pith. sign in

arxiv: 2605.24383 · v1 · pith:ZE43YGVEnew · submitted 2026-05-23 · 💻 cs.AI · cs.CY· cs.SE

A governance horizon for ethical-use constraints in open-weight AI models

Weiwei Xu , Hengzhi Ye , Haoran Ye , Kai Gao , Vladimir Filkov , Minghui Zhou This is my paper

Pith reviewed 2026-06-30 13:50 UTC · model grok-4.3

classification 💻 cs.AI cs.CYcs.SE
keywords ethical-use constraintsopen-weight modelsgovernance horizonderivation lineagesmetadata disclosureAI policy designorphan componentssupply chain accountability
0
0 comments X

The pith

Ethical-use constraints on open-weight AI models lose traceability after roughly seven generations of derivation, defining a governance horizon.

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

The authors audit millions of model repositories to test if voluntary ethical constraints propagate through open-weight AI lineages. They find that evidence of restrictions decays exponentially with a half-life of 1.31 derivation steps, so that beyond seven generations eighty percent of models lack enough public information for governance decisions. This limit arises because many lineages contain orphan components with no upstream intent to inherit, and because disclosure relies on repeated metadata statements rather than automatic propagation. The study shows that changing policy design, such as requiring declarations instead of relying on inheritance, can extend the reach more effectively than increasing enforcement alone. Comparison to other open ecosystems indicates the problem is tied to how AI models derive from one another.

Core claim

Disclosure-based governance for ethical-use constraints in open-weight AI has a shallow reach determined by the topology of model derivation. Restriction evidence decays with a half-life of 1.31 steps, and the governance horizon is reached at seven downstream generations where at least 80% of descendants lack sufficient public evidence. Orphan lineage components without inheritable upstream intent remain undecidable under inheritance-only policies, and these create bottlenecks that cannot be recovered by enforcement. Policy interventions that introduce mandatory declarations can shift the horizon at moderate enforcement levels, unlike inheritance-only designs that require near-complete enfor

What carries the argument

The governance horizon, the depth at which 80% of descendant models lack sufficient public evidence for governance determination due to decaying restriction evidence across derivation steps.

If this is right

  • Near-complete enforcement is needed for inheritance-only designs to extend the governance horizon.
  • Mandatory-declaration designs shift the horizon at moderate enforcement rates.
  • Orphan components in lineages remain undecidable regardless of enforcement under inheritance rules.
  • Unresolved upstream nodes create direct downstream undecidability that inheritance cannot fix.
  • The collapse is specific to open-weight derivation topology, as shown by comparison with ecosystems using explicit declarations.

Where Pith is reading between the lines

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

  • Platforms might explore embedding governance signals directly into model artifacts to bypass metadata decay.
  • Regulators could consider standards requiring provenance tracking that survives multiple derivations.
  • Similar horizons may limit governance in other rapidly forking open-source domains without structural changes.
  • Voluntary disclosure alone cannot support long-term accountability in AI supply chains.

Load-bearing premise

The metadata on Hugging Face repositories accurately indicates the presence or absence of ethical-use constraints and that derivation relationships between models can be reliably determined from that metadata.

What would settle it

An independent audit that examines model files directly and finds that ethical constraints are preserved and documented in a majority of models beyond seven generations, contrary to the metadata-based decay pattern.

read the original abstract

Ethical constraints on open-weight AI models are both a reflection of societal concerns and a foundation for AI governance policy. They are expected to propagate to downstream derivatives while implemented as voluntary metadata disclosures that must be restated at each generation of reuse. We audit 2,142,823 model repositories on Hugging Face Hub to test whether this disclosure-based governance infrastructure can sustain traceability across deep model lineages. Restriction evidence decays with a half-life of 1.31 derivation steps ($R^2$=0.98), and beyond seven downstream generations at least 80% of descendant models lack sufficient public evidence for a governance determination, a depth boundary we formalize as the governance horizon. Platform-level interventions to restore missing licence metadata reveal that policy design (not enforcement alone) is the binding factor: inheritance-only designs require near-complete enforcement to move the horizon, whereas a mandatory-declaration design that explicitly resolves orphan lineage components shifts the horizon already at moderate enforcement. The structural bottleneck is lineages with no inheritable upstream intent: such orphan components remain undecidable under any inheritance-only policy regardless of enforcement rate, and unresolved upstream nodes additionally create direct downstream undecidability bottlenecks that inheritance rules alone cannot recover. Comparison with PyPI, where governance signals are carried by explicit machine-readable declarations, corroborates that the collapse is topology-specific to open-weight derivation rather than inherent to open ecosystems. These results establish that disclosure-based governance has a shallow, structurally determined reach in open-weight AI, and that achieving deep supply-chain accountability requires provenance mechanisms propagating governance signals through derivation itself.

