2
  • The paper contributes a generative-music market framework, an empirical benchmark of attribution informativeness, and implications for industry policy and governance.The benchmark covers model paradigms and content modalities used by commercial music AI systems.Source: §1 · p. 2

2 Related work

Prior work addresses data markets and contract design, data-attribution methods, and generative music models. This paper connects these strands by incorporating attribution informativeness into payment design for music data.

Data pricing and contract design

  • Data-market and contract-theory work offers pricing tools, but the cited approaches do not combine ML attribution with contract design or explicitly accommodate creator-level differences in signal informativeness.Zhang et al. [61] assume risk-neutral creators and a reliable attribution score.Source: §2 · p. 3

Data attribution and music attribution

  • Music attribution research applies diverse methods to generative music, while this paper uses attribution informativeness to shape payment structures.Prior work broadly seeks to improve attribution informativeness; music applications include royalty assignment [6] [12] [17].Source: §2 · p. 3

Generative music models

  • Generative music uses codec language models and latent-audio diffusion models, with growing commercial value motivating study of economic sustainability.The paper focuses on payment mechanisms for training data used by these models.Source: §2 · p. 3

3 A framework for attribution-aware compensation

The framework models contracts between a music AI platform and creators licensing distinct catalogs, with payment design shaped by attribution quality and creator preferences. It derives contract and welfare implications and examines how market structure, creator incentives, and transparency affect them.

3.1 Framework for musical market and data attribution

  • The framework models a music AI platform licensing catalogs from creators and paying each through contracts tied to attribution of generated outputs.Catalogs are distinct and licensed wholesale, and attribution informativeness is assumed accessible to both creators and the platform in the transparent-information case.Source: §3.1 · pp. 3–4
  • Informativeness measures how well measured attribution tracks a catalog’s true contribution; under classical measurement error, it ranges from 0 for no information to 1 for a perfect linear predictor.The measure is an OLS slope, and the [0, 1] bound applies under the stated classical-noise regime.Source: §3.1 · p. 4
  • Creators receive a fixed fee, royalties, or both, and their mean–variance utility captures a preference for stable payments.Creators participate when contract utility meets their outside option; the analysis considers both symmetric and asymmetric information about attribution quality.Source: §3.1 · pp. 4–5

3.2 Equilibrium contract and welfare cost

  • The attribution-aligned benchmark shifts from royalty-dominant contracts at high informativeness to fixed-fee-dominant contracts at low informativeness.The royalty rate balances payment–contribution alignment against risk from noisy royalty exposure; the fixed fee implements the transfer needed for participation.Source: §3.2 · pp. 5–6
  • The welfare loss from imperfect attribution is concentrated among low-informativeness creators, and the marginal welfare gain from improvement is largest for the noisiest creators.The loss scales convexly with (1 − I_j)^2, and the proposed water-filling allocation directs most marginal improvement effort toward them.Source: §3.2 · p. 6

3.3 Market and policy implications

  • In multi-platform competition, a trailing platform earns no private return from improving attribution until it overtakes the leader, although total welfare is weakly increasing throughout.This result is stated for the specified equilibrium with Gaussian noise and homogeneous demand.Source: §3.3 · pp. 6–7
  • Uniform contracts can drive distinctive, high-outside-option creators to exit, and losses compound because their styles cannot be replaced by remaining catalogs.Under non-additive platform value, the welfare loss from exits is superlinear in the number of exiting creators.Source: §3.3 · p. 7
  • Attribution scores can be manipulated by smoothing their spread, while attribution transparency can reduce welfare losses from differences in creators’ beliefs about signal quality.The paper states that no linear contract deters score smoothing and identifies independent pipeline oversight as an institutional remedy.Source: §3.3 · pp. 7–8

4 Empirical instrument and measurement

Across audio and symbolic music settings, current attribution is too noisy for royalty-based contracts to dominate fixed fees for nearly all measured creators. The market simulation links greater attribution precision to welfare gains, but finds that substantial improvement is needed and platform incentives to invest can remain weak under competition.

