acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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 Attribution-Aware compensation in generative-music markets, where each creator's payment depends on an attribution signal 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 only the most precise signal in the market captures those gains. Empirically, 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.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Data attribution, generative music, creator compensation, contract theory, data markets, attribution uncertainty.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_mxuCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper attempts to formalize the problem of profit distribution regarding how royalties should be paid to music rights holders whose work is used to train generative AI, by combining training data attribution with economic theory. The main result is that royalty-based compensation is preferable when contribution estimation is accurate, while fixed-fee licensing is preferable when estimation is inaccurate. However, because current estimation methods are highly noisy, the theory implies that fixed-fee-dominant contracts are appropriate in practice.

Critical strengths
  • The subject addressed in this paper has become increasingly important in recent years. At the same time, as far as the reviewer understands, there is still no established method for how to translate the results obtained from training data attribution into practical social implementation. In this regard, the introduction of economic theory to connect machine learning results to real-world applications is a meaningful contribution.
Critical weaknesses

The starting point of this paper is interesting. However, as the reviewer read further, the following points raised doubts about the paper's actual contribution to the field.

  • The theoretical contribution feels somewhat overstated.

    • The paper presents numerous theorems and propositions, but most of them appear to follow relatively directly from existing contract theory, or from the assumptions and objective functions set by the authors. Furthermore, the paper does not sufficiently discuss whether these assumptions themselves (such as the distributional premises regarding contribution and attribution scores) are adequately justified. Although the reviewer is not an expert in contract theory, these contributions do not seem to be well organized.
    • The central conclusion, namely that higher accuracy in contribution estimation favors royalties while lower accuracy favors fixed compensation, can be anticipated from existing contract theory, and its mathematical derivation is relatively simple. As a result, it is unclear what non-trivial insight each result adds to existing theory, and it is difficult to see any novelty beyond "applying existing theory to music generation models."
    • Consequently, the reviewer has the impression that the paper derives many theorems and propositions from a simple mathematical model and then assigns them market-institutional interpretations, thereby making the theoretical contribution appear larger than it actually is.
  • The validation of the ground truth is insufficient.

    • The paper uses the loss difference after retraining with leave-one-catalog-out (LOCO) models as the ground truth for attribution estimation, but it lacks the validation needed to support the appropriateness of this choice.
    • In particular, in the audio generation scenario, the full model and the LOCO models use different random seeds. As a result, the observed loss differences may contain not only the effect of excluding the target catalog but also stochastic variation from retraining. The paper should verify the loss variation when the same data is trained with different seeds, as well as the impact on data unrelated to the excluded catalog. However, these effects are not adequately separated and validated. Therefore, treating the LOCO loss difference as a reliable ground truth remains questionable.
    • Without such validation, it is difficult to attribute the observed low attribution accuracy to the performance limitations of existing methods. This weakens the basis for the paper's central empirical claim, namely the conclusion that current estimation methods are too noisy and that fixed-fee contracts are therefore appropriate.
Questions for authors
  • Among the many theorems and propositions presented, which ones are non-trivial results that cannot be derived from existing contract theory? Which points do you consider to be particularly significant contributions to this field?
4 comments
AuthorsOpenReview ↗

W1 & Q1. Theoretical contribution and assumptions

W1.1. Which results contribute beyond the familiar informativeness principle?

We thank the reviewer for requesting a more explicit account. We respectfully clarify that the contribution goes beyond the qualitative intuition that noisy pay is costly. The paper determines how measured ML attribution should enter creator payments, quantifies the resulting alignment loss, and derives consequences for catalog licensing. Three insights are particularly significant.

Attribution rankings alone do not determine a payment rule. Theorem 1 makes the royalty coefficient depend on the signal's calibration and creator-specific precision. Proposition 1 quantifies the alignment value lost relative to perfect attribution. A benchmark must therefore measure how reliably each catalog's score tracks its contribution across outputs, beyond ordering catalogs within an output. In our audio experiment, CLAP has higher mean Spearman correlation than EKFAC, while EKFAC has higher median in-sample calibrated SNR. We measure the relevant quantities on 50 audio catalogs and 100 symbolic catalogs. At the illustrative risk aversion of α=2\alpha=2, the median fitted audio coefficients are approximately 0.011 for EKFAC and 0.0045 for D-TRAK: within the model, the median creator's pay responds only weakly to changes in calibrated attributed revenue.