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

3 major / 2 minor

Summary. The paper audits 2,142,823 Hugging Face model repositories to examine propagation of ethical-use constraint disclosures across open-weight model lineages. It reports an exponential decay in restriction evidence with a half-life of 1.31 derivation steps (R²=0.98) and defines a governance horizon at seven downstream generations, beyond which at least 80% of descendant models lack sufficient public evidence for governance determinations. The authors model policy interventions (inheritance-only vs. mandatory-declaration designs), identify orphan lineage components as structural bottlenecks, and contrast results with PyPI to argue the collapse is specific to derivation topology rather than open ecosystems generally.

Significance. If the lineage reconstruction and metadata fidelity assumptions hold, the work supplies a quantitative, large-scale empirical demonstration that disclosure-based governance has limited reach in open-weight AI supply chains, with the horizon determined by topology rather than enforcement intensity alone. The high R² fit, explicit comparison to PyPI, and distinction between orphan and inheritable components provide falsifiable structure for policy analysis. The scale of the audit (over 2M repositories) is a notable strength.

major comments (3)
  1. [Methods (lineage inference procedure)] The half-life of 1.31 steps and the 80%-undecidable threshold at depth 7 are computed from inferred derivation steps; the Methods section provides no ground-truth validation, error-rate estimation, or sensitivity analysis for the metadata-based lineage mapping (base_model fields, card text, etc.). If false-positive or false-negative rates exceed ~15%, both the exponential parameter and the governance-horizon claim become unreliable.
  2. [Policy intervention modeling] The abstract states that 'platform-level interventions to restore missing licence metadata' were simulated, yet no section details the exact enforcement-rate model, how orphan components are handled in the simulation, or the parameter values used to generate the 'near-complete enforcement' vs. 'moderate enforcement' outcomes.
  3. [Results (decay analysis)] Table or figure reporting the decay fit (R²=0.98) does not indicate whether the fit was performed on binned counts, per-lineage averages, or after excluding models with missing base_model metadata; this choice directly affects the reported half-life and the depth-7 threshold.
minor comments (2)
  1. [Abstract and §4] The exact operational definition of 'sufficient public evidence for a governance determination' should be stated explicitly, including any thresholds on licence text length or presence of specific keywords.
  2. [Data collection] Confirm whether the 2,142,823 figure is the total scraped repositories or the final filtered set used for lineage reconstruction.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We address each major point below and will revise the manuscript accordingly to improve methodological transparency and clarity.

read point-by-point responses

  1. Referee: [Methods (lineage inference procedure)] The half-life of 1.31 steps and the 80%-undecidable threshold at depth 7 are computed from inferred derivation steps; the Methods section provides no ground-truth validation, error-rate estimation, or sensitivity analysis for the metadata-based lineage mapping (base_model fields, card text, etc.). If false-positive or false-negative rates exceed ~15%, both the exponential parameter and the governance-horizon claim become unreliable.

    Authors: We agree that ground-truth validation of the lineage mapping is not provided and would be ideal. At the scale of 2.1M repositories, no public ground-truth derivation graph exists, which limits direct error-rate computation. We will add a sensitivity analysis in the revised Methods section that perturbs the inference rules (e.g., requiring explicit base_model vs. card-text matching) and reports the resulting range for the half-life and governance horizon. This will quantify robustness without claiming external validation. revision: yes

  2. Referee: [Policy intervention modeling] The abstract states that 'platform-level interventions to restore missing licence metadata' were simulated, yet no section details the exact enforcement-rate model, how orphan components are handled in the simulation, or the parameter values used to generate the 'near-complete enforcement' vs. 'moderate enforcement' outcomes.

    Authors: The simulation parameters and orphan-handling logic are described in the policy-intervention subsection of Results, but the description is too terse. We will expand this subsection to state the exact enforcement rates modeled (e.g., 95% vs. 60%), the rule for propagating declarations through orphans, and the precise definition of 'near-complete' versus 'moderate' enforcement used to generate the reported horizon shifts. revision: yes

  3. Referee: [Results (decay analysis)] Table or figure reporting the decay fit (R²=0.98) does not indicate whether the fit was performed on binned counts, per-lineage averages, or after excluding models with missing base_model metadata; this choice directly affects the reported half-life and the depth-7 threshold.