4 Empirical instrument and measurement

  • The empirical instrument compares attribution methods with leave-catalog-out ground truth across audio diffusion and symbolic music generation, measuring informativeness, calibration, and calibrated SNR.The audio setting uses 22 catalogs and EKFAC/D-TRAK; the symbolic pilot uses 100 artists and LoGra/TRAK.Source: §4 · p. 8, §4.1 · p. 8, §4.2 · pp. 8–9

4.1 Setting 1 — Latent-diffusion audio attribution

  • EKFAC and D-TRAK match on Spearman correlation but their calibration scales differ by roughly 10^3, with some creators’ attribution-score signs reversing across methods.Because welfare loss depends on calibration, rank agreement alone can mask materially different contract prescriptions.Source: §4.1 · p. 8
  • Current audio methods close about 10% of the welfare gap, and at most 1 of 22 creators clears the royalty threshold under either method.The other 21 creators are where each unit of precision improvement pays off most, according to the authors.Source: §4.1 · p. 8, Figure 2 · p. 9

4.2 Setting 2 — Symbolic music attribution

  • Across audio diffusion and symbolic generation, no creator clears the royalty threshold under either symbolic method, and calibrated SNR is roughly an order of magnitude below it.The authors report the same shortfall across four attribution methods and two model domains.Source: §4.2 · p. 9, Figure 2 · p. 9
  • On symbolic data, LoGra outperforms TRAK, and their calibration gap is ∼10^13, larger than the ∼10^3 gap between audio methods.Methods with similar rank correlations can therefore prescribe materially different contracts.Source: §4.2 · p. 9

4.3 Musical market simulation

  • At α=2, no audio creator clears the royalty threshold under either method, and the symbolic population is similarly fixed-fee-dominant.Only one 31-track classical catalog is royalty-eligible under D-TRAK at α≤1.Source: §4.3 · p. 9, Figure 2 · p. 9
  • 25–40× improvement in calibrated SNR is needed to close 50% of the welfare gap, raising median creator SNR from ≈0.05 to ≈1.The simulation treats these levels as a precision target for royalty-based compensation to become viable.Source: §4.3 · p. 9
  • In the two-platform simulation, EKFAC wins for 14 of 22 creators and D-TRAK for 8, while the median surplus gap is too small to make either platform overtake the other through improvement.The results instantiate the framework’s underinvestment prediction on the measured pool.Source: §4.3 · pp. 9–10, Figure 3 · p. 10

5 Conclusion

The conclusion frames creator compensation as a contract choice driven by how accurately a platform can trace outputs to creator catalogs. The supplied appendix passages clarify that the general results allow arbitrary attribution-noise structures, while the closed-form SNR shortcuts require additional conditions.

5 Conclusion

  • The framework makes each creator’s contract form depend on how accurately the platform traces generated outputs back to that creator’s catalog.The conclusion identifies extensions to growing catalogs, cross-media models, and attribution oversight as future work.Source: §5 · p. 10

A Gaussian special case and simplifying conditions

  • The main theoretical results hold for any noise structure through the general informativeness measure I_j, but the closed-form SNR expressions require three additional conditions.Source: A Gaussian special case and… · p. 14
  • With the Gaussian shortcut, informativeness is I_j = SNR_j/(1 + SNR_j), and multiplicative bias can be calibrated by dividing the measured estimate by its method-specific scale.The welfare-loss formula’s shortcut additionally uses the stated revenue-independence condition.Source: A Gaussian special case and… · p. 14

B.1 Proof of Theorem 1

  • Theorem 1’s proof maximizes a risk-adjusted alignment criterion distinct from platform profit, which yields the pure fixed-fee baseline.The optimal royalty rate is projected onto the feasible interval [0, 1].Source: §B.1 · pp. 14–15

B.2 Proof of Proposition 1

  • The proof of Proposition 1 characterizes benchmark loss as the quadratic gap from the perfect-attribution benchmark rate.The Gaussian version substitutes I_j = SNR_j/(1 + SNR_j) and the simulations’ variance convention.Source: §B.2 · p. 15

B.3 Endogenous quality choice

The endogenous-quality extension lets creators choose effort that scales their contribution, so royalties affect both attribution alignment and effort incentives. Its effort-only optimal royalty is unique, while the full welfare optimum trades off effort surplus against alignment loss.