Improving attribution can benefit the creator before it earns the improving platform any catalog rent. Proposition 2 establishes this under exclusive, whole-catalog licensing and the stated rate and fee conditions. With equal base catalog values and a runner-up offer that covers the outside option, an improving runner-up forces the leader to offer the creator better terms, yet earns no rent until it overtakes the leader. Theorem 2 uses standard Bertrand reasoning after deriving each platform's available surplus from its attribution precision and the creator's payment risk. Applying the model to measured precision assigns 35 of the 50 catalogs to the EKFAC platform and 15 to D-TRAK. This illustrates how the preferred platform differs across creators under the model.

Changing the payment rule cannot recover information absent from the signal. Appendix K bounds the alignment attainable by any finite-variance payment based on the available signals and revenue. Under its joint-Gaussian conditions, a linear payment attains the bound. Outside those conditions, nonlinear payments may exploit information that a linear royalty misses. The result therefore identifies both the limit imposed by the signal and the scope for improving the contract's form.

Section 2 compares our question with contribution-based allocation [2], valuation-based pricing [3], and music attribution for royalties [4]. Section 3.2 explicitly credits the classical informativeness principle [1]. The contribution is to connect measured attribution quality to contract design and licensing outcomes through these characterizations and catalog-level measurements.

W1.2. Are the distributional assumptions needed and justified?

We thank the reviewer for raising this concern. We clarify that the main royalty formula does not require Gaussian scores, Gaussian noise, or classical additive error. Given the stated mean-variance preferences and alignment criterion, Theorem 1 uses finite second moments and positive signal variance, with period revenue known when the contract is written.

Classical additive error gives the SNR shortcut in Section 3.1. Gaussian assumptions enter selected examples and extensions. Our measurements show why raw-score calibration needs checking. EKFAC and D-TRAK scores differ in scale by roughly a thousandfold, and 19 of 50 EKFAC catalogs have negative estimated slopes. We report calibration scale and sign rather than presuming unit calibration of raw scores.

The alignment objective is also explicit: Equation (3) and Lemma 1 separately analyze financial profit alone. The empirical measurements estimate attribution quality. They do not establish creators' preferences or platforms' objectives.

AuthorsOpenReview ↗

W2. Validation of the LOCO contribution reference

W2.1. Why use leave-one-catalog-out loss differences as the contribution reference?

We thank the reviewer for asking about the contribution reference. We clarify that LOCO measures the effect of excluding the rights holder's entire catalog, matching the unit of the licensing decision. Appendix G.1 evaluates the full and leave-out models on the same full-model outputs, so the comparison holds the evaluation material fixed.

This defines the benchmark's contribution target in model-loss units. It does not measure listener judgment or revenue. The seed checks below address how sensitive that reference and the resulting attribution measurements are to training randomness.

W2.2. Could training-seed variation explain the observed audio loss differences?

We thank the reviewer for requesting this validation. In the original audio experiment, the full model used seed 42 and each leave-out model used seed 42 plus the catalog index (Table 3). The loss differences can therefore include training variation, which query-bootstrap intervals do not capture.

To address this concern, we repeated full-data training at seeds 43 and 44 and retrained five selected leave-out datasets at seed 42, using the original training configuration. We evaluated all models on the same 200 outputs and kept the attribution scores and loss settings fixed. We compared losses between fits trained on identical data and recomputed calibrated SNR using the matched-seed removal effects.

CheckResult
Full-data differences, mean / across-query SD (×10−3\times10^{-3})-0.006 / 0.185
Leave-out differences, median mean / median across-query SD over five catalogs (×10−3\times10^{-3})0.006 / 0.173
Median calibrated SNR on the same five catalogs, EKFAC / D-TRAK0.0850 / 0.0182

The paired loss differences measure training sensitivity relative to the original removal effects. Comparing calibrated SNR under the original and matched-seed references tests whether the low-coefficient finding persists on the same five catalogs. At α=2\alpha=2, calibrated SNR below 0.25 corresponds to a royalty coefficient below 0.1.

W2.3. What is the effect on data unrelated to the excluded catalog?

We thank the reviewer for requesting this comparison. We classify outputs using the recorded catalog of origin of each evaluation prompt. Outputs whose prompts originate from the excluded catalog are classified as related; outputs whose prompts originate from other catalogs are classified as unrelated. We use the seed-42 full and leave-out models for the removal comparison. Within each catalog, we average loss differences separately for the two groups, then average equally across the five catalogs. We compute the full-data seed baseline on the same groups, averaging the seed-43 and seed-44 differences from seed 42.