    Authors: The exponential fit was performed on binned aggregate counts of restriction evidence by derivation depth, after excluding repositories lacking any base_model metadata. We will revise the figure caption and the corresponding Methods paragraph to state this explicitly and to note the number of models excluded by the metadata filter. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical decay and horizon are direct statistical summaries of audited metadata

full rationale

The paper's central results—the 1.31-step half-life and seven-generation governance horizon—are obtained by counting restriction evidence across inferred derivation depths in 2.1 million Hugging Face repositories and fitting an exponential decay model to those observed proportions. The half-life is the fitted parameter of that empirical curve (R²=0.98), and the horizon is the depth at which the observed or fitted fraction without sufficient evidence reaches 80%. Neither quantity is defined in terms of itself, nor does any equation reduce the output to a prior self-citation or ansatz. Lineage reconstruction from base_model fields and card text is an input assumption whose accuracy is external to the derivation; the paper does not invoke uniqueness theorems, self-citations, or prior author work to force the functional form. The PyPI comparison is likewise an independent external benchmark. The derivation chain therefore remains self-contained against the supplied data.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical fit of decay and the assumption about metadata accuracy and lineage inference from platform data.

free parameters (1)
  • half-life of 1.31 derivation steps = 1.31
    Fitted to the observed decay of restriction evidence across derivation steps from the audit data.
axioms (1)
  • domain assumption Derivation relationships can be reliably inferred from Hugging Face repository metadata
    The audit of 2,142,823 repositories relies on this to track lineages and measure decay.

pith-pipeline@v0.9.1-grok · 5822 in / 1102 out tokens · 33319 ms · 2026-06-30T13:50:49.821844+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

52 extracted references · 18 canonical work pages · 3 internal anchors

  1. [1]

    On the Opportunities and Risks of Foundation Models

    Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., Arx, S., Bern- stein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  2. [2]

    LLaMA: Open and Efficient Foundation Language Models

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozi` ere, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

    In: International Conference on Learning Representations (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022). https://openreview.net/forum?id=nZeVKeeFYf9 19

    2022

  4. [4]

    ACM Transactions on Software Engineering and Methodology (2025)

    Stalnaker, T., Wintersgill, N., Chaparro, O., Heymann, L.A., Di Penta, M., Ger- man, D.M., Poshyvanyk, D.: An empirical analysis of machine learning model and dataset documentation, supply chain, and licensing challenges on hugging face. ACM Transactions on Software Engineering and Methodology (2025)

    2025

  5. [5]

    In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pp

    Goddard, C., Siriwardhana, S., Ehghaghi, M., Meyers, L., Karpukhin, V., Bene- dict, B., McQuade, M., Solawetz, J.: Arcee’s mergekit: A toolkit for merging large language models. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pp. 477–485 (2024)

    2024

  6. [6]

    Advances in neural information processing systems36, 7093–7115 (2023)

    Yadav, P., Tam, D., Choshen, L., Raffel, C.A., Bansal, M.: Ties-merging: Resolv- ing interference when merging models. Advances in neural information processing systems36, 7093–7115 (2023)

    2023

  7. [7]

    arXiv preprint arXiv:2402.05979 (2024)

    McDuff, D., Korjakow, T., Cambo, S., Benjamin, J.J., Lee, J., Jernite, Y., Ferran- dis, C.M., Gokaslan, A., Tarkowski, A., Lindley, J., et al.: On the standardization of behavioral use clauses and their adoption for responsible licensing of ai. arXiv preprint arXiv:2402.05979 (2024)

    work page arXiv 2024

  8. [8]

    https://www.licenses.ai/blog/2022/8/18/ naming-convention-of-responsible-ai-licenses

    Contractor, D., Ferrandis, C.M., Lee, J., McDuff, D.: From RAIL to Open RAIL: Topologies of RAIL Licenses. https://www.licenses.ai/blog/2022/8/18/ naming-convention-of-responsible-ai-licenses. Responsible AI Licenses blog post, accessed 2026-03-14 (2022)

    2022

  9. [9]

    Accessed 2026- 03-14 (2024)

    Meta Platforms: Llama 3.1 Community License Agreement. Accessed 2026- 03-14 (2024). https://github.com/meta-llama/llama-models/blob/main/models/ llama3 1/LICENSE

    2026

  10. [10]

    In: Proceed- ings of the Conference on Fairness, Accountability, and Transparency, pp

    Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I.D., Gebru, T.: Model cards for model reporting. In: Proceed- ings of the Conference on Fairness, Accountability, and Transparency, pp. 220–229 (2019)

    2019

  11. [11]

    everyone wants to do the model work, not the data work

    Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P., Aroyo, L.M.: “everyone wants to do the model work, not the data work”: Data cascades in high-stakes AI. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, pp. 1–15 (2021)

    2021

  12. [12]

    Nature Machine Intelligence6(8), 975–987 (2024)

    Longpre, S., Mahari, R., Chen, A., Obeng-Marnu, N., Sileo, D., Brannon, W., Muennighoff, N., Khazam, N., Kabbara, J., Perisetla, K.,et al.: A large-scale audit of dataset licensing and attribution in ai. Nature Machine Intelligence6(8), 975–987 (2024)

    2024

  13. [13]

    arXiv preprint arXiv:2303.15772 (2023) 20

    Bommasani, R., Soylu, D., Liao, T.I., Creel, K.A., Liang, P.: Ecosystem graphs: The social footprint of foundation models. arXiv preprint arXiv:2303.15772 (2023) 20

    work page arXiv 2023

  14. [14]

    Science381(6654), 158–161 (2023)

    Samuelson, P.: Generative ai meets copyright. Science381(6654), 158–161 (2023)

    2023

  15. [15]

    Science380(6650), 1110–1111 (2023)

    Epstein, Z., Hertzmann, A., Investigators of Human Creativity, Akten, M., Farid, H., Fjeld, J., Frank, M.R., Groh, M., Herman, L., Leach, N.,et al.: Art and the science of generative ai. Science380(6650), 1110–1111 (2023)

    2023

  16. [16]

    arXiv preprint arXiv:2310.12941 (2023)

    Bommasani, R., Klyman, K., Longpre, S., Kapoor, S., Maslej, N., Xiong, B., Zhang, D., Liang, P.: The foundation model transparency index. arXiv preprint arXiv:2310.12941 (2023)

    work page arXiv 2023

  17. [17]

    1: May 2024

    Bommasani, R., Klyman, K., Kapoor, S., Longpre, S., Xiong, B., Maslej, N., Liang, P.: The foundation model transparency index v1. 1: May 2024. arXiv preprint arXiv:2407.12929 (2024)

    work page arXiv 2024

  18. [18]

    Communications of the ACM64(12), 86–92 (2021)

    Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J.W., Wallach, H., Daum´ e III, H., Crawford, K.: Datasheets for datasets. Communications of the ACM64(12), 86–92 (2021)

    2021

  19. [19]

    Longpre, S., Mahari, R., Obeng-Marnu, N., Brannon, W., South, T., Gero, K., Pentland, S., Kabbara, J.: Data authenticity, consent, & provenance for ai are all broken: what will it take to fix them? arXiv preprint arXiv:2404.12691 (2024)

    work page arXiv 2024

  20. [20]

    arXiv preprint arXiv:2602.10758 (2026)

    Wang, B., Chen, Y., Shi, J., Li, M., Lyu, Y., Wu, Y., Lin, Y., Yang, Z.: Hidden licensing risks in the llmware ecosystem. arXiv preprint arXiv:2602.10758 (2026)

    work page arXiv 2026

  21. [21]

    In: Proceedings of the 21st International Conference on Mining Software Repositories, pp

    Jiang, W., Yasmin, J., Jones, J., Synovic, N., Kuo, J., Bielanski, N., Tian, Y., Thiruvathukal, G.K., Davis, J.C.: Peatmoss: A dataset and initial analysis of pre- trained models in open-source software. In: Proceedings of the 21st International Conference on Mining Software Repositories, pp. 431–443 (2024)

    2024

  22. [22]

    In: NeurIPS 2025 Workshop on Regulatable ML (2025)

    Oderinwale, H., Laufer, B., Kleinberg, J.: Anatomy of a machine learning ecosys- tem: 2 million models on hugging face. In: NeurIPS 2025 Workshop on Regulatable ML (2025)

    2025

  23. [23]

    Official Journal of the European Union, L 2024/1689

    European Parliament and Council of the European Union: Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence. Official Journal of the European Union, L 2024/1689. Accessed 2026-03-14 (2024). https://eur-lex.europa.eu/eli/ reg/2024/1689/oj

    2024

  24. [24]

    arXiv preprint arXiv:2107.03721 (2021)

    Veale, M., Borgesius, F.Z.: Demystifying the draft eu artificial intelligence act. arXiv preprint arXiv:2107.03721 (2021)

    work page arXiv 2021

  25. [25]

