pith. sign in

arxiv: 2607.00641 · v1 · pith:J3P3AIAAnew · submitted 2026-07-01 · 💻 cs.CY · cs.LG

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 This is my paper

Pith reviewed 2026-07-02 06:26 UTC · model grok-4.3

classification 💻 cs.CY cs.LG
keywords generative musicdata attributioncreator compensationAI training datawelfare analysisroyalty contractsfixed-fee licensingmulti-platform competition
0
0 comments X

The pith

A market framework pays creators for training generative music AI based on attribution scores for their catalogs, with accuracy determining royalty versus fixed-fee contracts and welfare outcomes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a compensation framework for creators whose recordings are used to train generative music models. Payments depend on data-attribution scores applied to entire creator catalogs rather than single songs, and incorporate the signal-to-noise ratio of those scores. This produces a closed-form payment rule per creator while quantifying the welfare losses from inaccurate attribution for both creators and platforms. The optimal contract type shifts with attribution informativeness, and competitive dynamics limit welfare gains to the platform holding the most precise signal.

Core claim

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.

What carries the argument

The attribution-aware compensation mechanism that takes catalog-level data-attribution scores and their informativeness (signal-to-noise ratio) as direct inputs to the payment rule.

If this is right

  • Better attribution directly increases welfare for both creators and the platform.
  • Noisy attribution signals shift optimal payments toward fixed-fee licensing and lower overall welfare.
  • Under multi-platform competition, a platform captures gains from attribution improvements only when its signal is the most precise.
  • Empirical tests with acoustic and symbolic music models show that noisy signals reduce welfare and favor fixed-fee contracts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Platforms may invest in attribution technology to gain a competitive edge by becoming the most precise signal holder.
  • The same catalog-level attribution logic could apply to compensating creators in other generative domains such as images or text if analogous scores exist.
  • Creators might adjust their output strategies to maximize measured attribution if the payment rule is implemented.

Load-bearing premise

Attribution scores can be meaningfully computed for entire creator catalogs and their informativeness can be accurately estimated to serve as reliable inputs to the payment mechanism and welfare calculations.

What would settle it

An experiment or simulation in which switching to a higher-informativeness attribution method produces no measurable increase in creator or platform welfare relative to a lower-informativeness baseline.

Figures

Figures reproduced from arXiv: 2607.00641 by Beibei Li, Chris Donahue, Jiaqi W. Ma, Junwei Deng, Luyang Zhang, Xirui Jiang.

Figure 1
Figure 1. Figure 1: The attribution-aware compensation framework for generative music. Each creator’s [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Attribution measurement on two settings, audio diffusion (Stable Audio Open with EKFAC [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Framework prescription on the measured creator pool, [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-creator detail behind the main-text headline numbers. [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Scaling and marginal-welfare diagnostics. [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Bertrand competition between EKFAC and D-TRAK as two platforms. [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
read the original 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 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper develops a market framework for compensating creators in generative music AI markets. Each creator receives a payment based on a catalog-level data-attribution score whose informativeness (signal-to-noise ratio) enters the mechanism as a parameter. The framework produces a closed-form payment rule, quantifies welfare losses from inaccurate attribution for both creators and the platform, shows that the welfare-optimal contract switches between royalty-based and fixed-fee licensing depending on attribution informativeness, demonstrates direct welfare gains from better attribution, and establishes that under multi-platform competition a platform captures those gains only when its attribution signal is the most precise. The theoretical results are grounded by training acoustic and symbolic music models and measuring attribution informativeness against leave-one-catalog-out ground truth.