Output groupMean catalog-removal effect (×10−3\times10^{-3})Mean full-data seed difference (×10−3\times10^{-3})
Related outputs2.074-0.018
Unrelated outputs2.124-0.003

After subtracting each group's full-data baseline, the related-minus-unrelated difference is −0.035×10−3-0.035\times10^{-3}. Related outputs do not show a larger mean removal effect in this comparison.


References

[1] Bengt Holmstrom. Moral Hazard and Observability. The Bell Journal of Economics, 1979.

[2] Jiachen T. Wang, Zhun Deng, Hiroaki Chiba-Okabe, Boaz Barak, and Weijie J. Su. An Economic Solution to Copyright Challenges of Generative AI. 2024.

[3] Luyang Zhang, Cathy Jiao, Beibei Li, and Chenyan Xiong. Fairshare Data Pricing via Data Valuation for Large Language Models. NeurIPS, 2025.

[4] Junwei Deng, Xirui Jiang, Shiyuan Zhang, Shichang Zhang, Himabindu Lakkaraju, Ruijiang Gao, Chris Donahue, and Jiaqi W. Ma. Computational Copyright: Towards A Royalty Model for Music Generative AI. 2025.

Thank you for your responses and the additional experiments.

  • I think narrowing the claims has improved the overstatement issue. However, although the manuscript has been revised, I still find the current version a bit difficult to read.
  • I appreciate the additional empirical evidence, but I feel it is still unclear whether the low attribution accuracy comes from the methods or from the ground truth. In the W2.3 results, the removal effect is nearly the same for outputs related to the excluded catalog and for unrelated outputs (2.074 vs. 2.124, ×1e-3). This suggests that the LOCO loss differences may mainly capture a uniform effect of reduced training data, rather than the catalog-specific contribution. If so, the low SNR may reflect a limitation of the ground truth rather than poor performance of the attribution methods.

Please let me know if I have misunderstood this point. though I should not judge the value of the paper on this point alone, I believe the significance of its contribution depends relatively heavily on this result.

AuthorsOpenReview ↗

Response to the additional question of Reviewer mxuC

We thank the reviewer for noting that the narrower claims address the overstatement concern. We also appreciate the comment on readability and will keep improving the writing.

On the question of the ground truth, we respectfully clarify that our additional checks indicate the LOCO loss differences are catalog-specific and that a uniform effect of reduced training data does not explain the low SNR. The W2.3 averages are similar, but that does not mean removing different catalogs has the same effect on a given output.

First, the removal effect on an output depends on which catalog is removed, so it is not uniform. Of the 200 outputs, 23 come from prompts that originate in the five catalogs of our seed check. Using the original models, the catalog of origin has the largest removal effect of the five on 22 of these outputs. The EKFAC and D-TRAK scores put it first on 7 and 8, where chance would give about 5. Prompt origin is only a proxy for influence, but the removal effects agree with it far more often than the scores do.

Second, these differences between catalogs are reproducible across training seeds. We subtract each output's mean removal effect across the five catalogs, which leaves the catalog-specific part. For each catalog, we correlate this part between the original and matched-seed models across the 200 outputs. The median correlation is 0.991.

Third, the SNR stays low when the uniform part is removed, so a uniform effect does not explain it. On the original 50-catalog benchmark, we subtract each output's mean removal effect across the 50 catalogs and recompute calibrated SNR with the scores unchanged. The median moves from 0.0232 to 0.0257 for EKFAC and from 0.0091 to 0.0098 for D-TRAK. This subtraction also removes any real contribution that catalogs share, so we treat it as a sensitivity check.

These checks do not establish that the LOCO reference is error-free. They indicate that its catalog-specific part is reproducible and that the attribution scores track it poorly.

Reviewer_MemsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In this paper the authors study how to move towards an attribution-based system for paying creators in the context of AI music platforms. In their framework musicians license catalogs to platforms in an exclusive manner. The platforms offer a mix of a fixed rate payment and a royalty based payment. Musicians require their average payment to be higher than a threshold value that they could get by outsourcing through other channels and are adverse to risks, preferring fixed rate payments. At the same time it is proven (in Lemma 1) that also companies prefer fixed rate payments since they maximise overall profit. The authors than study what happens if we assume that attributing the work of the authors should be valued as much as having profit (for reputation or for regulations). Here the authors prove that in that case the welfare-optimal royalty is given by I/αI / \alpha, where II is the informativeness (which is related to the quality of the attribution estimators, better its SNR, higher the informativeness) and α\alpha the risk adversion of the creator. It is shown then in Section 3 that at the level of the market 1. under this system, the creator receives the greater of the outside option and the second-highest surplus offering platform, and the winning platform keeps the remainder 2. a platform improving its attribution system, can indirectly make its rival improve the rent of its musicians (attribution moat). The authors simulate the scenario in an experimental section where they finetune a diffusion (waveform) music generation and a midi model both with a set of 50 catalogs and 50 versions removing a artists in each (in order to compute an attribution "ground truth" value) and they compare 4 attribution models on them looking at the economic repercursions. All benchmarked models result in very high noise levels so they offer little gain in improving the joint value.

