- 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
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.
| Check | Result |
|---|---|
| Full-data differences, mean / across-query SD () | -0.006 / 0.185 |
| Leave-out differences, median mean / median across-query SD over five catalogs () | 0.006 / 0.173 |
| Median calibrated SNR on the same five catalogs, EKFAC / D-TRAK | 0.0850 / 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 , 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 group | Mean catalog-removal effect () | Mean full-data seed difference () |
|---|---|---|
| Related outputs | 2.074 | -0.018 |
| Unrelated outputs | 2.124 | -0.003 |
After subtracting each group's full-data baseline, the related-minus-unrelated difference is . 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.
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.
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, 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.