Significance. If the central derivation and empirical estimates hold, the work supplies a principled, parameter-light mechanism for attribution-aware compensation that directly ties attribution quality to contract design and welfare. The closed-form rule, the royalty-versus-fixed-fee threshold result, and the multi-platform competition finding are all falsifiable and policy-relevant. The leave-one-catalog-out empirical protocol for estimating the key SNR parameter is a concrete methodological contribution that can be replicated in other generative domains.

major comments (2)
  1. [§3 (payment rule) and §4 (welfare analysis)] The abstract and §4 claim that the payment rule and welfare expressions are closed-form and that the royalty/fixed-fee threshold depends on catalog-level SNR, yet the manuscript provides neither the explicit derivation steps nor the intermediate equations that map the primitives (catalog attribution scores and SNR) to the final payment and welfare formulas. Without these steps it is impossible to confirm that the welfare cost of noisy attribution follows directly rather than from additional functional-form assumptions.
  2. [Empirical validation section] Experiments section: the leave-one-catalog-out results are reported only as point estimates of attribution informativeness; no sample sizes, standard errors, or sensitivity checks on the SNR values are shown. Because the model predicts that even moderate changes in SNR shift the optimal contract and welfare numbers, the absence of uncertainty quantification makes it difficult to assess whether the empirical finding that "noisy attribution pushes payment toward fixed-fee licensing" is robust.
minor comments (3)
  1. [§2] The definition of catalog-level attribution score and its aggregation from song-level scores should be stated explicitly before the payment rule is introduced.
  2. [Figures 3–5] Figure captions for the welfare plots should include the exact SNR values used and the number of catalogs in each panel.
  3. [§1.2] The related-work discussion of prior compensation schemes should cite the specific functional forms used in those papers so readers can see precisely how the SNR parameter differentiates the present mechanism.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our theoretical derivations and empirical robustness. We address each major comment below and will incorporate the suggested additions in the revised manuscript.

read point-by-point responses

  1. Referee: [§3 (payment rule) and §4 (welfare analysis)] The abstract and §4 claim that the payment rule and welfare expressions are closed-form and that the royalty/fixed-fee threshold depends on catalog-level SNR, yet the manuscript provides neither the explicit derivation steps nor the intermediate equations that map the primitives (catalog attribution scores and SNR) to the final payment and welfare formulas. Without these steps it is impossible to confirm that the welfare cost of noisy attribution follows directly rather than from additional functional-form assumptions.

    Authors: We agree that the main text did not include sufficient intermediate steps for the closed-form derivations. In the revision we will expand §3 to show the full mapping from catalog attribution scores and SNR to the payment rule, including all intermediate equations, and then derive the welfare expressions in §4 directly from those primitives. This will confirm that the welfare cost of noisy attribution follows from the model without additional functional-form assumptions. revision: yes

  2. Referee: [Empirical validation section] Experiments section: the leave-one-catalog-out results are reported only as point estimates of attribution informativeness; no sample sizes, standard errors, or sensitivity checks on the SNR values are shown. Because the model predicts that even moderate changes in SNR shift the optimal contract and welfare numbers, the absence of uncertainty quantification makes it difficult to assess whether the empirical finding that "noisy attribution pushes payment toward fixed-fee licensing" is robust.

    Authors: We acknowledge that the empirical section lacked uncertainty quantification. In the revision we will report the sample sizes for the leave-one-catalog-out protocol, add standard errors to the attribution informativeness estimates, and include sensitivity checks that vary SNR around the point estimates to show the resulting shifts in optimal contract type and welfare values. This will allow readers to assess the robustness of the finding that noisy attribution favors fixed-fee licensing. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained with independent empirical inputs

full rationale

The framework derives closed-form payment rules, welfare costs, and optimal contract forms (royalty vs. fixed-fee) directly from primitives consisting of catalog-level attribution scores and their informativeness (signal-to-noise ratio) as exogenous inputs to the mechanism. These quantities are not defined in terms of the derived payments or welfare values. The informativeness parameter is measured separately via leave-one-catalog-out experiments on trained music models against ground-truth attribution, providing an external benchmark independent of the theoretical outputs. No equation reduces by construction to a fitted parameter renamed as a prediction, no self-citation is load-bearing for the central claims, and no ansatz or uniqueness result is smuggled in. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on standard contract theory assumptions and the empirical measurability of attribution quality; no new entities are postulated.