Critical strengths
  1. The method tries to solve a really pressing problem, where many artists are reluctant to share their catalog with AI platforms such SUNO given low incentives for royalties. This paper proposes in a quantitative way an economic mechanism were the artists are attributed more fairly (even, as shown in Proposition 2, when they don't have to change their platform), combining the fields of data attribution and payment desing.
  2. Good empirical evaluation (given the assumptions) across two different tasks (waveform audio generation and midi generation) showcasing the unsuitability of current attribution models. While the non-existence of good attribution methods may hinder the adoption of attribution-based Still raising awareness about this alternative payment policy and pushing the community to improve on attribution methods is a positive factor.
Critical weaknesses
  1. Principal weakness of the paper is that, although apparently well structured and without grammatical flaws, it's reading pace and flow is extremely low, bordering non comprehensibility when reading it without the assistance of an AI model. This highly suggested that the model has been generated with an LLM in a heavy manner. This kinda goes against what the authors claim in the AI USE STATEMENT, where they claim a minor usage. The main problem is that many phrases are very dry and their contextualisation is difficult to get. I will show different examples:

    • At line 323, when one first reads this thinks, why should the creators report their scores? Aren’t the scores produced by the platform? Then after the proposition one sees that this should not happen. This paragraph should be opened assuming this would be a hypotetical scenario, not what really happens.
    • At line 319: "the largest gains go to the creators measured least precisely today", specifying "today" is completely out of context for this phrase. Looks like an AI hallucination
    • Lines 299 to 301: This phrase is difficult to understand in this way. I would say: “Improving its attribution without overtaking the leader does not increase total joint surplus, but forces the winning platform to offer better terms, transferring some of its profit to the creator.”
    • Line 248: "Attribution precision determines the royalty rate; the creator’s outside option sets the required utility." these are typical AI style phrases
    • At line 282: The authors should reference the creator utility as Uj(θj)U_j(\theta_j) and should reference the platform payoff as defined in Appendix E.1, p. 23, lines 1210-1211, assumption (C0). It is very confusing to find these terms without reference, especially since the platform payoff in the previous section is defined as in Eq. 3 not considering the alignment term.
    • Line 253: Not clear phrase. What is the transfer level?
    • In synthesis all part 3.3 is highly condensed with all the required definitions and explanations relegated to the appendix. Since many concepts are introduced in the appendix (like the definition for the platform payoff in prev. point), it is very difficult to read altogether.

    Another indicative clue is that Table 1, that should describe music payment models, sits near the conclusions and is not reference anywhere in the paper. This is a clear AI use artifact and the authors have not revised the paper much (it would have been obvious to move reference it).

    Nevertheless I haven't had much difficulty comprehending the first three pages. This would suggest that AI was used more intensively starting with the methodological section.

  2. A second weakness is the apparent lack of novelty in the presented analysis. While I agree that the settings in which the economic analysis is applied is a novel one, and that the authors combine the data-attribution field with the payment design field, many of the reported analyses could work verbatim in other scenarios. For example under similar assumptions, Proposition 3 applies almost verbatim to employees of a certain company and their compensation. Its mathematical conclusion does not depend on the AI music market scenario. A more substantive contribution here could come from demonstrating that these conditions hold in music AI, measuring the resulting losses, or deriving a consequence that depends on its particular attribution or licensing structure. Another example is the Theorem 2, where the reasoning follows exactly a second-price Bertrand auction.

  3. The methodological framework is based on different assumptions that could not be true in practice. The most important one is that the platforms should value both making profit and giving attribution (the whole analysis is based on this), but it's not clear why that should be the case if platforms value profit in first case (it could be because of external enforcement). Another one is that creators and the platform agree about how informative attribution scores are (a creator may distrust these estimates and evaluate royalties as not acceptable).