    Technical report, U.S

    National Telecommunications and Information Administration: Dual-use foun- dation models with widely available model weights report. Technical report, U.S. Department of Commerce (July 2024). https://www.ntia.gov/ programs-and-initiatives/artificial-intelligence/open-model-weights-report 21

    2024

  26. [26]

    2023 , number =

    National Institute of Standards and Technology: Artificial intelligence risk man- agement framework (AI RMF 1.0). Technical Report NIST AI 100-1, U.S. Department of Commerce (2023). https://doi.org/10.6028/NIST.AI.100-1

    work page doi:10.6028/nist.ai.100-1 2023

  27. [27]

    https://aipolicy.substack.com/p/ supply-chains-2

    Cen, S.H., Hopkins, A., Ilyas, A., Madry, A., Struckman, I., Videgaray, L.: AI supply chains (and why they matter). https://aipolicy.substack.com/p/ supply-chains-2. On AI Deployment series, Substack (2023)

    2023

  28. [28]

    Technical Report 44, OECD Publishing, Paris (2025)

    OECD, GPAI: Ai openness: A primer for policymakers. Technical Report 44, OECD Publishing, Paris (2025). https://doi.org/10.1787/02f73362-en . https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/08/ ai-openness 958d292b/02f73362-en.pdf

    work page doi:10.1787/02f73362-en 2025

  29. [29]

    Law, Innovation and Technology16(2), 341–391 (2024)

    Gorwa, R., Veale, M.: Moderating model marketplaces: Platform governance puzzles for ai intermediaries. Law, Innovation and Technology16(2), 341–391 (2024)

    2024

  30. [30]

    closed: Emerging consensus and key questions for foundation ai model governance

    Bateman, J., Baer, D., Bell, S.A., Brown, G.O., Cu´ ellar, M.-F., Ganguli, D., Henderson, P., Kotila, B., Lessig, L., Lundblad, N.B., et al.: Beyond open vs. closed: Emerging consensus and key questions for foundation ai model governance. Technical report, Carnegie Endowment for International Peace (2024)

    2024

  31. [31]

    US Copyright Office (2023)

    Mahari, R., Longpre, S., Bollacker, K., Muennighoff, N., Khazam, N., Pentland, S.: Comment to US Copyright Office on Data Provenance and Copyright. US Copyright Office (2023). https://dspace.mit.edu/handle/1721.1/154171

    2023

  32. [32]

    In: Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pp

    Solaiman, I.: The gradient of generative ai release: Methods and considerations. In: Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pp. 111–122 (2023)

    2023

  33. [33]

    Accessed 2026-03-14 (2024)

    Open Source Initiative: The Open Source AI Definition, Version 1.0. Accessed 2026-03-14 (2024). https://opensource.org/ai/open-source-ai-definition

    2026

  34. [34]

    Xinyi Xu, Zhaoxuan Wu, Rui Qiao, Arun Verma, Yao Shu, Jingtan Wang, Xinyuan Niu, Zhenfeng He, Jiangwei Chen, Zijian Zhou, et al

    White, M., Haddad, I., Osborne, C., Liu, X.-Y.Y., Abdelmonsef, A., Varghese, S., Hors, A.L.: The model openness framework: Promoting completeness and openness for reproducibility, transparency, and usability in artificial intelligence. arXiv preprint arXiv:2403.13784 (2024)

    work page arXiv 2024

  35. [35]

    arXiv preprint arXiv:2311.09227 (2023)

    Seger, E., Dreksler, N., Moulange, R., Dardaman, E., Schuett, J., Wei, K., Winter, C., Arnold, M., ´O h ´Eigeartaigh, S., Korinek, A., Anderljung, M.: Open-sourcing highly capable foundation models: An evaluation of risks, benefits, and alternative methods for pursuing open-source objectives. arXiv preprint arXiv:2311.09227 (2023)

    work page arXiv 2023

  36. [36]

    SSRN preprint

    Widder, D.G., West, S., Whittaker, M.: Open (for business): Big tech, concen- trated power, and the political economy of open AI. SSRN preprint. Accepted to appear in Nature (2023). https://doi.org/10.2139/ssrn.4543807 . https://ssrn. 22 com/abstract=4543807

    work page doi:10.2139/ssrn.4543807 2023

  37. [37]

    Journal of Machine Learning Research24(400), 1–79 (2023)

    Henderson, P., Li, X., Jurafsky, D., Hashimoto, T., Lemley, M.A., Liang, P.: Foundation models and fair use. Journal of Machine Learning Research24(400), 1–79 (2023)