free parameters (1)
  • attribution informativeness (SNR)
    Serves as input parameter determining royalty vs fixed-fee contract type per creator catalog.
axioms (1)
  • domain assumption Economic agents (creators and platforms) maximize expected welfare under the derived payment rules.
    Invoked to determine optimal contract form and welfare costs of inaccurate attribution.

pith-pipeline@v0.9.1-grok · 5835 in / 1420 out tokens · 32782 ms · 2026-07-02T06:26:29.818141+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 11 canonical work pages · 2 internal anchors

  1. [1]

    Acemoglu, A

    D. Acemoglu, A. Makhdoumi, A. Malekian, and A. Ozdaglar. Too much data: Prices and inefficiencies in data markets.American Economic Journal: Microeconomics, 14(4):218–256, 2022

    2022

  2. [2]

    Agarwal, M

    A. Agarwal, M. Dahleh, and T. Sarkar. A marketplace for data: An algorithmic solution. In Proceedings of the 2019 ACM Conference on Economics and Computation, pages 701–726, 2019

    2019

  3. [3]

    MusicLM: Generating Music From Text

    A. Agostinelli, T. I. Denk, Z. Borsos, J. Engel, M. Verzetti, A. Caber, N. Zeghidour, and C. Frank. MusicLM: Generating music from text, 2023. arXiv preprint arXiv:2301.11325

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    G. A. Akerlof. The market for “lemons”: Quality uncertainty and the market mechanism.The Quarterly Journal of Economics, 84(3):488–500, 1970

    1970

  5. [5]

    J. Bae, W. Lin, J. Lorraine, and R. Grosse. Training data attribution via approximate unrolled differentiation. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  6. [6]

    Barnett, H

    J. Barnett, H. Flores Garcia, and B. Pardo. Exploring musical roots: Applying audio embed- dings to empower influence attribution for a generative music model, 2024. arXiv preprint arXiv:2401.14542

    work page arXiv 2024

  7. [7]

    Bergemann, A

    D. Bergemann, A. Bonatti, and A. Smolin. The design and price of information.American Economic Review, 108(1):1–48, 2018

    2018

  8. [8]

    Bogdanov, M

    D. Bogdanov, M. Won, P. Tovstogan, A. Porter, and X. Serra. The MTG-Jamendo dataset for automatic music tagging. InMachine Learning for Music Discovery Workshop, International Conference on Machine Learning (ICML), 2019

    2019

  9. [9]

    Caillon, B

    A. Caillon, B. McWilliams, C. Tarakajian, I. Simon, I. Manco, J. Engel, N. Constant, Y . Li, et al. Live music models. InAdvances in Neural Information Processing Systems (NeurIPS), Creative AI Track, 2025. arXiv preprint arXiv:2508.04651

    work page arXiv 2025

  10. [10]

    Y . Chen, H. Liu, and A. Karbasi. Selling data to a machine learner. InInternational Conference on Machine Learning (ICML), 2022. 10

    2022

  11. [11]

    What is your data worth to gpt? llm-scale data valuation with influence functions.arXiv preprint arXiv:2405.13954, 2024

    Sang Keun Choe, Hwijeen Ahn, Juhan Bae, Kewen Zhao, Minsoo Kang, Youngseog Chung, Adithya Pratapa, Willie Neiswanger, Emma Strubell, Teruko Mitamura, et al. What is your data worth to gpt? llm-scale data valuation with influence functions.arXiv preprint arXiv:2405.13954, 2024

    work page arXiv 2024

  12. [12]

    W. Choi, J. Koo, K. W. Cheuk, J. Serrà, M. A. Martinez-Ramirez, Y . Ikemiya, N. Murata, Y . Takida, W.-H. Liao, and Y . Mitsufuji. Large-scale training data attribution for music generative models via unlearning. InAdvances in Neural Information Processing Systems (NeurIPS), Creative AI Track, 2025

    2025

  13. [13]

    L. W. Cong, W. Wei, D. Xie, and L. Zhang. Data, ai, and economic growth. Nber working paper, NBER, 2022

    2022

  14. [14]