Questions for authors
  • Line 290: It says that competition changes the fixed fee. Is this because the royalties are always selected using the formula of Theorem 1 so they are not changed arbitrarily? If the total value considers the rent and the attribution, it would imply that better attribution is not related to final payment to the creator. But this contradicts the fact that better attribution increases the royalties value. So in the end still the authors would be payed more. Can you please explain better?
  • How frequently do the empirically calibrated contracts produce negative unconstrained fixed fees, and how much would enforcing the full contract constraints change the welfare-optimal royalty coefficients?
  • In Proposition 1: maybe using small rr for Pearson coefficient would be better since RR already is used for the risk premium previously
  • Line 250: Why write this form when we could just use ρj∗E[vj]\rho_j^*E[v_j]?
  • What economic or normative justification supports the particular covariance-based alignment objective in Equation (4), and how sensitive are the optimal royalty rates to alternative formulations of social welfare?
  • Why profit here is Π\Pi capitalised while it was not before?
  • Eq 3: This quantity should be defined as the financial profit, since it is used later in the definition of the joint surplus
  • Line 215: I don’t know if to put the delta formalism here since it’s not used anywhere in the paper. Also it is not really modelled. Additionally the authors rely on the assumption that improved quality of the AI model automatically implies people listening more to AI models, which is not necessarily the case.
  • Line 164: Why additive here? Shouldn’t be something related to diversity?
5 comments
AuthorsOpenReview ↗

Weakness 1. Readability and AI-use disclosure

We thank the reviewer for the detailed reading and the specific suggestions for improving clarity. We respectfully disagree that the cited wording and organization issues show that the manuscript was heavily generated by an LLM, or that our disclosure was inaccurate. The authors wrote, reviewed, verified, and revised the manuscript over multiple rounds. A hard-to-follow passage, the word "today," or a missing table reference does not show that the text was generated by AI.

We further would like to point out that our AI Use Statement does not describe our AI use as minor. The submission discloses that AI tools were used to "polish the writing," help identify potentially missing related work, assist with drafting parts of the related-work section and selected experimental subsections, and proofread mathematical arguments after the authors completed the proofs. It also states that the authors "reviewed and rewrote all AI-assisted text" and take responsibility for the final content. This discloses drafting assistance, not only grammar correction. We nevertheless appreciate the reviewer's specific suggestions and have adopted them.

  • We thank the reviewer for the wording suggestions. We have removed "today" (line 319), clarified the surplus transfer (lines 299 to 301), rewritten the royalty-rate and outside-option sentence (line 248), and replaced "transfer level" with a direct explanation of the fixed fee (line 253).

  • We thank the reviewer for asking why creators would report scores (line 323). We clarify that Proposition 4 examines a hypothetical reporting arrangement, not the baseline procedure. The revised paragraph now says this first, then explains why payment should not rely on unverified creator reports.

  • We thank the reviewer for requesting clearer payoff definitions (line 282 and Section 3.3). We clarify that creator utility is defined in Equation (2) and financial profit in Equation (3). The main text now also defines the competition payoff as financial profit plus the alignment term of assumption (C0), and joint surplus as this payoff plus creator utility.

  • We thank the reviewer for pointing out the missing reference to Table 1. We respectfully disagree that this is a "clear AI use artifact." An omitted cross-reference also occurs in human-written manuscripts, so neither the omission nor the table's location shows AI authorship. We have added the reference in Section 3.1, where the contract forms are defined.

AuthorsOpenReview ↗

Weakness 2. Novelty beyond classical contract theory

We thank the reviewer for raising this concern. We respectfully clarify that our core contribution is to combine music data attribution with contract theory, a combination the reviewer also recognizes as novel. The paper provides a framework that turns measured attribution quality into creator contracts, and derives guidance for platforms, creators, attribution researchers, and regulators (Section 3.3). Within this framework, the paper measures the losses that current attribution precision implies. It also derives consequences that depend on the attribution signal and on catalog licensing. These address two of the three directions the reviewer suggests.

Specifically, Our framework connects measured attribution quality to creator payments and catalog licensing.

  • We derive how strongly a creator's pay should respond to a noisy attribution score (Theorem 1). Prior work connects music attribution to royalty allocation [4], and contribution-based revenue allocation [2] and valuation-based data pricing [3] use contribution estimates to determine payments. These works take the estimate as given. Theorem 1 sets each creator's royalty by the score's informativeness and the creator's risk aversion. Section 3.2 credits the classical informativeness principle [1] on which it builds.