B.3 Endogenous quality choice

  • Under multiplicative quality scaling, informativeness is invariant to effort, and the platform commits to a contract before the creator chooses effort.The extension assumes participation’s outside option does not depend on the royalty and that the participation constraint binds.Source: §B.3 · p. 15
  • The effort-only welfare objective has a unique maximizing royalty in (0, 1), and effort responds positively to the royalty below its threshold.The effort cost parameter and risk aversion both increase the denominator governing optimal effort.Source: §B.3 · pp. 15–16
  • The welfare-optimal royalty trades off alignment loss against the moral-hazard role of royalties; at pilot parameters, it is about 0.98 for all three creators, an order of magnitude above the alignment-only rate.As effort costs grow, effort welfare vanishes and the optimum converges to the static alignment rate.Source: §B.3 · p. 16

C Incentive compatibility: full analysis

  • The incentive-compatibility appendix develops a three-tier creator-manipulation hierarchy, private-risk-aversion screening, and detection tests for variance shrinkage.It also states that the hierarchy is invariant to competition under exclusive licensing.Source: C Incentive compatibility: full analysis · p. 16

C.1 Proof of the three-tier hierarchy

The appendix specifies manipulation incentives and related screening, detection, and competition results. Its conditions distinguish distortions that monitoring can deter from variance shrinkage that linear contracts cannot deter.

C.1 Proof of the three-tier hierarchy

  • Tier 2 mean inflation is not self-policing and requires sufficiently strong external monitoring, while Tier 3 variance shrinkage cannot be deterred by any linear contract with a positive royalty.The Tier 3 result assumes the attribution estimate is independent of demand, or demand is constant.Source: §C.1 · p. 17

C.2 SNR-squared manipulation–welfare ratio

  • Under uniform demand, the maximum Tier 3 manipulation gain relative to welfare loss is SNR_j^2; with stochastic demand, SNR_j^2 is an upper bound, and monitoring expenditure should scale super-linearly with attribution precision.The bound is attained in the uniform-demand limit.Source: §C.2 · pp. 17–18

C.3 Statistical detection of Tier-3 manipulation

  • Under a Gaussian model, a variance-ratio test has power ≥0.95 at T≥50 for shrinkage ϵ≥0.3, complementing institutional prevention with statistical detection.Source: §C.3 · p. 18

C.4 Mussa–Rosen screening with private risk aversion

  • With private risk aversion satisfying the monotone hazard rate condition, the platform’s revenue-maximizing incentive-compatible menu is implementable in dominant strategies without ironing.Under stochastic demand with Cov(â_jπ, π)≠0, separation between signal incentives and menu incentives is at most approximate.Source: §C.4 · p. 18

C.5 Dynamic self-policing and competition invariance

  • Under exclusive licensing, the three-tier manipulation hierarchy is invariant to the number of Bertrand competitors and creators’ outside options; proportional detection suppresses bias but cannot achieve exact truthfulness.Source: §C.5 · pp. 18–19

C.6 Failure modes

  • Attribution pooling is unprofitable when Cov(â_i, â_j)>−Var(â_j)/2; the passage says this condition is typically satisfied in music markets because stylistic similarity drives positive cross-catalog correlation.Source: §C.6 · p. 19

D Welfare cost of uniform contracts

Uniform contracts misprice attribution risk and can drive distinctive creators to exit, producing risk-compensation and lost-contribution welfare costs. The multi-platform extension considers exclusive licensing, platform-specific attribution informativeness, and additive surplus as its baseline.

D Welfare cost of uniform contracts

  • Creators with high attribution SNR are offered royalties that are too low, while low-SNR creators bear excessive attribution-noise risk; distinctive creators can be harmed most by this mismatch.Distinctive creators are also described as having strong standalone careers and complementary value in generation.Source: D Welfare cost of uniform contracts · p. 19
  • Uniform contracts create excess risk-premium costs for participating creators and lose exiting creators’ entire contribution, with joint exit causing superlinear losses when creators are complements.The proof relates the superlinear loss to distinctive creators’ concentration among those exiting.Source: D Welfare cost of uniform contracts · pp. 19–20

E Multi-platform Bertrand competition

  • The multi-platform extension formalizes Bertrand equilibrium, the attribution moat, and welfare comparisons across K ≥2 platforms.It extends creator-specific informativeness across platforms.Source: E Multi-platform Bertrand competition · p. 20

E.1 Setting and surplus decomposition

  • The platform setting assumes exclusive licensing, heterogeneous informativeness across platforms, and additive separability of each platform’s surplus; centered Gaussian noise is assumed where indicated.These are the stated conditions C1–C4 for the multi-platform analysis.Source: §E.1 · p. 20

E.2 Bertrand equilibrium

In the Bertrand equilibrium, platform-specific royalty incentives are independent of rivals, while creator allocation and platform rents depend on relative surpluses. The appendix also characterizes the attribution moat, welfare comparisons, royalty invariance, and the boundary posed by complementary creators.