    Copet, F

    J. Copet, F. Kreuk, I. Gat, T. Remez, D. Kant, G. Synnaeve, Y . Adi, and A. Défossez. Simple and controllable music generation. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  15. [15]

    T. M. Cover and J. A. Thomas.Elements of Information Theory. Wiley-Interscience, 2nd edition, 2006

    2006

  16. [16]

    Deezer: AI-generated tracks now represent 44% of all new uploaded mu- sic

    Deezer. Deezer: AI-generated tracks now represent 44% of all new uploaded mu- sic. Deezer Newsroom, 2026. URL https://newsroom.deezer.com/2026/04/ai- generated-tracks-represent-44-of-new-uploaded-music/ . Published April 20, 2026

    2026

  17. [17]

    J. Deng, X. Jiang, S. Zhang, S. Zhang, H. Lakkaraju, R. Gao, C. Donahue, and J. Ma. Com- putational copyright: Towards a royalty model for music generative AI, 2023. arXiv preprint arXiv:2312.06646

    work page arXiv 2023

  18. [18]

    Jukebox: A Generative Model for Music

    P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and I. Sutskever. Jukebox: A generative model for music, 2020. arXiv preprint arXiv:2005.00341

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  19. [19]

    Dieleman, A

    S. Dieleman, A. van den Oord, and K. Simonyan. The challenge of realistic music generation: modelling raw audio at scale. InAdvances in Neural Information Processing Systems (NeurIPS), 2018

    2018

  20. [20]

    Melody transcription via generative pre- training

    Chris Donahue, John Thickstun, and Percy Liang. Melody transcription via generative pre- training. InISMIR, 2022

    2022

  21. [21]

    Hookpad aria: A copilot for songwriters.arXiv preprint arXiv:2502.08122, 2025

    Chris Donahue, Shih-Lun Wu, Yewon Kim, Dave Carlton, Ryan Miyakawa, and John Thickstun. Hookpad aria: A copilot for songwriters.arXiv preprint arXiv:2502.08122, 2025

    work page arXiv 2025

  22. [22]

    Engstrom, A

    L. Engstrom, A. Feldmann, and A. M ˛ adry. DsDm: Model-aware dataset selection with data- models. InInternational Conference on Machine Learning (ICML), 2024

    2024

  23. [23]

    Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI act)

    European Parliament and Council of the European Union. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI act). Technical report, European Parliament and Council of the European Union, 2024. Official Journal of the European Union

    2024

  24. [24]

    Evans, J

    Z. Evans, J. D. Parker, C. J. Carr, Z. Zukowski, J. Taylor, and J. Pons. Stable Audio Open, 2024. arXiv preprint arXiv:2407.14358

    work page arXiv 2024

  25. [25]

    Forsgren and H

    S. Forsgren and H. Martiros. Riffusion: Stable diffusion for real-time music generation, 2022. URLhttps://riffusion.com/about

    2022

  26. [26]

    Georgiev, J

    K. Georgiev, J. Vendrow, H. Salman, S. M. Park, and A. Madry. The journey, not the destination: How data guides diffusion models, 2023. arXiv preprint arXiv:2312.06205

    work page arXiv 2023

  27. [27]

    Ghorbani and J

    A. Ghorbani and J. Zou. Data Shapley: Equitable valuation of data for machine learning. In Proceedings of the 36th International Conference on Machine Learning, pages 2242–2251, 2019

    2019

  28. [28]

    Giordano, W

    R. Giordano, W. Stephenson, R. Liu, M. Jordan, and T. Broderick. A Swiss army infinitesimal jackknife. InProceedings of the 22nd International Conference on Artificial Intelligence and Statistics, pages 1139–1147, 2019

    2019

  29. [29]

    J. Gong, Y . Song, W. Zhao, S. Wang, S. Xu, J. Guo, and X. Yang. ACE-Step 1.5: Pushing the boundaries of open-source music generation, 2026. arXiv preprint arXiv:2602.00744

    work page arXiv 2026

  30. [31]