  • We show that attribution methods used for payment must be evaluated by calibrated precision, not by ranking (Proposition 1). Proposition 1 ties the welfare cost of attribution noise to how closely each creator's score tracks their contribution across queries, and the royalty also depends on the score's scale. Rank correlation compares catalogs within a query, so it measures neither, and the two can favor different methods. Additionally, our empirical results support this distinction. CLAP has higher mean Spearman correlation than EKFAC (0.046 versus 0.007), yet EKFAC has higher median in-sample calibrated SNR (0.023 versus 0.016). Thus, selecting by ranking alone would favor CLAP, while selecting by this calibrated-precision measure would favor EKFAC.

  • We apply the framework to generative music models and find that current attribution precision favors fixed-fee-dominant contracts, so attribution should not yet drive large royalties. In the 50-catalog audio experiment at α=2\alpha=2, the median royalty coefficient is about 0.011 for EKFAC and 0.0045 for D-TRAK, and these methods close about 9% of the alignment gap with in-sample calibration. On held-out queries the gains are 5% to 7%, with intervals that include zero (see our response to Reviewer NV5E, Q3).

Theorem 2 shows that attribution quality decides which AI platform wins a creator's catalog. A platform that measures a creator's contribution more precisely can offer that creator a better contract. The proof uses standard Bertrand reasoning. The new result is Proposition 2. A trailing platform that improves its attribution raises the creator's terms before it gains anything itself. In our audio data, a platform using EKFAC wins 35 of 50 catalogs and one using D-TRAK wins 15, with in-sample calibration.

Proposition 3 is general but it answers a design question central to this market. We agree that its mathematics does not depend on music, and we do not claim it as a new technique. Its role is to justify per-creator contracts. Music licensing largely pays creators on uniform terms, such as blanket licenses and pro-rata streaming rates (Table 1). Proposition 3 shows the cost of keeping uniform terms when attribution could personalize them. Creators with the best outside options leave first, and when their catalogs contribute complementary styles, the loss is at least the sum of their individual contributions. This is why the framework sets each creator's royalty from that creator's own attribution precision. Generality does not weaken its value. Its conditions involve outside options and catalog complementarities, which our data do not observe, so we do not test them empirically.

We have additionally included brief explanations after Theorem 2 (lines 317 to 321) and Proposition 3 (lines 345 to 346) in the revised PDF that state their scope and economic roles.

AuthorsOpenReview ↗

Weakness 3. Modeling assumptions

W3.1. Why would a profit-maximizing platform value attribution?

We thank the reviewer for raising this concern. We respectfully clarify that only part of the analysis relies on the alignment objective. In the effort extension, a profit-maximizing platform benefits from more precise attribution.

Better attribution raises a profit-maximizing platform's profit when creators choose effort. Lemma 1 holds each creator's contribution fixed, and in that case a profit-only platform pays a fixed fee. Appendix B.3 lets creators choose effort that raises both their contribution and their attribution score. A platform maximizing only financial profit then pays a positive royalty to induce that effort (Proposition 5). Noisy attribution makes royalties risky, and creators must be compensated for that risk. More precise attribution lowers this cost, so the platform pays a higher royalty and earns strictly higher profit (Appendix B.3, lines 998 to 1002). The size of this effect depends on creators' effort costs, which we do not measure, so our empirical conclusions concern the static setting.

The alignment objective is needed only for part of the analysis. Lemma 1, the effort extension, Propositions 3 and 4, and the Section 4 measurements of attribution quality (correlations, calibration scale, and calibrated SNR) use only financial profit, creator utility, or the attribution scores themselves. Theorem 1, Proposition 1, and the competition results add the alignment objective. It represents a platform, auditor, or regulator that values pay tracking contribution (Section 3.2), including through external enforcement as the reviewer suggests. Appendix E.1 varies its weight. The royalty scales with that weight, and a zero weight recovers the profit-only baseline.

We have additionally labeled Equation (3) as expected financial profit (line 220), defined the competition payoff in the main text (lines 307 to 309), and stated the profit result above in Appendix B.3 (lines 998 to 1002).

W3.2. What if creators distrust the platform's informativeness estimate?

We thank the reviewer for raising this concern. We acknowledge that the shared assessment is a simplifying assumption, and that creators who distrust the scores may find royalties less acceptable. Suppose a creator expects lower scores than the platform forecasts, or sees the scores as noisier. Under Equation (2), that creator values any royalty less. The platform then has to raise the fixed fee, lower the royalty, or lose the creator. Distrust of this kind pushes contracts toward fixed fees, the same direction as noisy attribution.