E.2 Bertrand equilibrium

  • Each platform’s optimal royalty is independent of the number of platforms and rivals’ offers; competition affects rents, not the royalty incentive slope.Creators join the platform offering the highest maximized surplus if it meets their effective outside option, and otherwise exit.Source: §E.2 · pp. 20–21

E.3 Vertical differentiation rent and the moat

  • A trailing platform earns no attribution-investment rent until it overtakes the leader; under heterogeneous demand, a leader’s advantage extends the flip dead zone, while a trailer’s advantage shrinks it.The dead zone is the range in which a trailer’s attribution improvement raises the creator’s outside option but does not make the trailer the winner.Source: §E.3 · p. 21

E.4 Welfare comparisons

  • Under centered noise and homogeneous demand, the competitive allocation minimizes total welfare loss among allocations satisfying C1–C3; expected marginal welfare gains from entry diminish as platforms are added.For informativeness drawn uniformly from [0, 1], the marginal gain from adding the (K+1)-th platform is proportional to 1/((K+1)(K+2)).Source: §E.4 · p. 22

E.5 Royalty invariance under moral hazard

  • With endogenous quality choice under Q1–Q4 and C1–C3, optimal royalties, equilibrium effort, and per-creator welfare depend on a platform’s own primitives, not platform count or creators’ outside options.This competition invariance holds if and only if the surplus function decomposes across creators.Source: §E.5 · p. 22

E.6 Tension with non-additive utility

  • Additive separability is a tractable baseline that conflicts with the non-additive utility driving superlinear selective-exit losses; general equilibrium corrections for complementarities remain unquantified.The headline moat result is stated to be robust to weak complementarities.Source: §E.6 · p. 22

F Discussion of modeling choices

The discussion identifies modeling simplifications, while the appendix reports per-creator attribution diagnostics and a welfare-based investment allocation.

F Discussion of modeling choices

  • The model assumes independent attribution errors and one-shot licensing; correlated errors and renegotiation with learning are extensions, while joint contract–investment optimization awaits empirical calibration of attribution costs.Mean-variance preferences support tractable linear-contract analysis; the qualitative increase in royalty optimality with SNR also holds under CARA and CRRA expected utility with regularity conditions.Source: F Discussion of modeling choices · pp. 22–23

G.1 Detailed empirical setup

  • The audio experiment fine-tunes Stable Audio Open on 475 tracks across 22 artist catalogs and evaluates 200 prompts using matched full and leave-one-catalog-out models.The ground-truth attribution target is model-internal loss change at catalog level, rather than listener-judged resemblance [42] [48] [37].Source: §G.1 · p. 23

G.2 Per-creator measurement diagnostics

  • Spearman ρ = 0.51 (p = 0.015) shows the methods agree on which creators are detectable, although their absolute scales differ by orders of magnitude.At T = 200, most per-creator bootstrap intervals span zero for both methods.Source: Figure 4 · p. 24, §G.2 · p. 24
  • Pearson r = 0.69 (p < 0.001) links catalog size with ground-truth signal variance; most creators fall in the precision range with the largest marginal welfare gains from better attribution.The water-filling rule directs most of an improvement budget to this steep-slope region.Source: §G.2 · p. 24, Figure 5 · p. 25

G.3 Optimal attribution-investment portfolio

  • The water-filling rule allocates budget to creator j only when marginal welfare gain M_j exceeds the budget shadow price; under square-root technology, shares follow a square law.The allocation maximizes total welfare-loss reduction.Source: §G.3 · p. 25
  • Classical receives 75.1% of the pilot attribution budget despite having the smallest Var(a∗), because its low SNR places it on the steepest part of the welfare-loss curve.The other pilot shares are 19.3% for pop and 5.5% for reggae.Source: §G.3 · p. 26

G.4 Multi-platform Bertrand competition between attribution methods

The section compares creator surplus under competing attribution methods, then extends the framework to bargaining and dynamic participation. It characterizes how contract terms, exit, and continued licensing respond to attribution quality and creator circumstances.