    Studying large language model generalization with influence functions.arXiv preprint arXiv:2308.03296, 2023

    R. Grosse, J. Bae, C. Anil, et al. Studying large language model generalization with influence functions, 2023. arXiv preprint arXiv:2308.03296

    work page arXiv 2023

  31. [32]

    Guruganesh, J

    G. Guruganesh, J. Schneider, and J. R. Wang. Contracts under moral hazard and adverse selection. InProceedings of the 22nd ACM Conference on Economics and Computation (EC), 2021

    2021

  32. [33]

    Hirshleifer

    J. Hirshleifer. The private and social value of information and the reward to inventive activity. The American Economic Review, 61(4):561–574, 1971

    1971

  33. [34]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  34. [35]

    Holmström

    B. Holmström. Moral hazard and observability.The Bell Journal of Economics, 10(1):74–91, 1979

    1979

  35. [36]

    Holmström and P

    B. Holmström and P. Milgrom. Aggregation and linearity in the provision of intertemporal incentives.Econometrica, 55(2):303–328, 1987

    1987

  36. [37]

    Ilyas, S

    A. Ilyas, S. M. Park, L. Engstrom, G. Leclerc, and A. Madry. Datamodels: Predicting predictions from training data. InProceedings of the 39th International Conference on Machine Learning, pages 9525–9587, 2022

    2022

  37. [38]

    Global music report 2025

    International Federation of the Phonographic Industry (IFPI). Global music report 2025. Technical report, IFPI, 2025

    2025

  38. [39]

    R. Jia, D. Dao, B. Wang, F. A. Hubis, N. Hynes, N. M. Gürel, B. Li, C. Zhang, D. Song, and C. Spanos. Towards efficient data valuation based on the Shapley value. InProceedings of the 22nd International Conference on Artificial Intelligence and Statistics, pages 1167–1176, 2019

    2019

  39. [40]

    C. I. Jones and C. Tonetti. Nonrivalry and the economics of data.American Economic Review, 110(9):2819–2858, 2020

    2020

  40. [41]

    Amuse: Human-ai collaborative songwriting with multimodal inspirations

    Yewon Kim, Sung-Ju Lee, and Chris Donahue. Amuse: Human-ai collaborative songwriting with multimodal inspirations. InProceedings of the 2025 CHI conference on human factors in computing systems, pages 1–28, 2025

    2025

  41. [42]

    P. W. Koh and P. Liang. Understanding black-box predictions via influence functions. In Proceedings of the 34th International Conference on Machine Learning, pages 1885–1894, 2017

    2017

  42. [43]

    Y . Kwon, E. Wu, K. Wu, and J. Zou. DataInf: Efficiently estimating data influence in LoRA- tuned LLMs and diffusion models. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  43. [44]

    Markowitz

    H. Markowitz. Portfolio selection.The Journal of Finance, 7(1):77–91, 1952

    1952

  44. [45]

    Maskin and J

    E. Maskin and J. Riley. Monopoly with incomplete information.The RAND Journal of Economics, 15(2):171–196, 1984

    1984

  45. [46]

    Bruno Kacper Mlodozeniec, Runa Eschenhagen, Juhan Bae, Alexander Immer, David Krueger, and Richard E. Turner. Influence functions for scalable data attribution in diffusion models. InInternational Conference on Learning Representations (ICLR), 2025. URL https:// openreview.net/forum?id=esYrEndGsr

    2025

  46. [47]

    Mussa and S

    M. Mussa and S. Rosen. Monopoly and product quality.Journal of Economic Theory, 18(2): 301–317, 1978

    1978

  47. [48]

    S. M. Park, K. Georgiev, A. Ilyas, G. Leclerc, and A. Madry. TRAK: Attributing model behavior at scale. InProceedings of the 40th International Conference on Machine Learning, pages 27074–27113, 2023

    2023

  48. [49]

    J. W. Pratt. Risk aversion in the small and in the large.Econometrica, 32(1/2):122–136, 1964

    1964

  49. [50]