Distrust also gives platforms a reason to make attribution checkable. Section 3.3 suggests disclosing the platform's measured informativeness (lines 370 to 372). The scores can also be computed by an independent auditor rather than by the platform (line 366). We leave a formal model of disagreement to future work.

Reviewer_NV5EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper models creator compensation in generative music at the catalog level under noisy data attribution. Creators have mean-variance risk-averse utility, and the platform maximizes profit plus a payment-contribution alignment term. Under this setup, the optimal linear royalty rate depends on the catalog's attribution signal-to-noise ratio and the creator's risk aversion. The paper also analyzes uniform contracts, Bertrand competition across platforms, and score manipulation, and evaluates existing attribution methods against leave-one-catalog-out retraining on Stable Audio Open across 50 catalogs and a symbolic transformer across 100 catalogs.

Critical strengths
  1. Linking data attribution to creator compensation at the catalog level makes sense. Most attribution papers only check rank agreement across songs. Most data market papers assume attribution scores are exact. Adding creator risk aversion shows why noisy or uncalibrated scores hurt welfare through payment variance. So catalog-level informativeness and scale calibration are the right quantities to study.

  2. Building the leave-one-catalog-out ground truth across 51 audio models and 101 symbolic models is a big empirical effort. The diagnostics in Section 4 and Appendix G clearly show that raw attribution scores have very different scales across methods, and rank agreement does not match calibrated welfare gains.

Critical weaknesses
  1. A small royalty rate does not mean fixed fees dominate total payment. Total payment is the fixed fee plus the royalty payment, and the fixed fee depends on the creator's outside option, expected royalties, and risk compensation. If the fixed fee is zero, all payment still comes from royalties even when the royalty rate is tiny. Theorem 1 only shows lower payment sensitivity to attribution, not fixed-fee dominance without extra assumptions on payment levels.

  2. The audio ground truth mixes catalog removal with training seed noise. Table 3 shows seed 42 for the full model and seeds 43 to 92 for the leave-catalog-out models, whereas the symbolic experiment uses seed 42 for every run. Without same-seed comparisons or repeated training runs, we cannot tell how much of the poor audio attribution score comes from the estimators versus training randomness. Also, resampling queries for bootstrap intervals does not capture this training variance.

  3. The calibration results do not show reliable gains for a typical creator. Table 2 fits and tests the 50 creator slopes on the exact same 200 queries. In the held-out check in Appendix G.3 Table 7, the bootstrap intervals cross zero for all four non-random methods' total gains and their differences from Random. Every method has a negative median gain across creators, and a single catalog makes up about 90% to 113% of the net total gain.

Questions for authors
  1. A small royalty rate does not mean fixed fees make up most of a creator's expected pay. The fixed fee also depends on outside options, expected attributed revenue, and risk compensation. Does fixed-fee-dominant just mean a near-zero royalty rate, or a high fixed-fee share of expected pay? Please clarify this, and either report the implied fixed-fee share or revise the claim to describe low sensitivity to attribution.

  2. Table 3 shows that the symbolic runs fix seed 42, while the audio runs use seed 42 for the full model and seeds 43 to 92 for the leave-catalog-out models. Why were the audio seeds varied? Can you run a matched-seed check on a small subset of catalogs, or repeat the full-data training across a few seeds, to show how large training variance is relative to the catalog-removal effect?

  3. In Table 7, the held-out aggregate gains have bootstrap intervals that include zero, median creator gains are negative, and one catalog drives almost all the net gain. Given these results, what claims about creator-level calibration benefits still hold? Please add the held-out gains and their bootstrap intervals to Table 2, and adjust your main claims to match this evidence.

5 comments
AuthorsOpenReview ↗

Q1 & W1. Does "fixed-fee-dominant" just mean a near-zero royalty rate, or a high fixed-fee share of expected pay?