    2023

  38. [38]

    Lemley, M.A., Casey, B.: Fair learning. Tex. L. Rev.99, 743 (2020)

    2020

  39. [39]

    Seton Hall L

    Gervais, D.J.: AI derivatives: The application to the derivative work right to literary and artistic productions of AI machines. Seton Hall L. Rev.52, 1111 (2022)

    2022

  40. [40]

    Transactions of the Association for Computational Linguistics6, 587–604 (2018)

    Bender, E.M., Friedman, B.: Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics6, 587–604 (2018)

    2018

  41. [41]

    In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pp

    Pushkarna, M., Zaldivar, A., Kjartansson, O.: Data cards: Purposeful and trans- parent dataset documentation for responsible AI. In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pp. 1776–1826 (2022)

    2022

  42. [42]

    In: Proceedings of the 32nd IEEE/ACM International Conference on Program Comprehension, pp

    Pepe, F., Nardone, V., Mastropaolo, A., Bavota, G., Canfora, G., Di Penta, M.: How do hugging face models document datasets, bias, and licenses? an empiri- cal study. In: Proceedings of the 32nd IEEE/ACM International Conference on Program Comprehension, pp. 370–381 (2024)

    2024

  43. [43]

    In: The Twelfth Interna- tional Conference on Learning Representations (2024)

    Yang, X., Liang, V.W., Zou, J.Y.: Navigating dataset documentations in AI: A large-scale analysis of dataset cards on HuggingFace. In: The Twelfth Interna- tional Conference on Learning Representations (2024)

    2024

  44. [44]

    Holistic Evaluation of Language Models

    Liang, P., Bommasani, R., Lee, T., Tsipras, D., Soylu, D., Yasunaga, M., Zhang, Y., Narayanan, D., Wu, Y., Kumar, A., et al.: Holistic evaluation of language models. arXiv preprint arXiv:2211.09110 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  45. [45]

    In: Forty-second International Conference on Machine Learning Position Paper Track (2025)

    Duan, M., Du, M., Zhao, R., Wang, M., Wu, Y., Shadbolt, N., He, B.: Posi- tion: Current model licensing practices are dragging us into a quagmire of legal noncompliance. In: Forty-second International Conference on Machine Learning Position Paper Track (2025)

    2025

  46. [46]

    arXiv preprint arXiv:2602.08816 (2026)

    Jewitt, J., Rajbahadur, G.K., Li, H., Adams, B., Hassan, A.E.: Permissive- washing in the open ai supply chain: A large-scale audit of license integrity. arXiv preprint arXiv:2602.08816 (2026)

    work page arXiv 2026

  47. [47]

    arXiv preprint arXiv:2509.09873 (2025) 23

    Jewitt, J., Li, H., Adams, B., Rajbahadur, G.K., Hassan, A.E.: From hugging face to github: Tracing license drift in the open-source ai ecosystem. arXiv preprint arXiv:2509.09873 (2025) 23

    work page arXiv 2025

  48. [48]

    arXiv preprint arXiv:2309.08133 (2023)

    Lee, K., Cooper, A.F., Grimmelmann, J.: Talkin’ ’bout AI generation: Copyright and the generative-AI supply chain. arXiv preprint arXiv:2309.08133 (2023)

    work page arXiv 2023

  49. [49]

    Journal of machine learning research22(164), 1–20 (2021)

    Pineau, J., Vincent-Lamarre, P., Sinha, K., Larivi` ere, V., Beygelzimer, A., d’Alch´ e-Buc, F., Fox, E., Larochelle, H.: Improving reproducibility in machine learning research (a report from the neurips 2019 reproducibility program). Journal of machine learning research22(164), 1–20 (2021)

    2019

  50. [50]

    In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pp

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M.,et al.: Transformers: State-of-the-art natu- ral language processing. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pp. 38–45 (2020)

    2020

  51. [51]

    Biometrika70(1), 41–55 (1983)

    Rosenbaum, P.R., Rubin, D.B.: The central role of the propensity score in observational studies for causal effects. Biometrika70(1), 41–55 (1983)

    1983

  52. [52]

    Robins, J.M., Rotnitzky, A., Zhao, L.P.: Estimation of regression coefficients when some regressors are not always observed. Journal of the American statistical Association89(427), 846–866 (1994) 24 S1 Dependency extraction and annotation validation This section provides additional details on the validation of the dependency-extraction pipeline described ...

    1994