G.4 Multi-platform Bertrand competition between attribution methods

  • EKFAC wins 14 creators and D-TRAK wins 8; their surplus rankings diverge for roughly one-third of creators, where the trailing platform’s private return is zero.These cases are the dead zones of Proposition 2.Source: §G.4 · p. 26
  • Dead-zone widths span four orders of magnitude, from ∼10−11 to ∼10−5, with a median around 10−7.Most creators have nearly indifferent surplus across methods, while a small minority drives the population-level dead-zone width.Source: §G.4 · p. 26, Figure 6 · p. 27

H Bargaining with large rights holders

  • Existing music-industry payment models occupy the pure-fixed-fee or pure-royalty boundaries, while attribution-aligned AI training licenses can use interior contracts.The contract space is θj = (fj, ρj).Source: §G.4 · p. 26
  • For irreplaceable creators, Nash bargaining gives the creator a share βj of incremental joint surplus, while the fixed fee implements the split and the royalty rate minimizes the risk premium.Noisy attribution shifts both parties toward fixed fees; creators with higher bargaining power capture more of the efficiency gain from better attribution.Source: H Bargaining with large rights holders · p. 27

I.1 Setup

  • The dynamic extension models creators who exit when realized payments fall below their outside option and may re-enter in a later period.The platform may add a retention rent to the static payment rule’s fixed fee.Source: I Dynamic participation · p. 27, §I.1 · p. 27

I.2 Static exit probability and the α-cancellation

  • Theorem 3 characterizes static exit probability, the platform’s Bellman value from an active creator, and steady-state participation under the dynamic extension.The steady-state participation expression depends on re-entry probability and static exit probability.Source: §I.2 · p. 28

I.3 Two-regime characterization

The dynamic model divides optimal retention policy into thin- and thick-margin regimes. It also links attribution informativeness and payment volatility to creator exit and participation.

I.3 Two-regime characterization

  • The platform optimally pays a positive retention rent in the thick-margin regime, while the thin-margin regime has zero rent and exits at the static rate.The thick-margin condition is µnet > µcrit; otherwise the optimum is s∗ = 0.Source: §I.3 · p. 28
  • Higher attribution informativeness reduces static exit, and higher signal variance also reduces exit through the payment buffer-to-standard-deviation ratio.The passage attributes the latter result to the risk premium growing quadratically in signal variance while payment standard deviation grows linearly.Source: §I.3 · p. 29
  • Improving informativeness raises steady-state participation and produces a marginal welfare gain.The platform’s retention rent is bounded above by the creator’s net margin.Source: §I.3 · pp. 28–29
  • For music platforms, losing distinctive long-tail catalogs can reduce stylistic diversity, and the steady-state participation of these catalogs is the operative variable for long-run generation quality.The passage predicts a persistent participation gap for the most valuable cohort under uniform compensation.Source: §I.3 · p. 29

J Synthetic data value chains

The section extends attribution to value created when a platform turns generated music into synthetic training data and downstream products. It argues that single-stage attribution misses some creator value and that full-chain measurement adds noise.

J Synthetic data value chains

  • The platform’s value chain can run from licensing original music through model training and synthetic-data generation to downstream value extraction.Downstream uses include fine-tuning specialized models, third-party licensing, and reducing dependence on future human-created music licensing.Source: J Synthetic data value chains · p. 29
  • A creator’s attributable value can include both the value vj of stage-one output and their marginal contribution to downstream synthetic-data value.The passage denotes the latter contribution as ∆vsyn_j.Source: J Synthetic data value chains · p. 29
  • Payment based only on vj systematically underpays creators whose data is valuable for synthetic generation.Synthetic data generated in a creator’s style may continue competing in the market without further licensing, compounding displacement over time.Source: J Synthetic data value chains · p. 29
  • Full-chain attribution compounds measurement noise, strengthening the case for fixed-fee components or hybrid payments that reflect future synthetic-generation value.The passage says the full-chain attribution signal has lower effective SNR than single-stage attribution and notes that pure per-output royalties can understate total value extraction.Source: J Synthetic data value chains · p. 29
Loading 2607.00641v1…