We thank the reviewer for this question. We clarify that "fixed-fee-dominant" means a royalty coefficient near zero (Lines 258–259). The fee's share of expected pay also depends on the outside option (Lines 259–261). At the median EKFAC coefficient of about 0.011 at α=2\alpha=2 (Lines 478–479), an extra dollar of calibrated attributed revenue raises pay by about 1.1 cents. For positive expected attributed revenue and a positive outside option, Equation (6) (Lines 264–268) gives the royalty share and therefore the fixed-fee share. Combining Equation (6) with the reported coefficients (Lines 478–480) gives the following thresholds:

  • The fee can be zero only if the outside option is at most about 1.1% of expected attributed revenue (first line of Equation (6), Line 265). Risk compensation can require a positive fee even below this level.
  • At that coefficient, an outside option of at least about 2.2% of expected attributed revenue guarantees a fixed-fee share of at least one half, and strictly more with positive risk compensation (Equation (6), Lines 265–267).
  • For the 48 of 50 EKFAC and 49 of 50 D-TRAK creators with coefficients below 0.1 (Lines 486–487), an outside option of at least 20% of expected attributed revenue guarantees a majority fixed-fee share.

These are conditions, not estimates of creators' outside options (Line 269). We have additionally made the coefficient definition explicit in Section 3.2 (Lines 258–269), added Equation (6) for payment shares, and shown that the fixed-fee share approaches one as informativeness vanishes under the stated assumptions (Lines 270–278). The empirical section now describes the median result as a royalty coefficient close to zero (Lines 417–418).

AuthorsOpenReview ↗

Q2 & W2. Why were the audio seeds varied, and can a matched-seed or repeated-training check quantify their effect?

We thank the reviewer for requesting this check. In the original audio experiment, the launcher set leave-out seeds to 42 plus the catalog index and the full-model seed to 42 (Table 3). Appendix G.1 now explains this assignment. We acknowledge that query-bootstrap intervals do not capture training variation.

To address this concern, we repeated full-data training at seeds 43 and 44 and retrained five selected leave-out datasets at seed 42, using the original training configuration. We evaluated all models on the same 200 outputs and kept the attribution scores and loss settings fixed. We compared losses between fits trained on identical data and recomputed calibrated SNR using the matched-seed removal effects.

CheckValues
Full-data differences, mean / across-query SD (×10−3\times10^{-3})-0.006 / 0.185
Leave-out differences, median mean / median across-query SD over five catalogs (×10−3\times10^{-3})0.006 / 0.173
Median calibrated SNR on five catalogs, EKFAC / D-TRAK0.0850 / 0.0182

In this example, both median SNRs remain below 0.25, corresponding to coefficients below 0.1 at α=2\alpha=2. It does not show that the actual seed changes are small or that the real five-catalog result will persist.

The symbolic setting already provides a comparison without the audio seed offset: all 101 fits use seed 42, yet median calibrated SNR is 0.017 for LoGra and 0.002 for TRAK.

AuthorsOpenReview ↗

Q3 & W3. What calibration benefits remain supported by the held-out results, and can their intervals be added to Table 2?

We thank the reviewer for this question. We acknowledge that the held-out results do not support creator-level calibration benefits, and the revised paper does not claim them. Two findings still hold.

  • Royalty coefficients are near zero. At α=2\alpha=2, the median fitted coefficients are 0.011 for EKFAC and 0.0045 for D-TRAK. The held-out results reinforce this: under Theorem 1, a calibration that does not carry over to new queries warrants a royalty near zero.
  • Scores must be calibrated before they set pay. Raw EKFAC and D-TRAK scores differ in scale by about a thousandfold.

We have added the held-out gains and intervals to Table 2, and the differences from Random to Table 7:

MethodIn-sample gainHeld-out gain [95% CI]Held-out minus Random [95% CI]
Random0.53%-0.63% [-1.9, -0.5]n/a
CLAP4.96%2.83% [-2.9, 7.8]3.46 [-1.5, 9.0]
Fingerprint5.80%3.55% [-1.8, 7.5]4.18 [-0.5, 8.7]
EKFAC9.36%5.08% [-5.4, 10.8]5.71 [-4.2, 12.0]
D-TRAK9.18%6.92% [-2.5, 15.5]7.55 [-1.3, 16.6]

Gain is the reduction in squared prediction error relative to a fitted constant, as a share of reference variance. Intervals are percentiles from a 5,000-replicate query bootstrap, with calibration refitted in each replicate. As the reviewer notes, every non-random interval and every difference from Random includes zero, the median catalog's held-out gain is negative under every method, and one catalog accounts for nearly all of each method's net gain.

We have revised the main claims accordingly (blue in the PDF):

  • Lines 419 to 426 state that held-out gains are not reliable for most creators, separately from the low royalty coefficients.
  • Lines 407 to 414 mark the SNR comparison as in-sample and note that EKFAC and D-TRAK swap order on held-out gain. Line 469 treats the CLAP comparison as a comparison of point estimates.
  • Line 475 states that the market calculations use in-sample calibration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.