    Pruthi, F

    G. Pruthi, F. Liu, S. Kale, and M. Sundararajan. Estimating training data influence by tracing gradient descent. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  50. [51]

    Major record labels file copyright in- fringement suits against AI music generators Suno and Udio

    Recording Industry Association of America (RIAA). Major record labels file copyright in- fringement suits against AI music generators Suno and Udio. Press release, Recording Industry Association of America (RIAA), 2024. Press release, June 2024. 12

    2024

  51. [52]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  52. [53]

    L. L. Scharf.Statistical Signal Processing: Detection, Estimation, and Time Series Analysis. Addison-Wesley, 1991

    1991

  53. [54]

    Serrà, E

    J. Serrà, E. Gómez, and P. Herrera. Audio cover song identification and similarity: Background, approaches, evaluation, and beyond. InAdvances in Music Information Retrieval, pages 307–332. Springer, 2010

    2010

  54. [55]

    C. E. Shannon. A mathematical theory of communication.The Bell System Technical Journal, 27(3):379–423, 1948

    1948

  55. [56]

    Silberling

    A. Silberling. AI music generator Suno hits 2M paid subscribers and $300M in annual recurring revenue. TechCrunch, 2026. URL https://techcrunch.com/2026/02/27/ai-music- generator-suno-hits-2-million-paid-subscribers-and-300m-in-annual- recurring-revenue/. Published February 27, 2026

    2026

  56. [57]

    Spijkervet and J

    J. Spijkervet and J. A. Burgoyne. Contrastive learning of musical representations. InProceedings of the 22nd International Society for Music Information Retrieval Conference (ISMIR), 2021

    2021

  57. [58]

    Anticipatory music transformer

    John Thickstun, David Hall, Chris Donahue, and Percy Liang. Anticipatory music transformer. TMLR, 2024

    2024

  58. [59]

    van den Oord, O

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  59. [60]

    A. L.-C. Wang. An industrial-strength audio search algorithm. InProceedings of the 4th International Conference on Music Information Retrieval (ISMIR), pages 7–13, 2003

    2003

  60. [61]

    Zhang, C

    L. Zhang, C. Jiao, B. Li, and C. Xiong. Fairshare data pricing via data valuation for large language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  61. [62]

    Zheng, T

    X. Zheng, T. Pang, C. Du, Q. Liu, J. Jiang, and M. Lin. Intriguing properties of data attribution on diffusion models. InInternational Conference on Learning Representations (ICLR), 2024. 13 A Gaussian special case and simplifying conditions The main results (Theorem 1, Propositions 1–3) hold for any noise structure through the general informativeness mea...

    2024

  62. [63]

    The excess isL uniform j =R uniform j −R ∗ j

    For creators who remain (j /∈ Sexit), the risk premium is Runiform j = αj 2 ¯ρ2 Var(vj), which is generically larger thanR ∗ j since¯ρ̸=ρ ∗ j . The excess isL uniform j =R uniform j −R ∗ j

  63. [64]

    Summing these two terms yields the uniform-contract welfare gap

    For creators who exit ( j∈ S exit), the platform loses their entire contribution, which by Claim 2 is superlinear. Summing these two terms yields the uniform-contract welfare gap. E Multi-platform Bertrand competition This appendix gives the formal results behind the attribution-moat analysis (Section 3.3): the proof of the multi-platform Bertrand equilib...

  64. [65]

    Budget shares follow a square law in marginal welfare gains

    Closed form under square-root technology.If gj(β) =γ j √β, the first-order approximation of the optimal allocation is βapprox j = (ϕjγj)2 P k(ϕkγk)2 ·B. Budget shares follow a square law in marginal welfare gains. 4.Comparative statics.∂β j/∂Sj <0,∂β j/∂αj <0,∂β j/∂Var(a ∗ j)>0. Proof. (i) L(S) is strictly convex for S >0 : L′′ = Λ(6S2 + 6S+ 2)/[S 3(1 +S)...