REVIEW 3 major objections 5 minor 28 references
The Mirage of Artificial Intelligence Terms of Use Restrictions
T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash
Pith's one-line read AI model terms of use are largely unenforceable because the weights and outputs they protect are not copyrightable, and copyright preemption blocks contract claims that try to create exclusive rights where none exist.
desk verdict A careful, genuinely useful synthesis of AI ToS enforceability; the mirage metaphor overstates a circuit-dependent conclusion, but the analysis is honest and should be published after qualification. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the combination of two doctrines. First, the human-authorship and functionality doctrines of copyright law remove model outputs and model weights from the set of protectable works, so there is no underlying exclusive right for a license to condition. Second, copyright preemption—both express preemption under § 301(a) and the recently revived conflict preemption from X Corp. v. Bright Data—converts that absence of copyright into a bar on state contract and tort claims that would give the same control over copying that copyright would have given had it existed. The paper's analysis pivots on how courts resolve the split between ProCD, which treats a two-party contract as an extra element that avoids preemption, and Genius/X Corp., which preempts contracts that protect copying of uncopyrightable material.
What would settle it
A federal court in a circuit that follows ProCD (for example, the Seventh Circuit) would falsify the paper's central claim if it enforced, with damages or an injunction, a terms-of-service clause forbidding a user from training a competing model on uncopyrightable AI outputs—without finding any copyright in the outputs or weights.
Extended reading notes
Core claim
The central discovery is that the enforceability of AI terms of use collapses because the two artifacts those terms purport to protect—model weights and model outputs—sit largely outside copyright law. Model outputs are generated by automated processes and accordingly fail the human-authorship requirement under Copyright Office guidance and recent case law; model weights are similarly machine-produced and are functional artifacts that § 102(b) excludes from protection. With no copyright in the underlying asset, the traditional mechanism for enforcing software licenses—conditioning a copyright grant on compliance—does not work, and contract claims that try to recreate copyright-like control over uncopyrightable material increasingly run into express and conflict preemption. The paper therefore concludes that anti-competitive restrictions are least likely to survive, that even narrow responsible-use terms may be preempted if Judge Alsup's conflict-preemption analysis takes hold, and that terms do not bind third parties who obtain outputs from the original user.
Load-bearing premise
The entire conclusion rests on courts adopting broad copyright preemption of contract claims over uncopyrightable outputs, so if ProCD's no-preemption approach dominates instead, the same clickwrap terms could be enforced even without copyright.
Editorial extensions
If this is right
- If the paper is right, "don't train a competing model on our outputs" clauses are the most vulnerable terms, because they are closest to reproducing copyright's exclusive rights and can also be attacked as copyright misuse or anticompetitive.
- Responsible-use restrictions will survive only when they are specific, focus on the purpose of use rather than copying, and target conduct with independent public-policy salience; vague or copying-centered restraints will face serious preemption challenges.
- Even enforceable terms would not bind third parties who receive model outputs from the original user, because contracts do not run with uncopyrightable information.
- Open-weight model licenses cannot rely on the copyleft enforcement mechanism of open-source software, since that mechanism depends on a copyright interest the model creator does not have.
- Policymakers relying on mandated license terms, such as watermark-preservation duties in the California AI Transparency Act, should expect weak or nonexistent private enforcement and should legislate prohibited uses directly.
Reading between the lines
- The authors leave implicit that the practical gap is largest for open-weight releases: closed API providers can still punish misuse through account revocation and access control even if their terms could not be enforced in court.
- A testable extension of the paper's logic is to track the first litigated cases enforcing an output anti-distillation clause: if those cases are withdrawn, settled quietly, or dismissed on preemption grounds, the "mirage" claim is strengthened.
- If the preemption reasoning generalizes, the same logic would undercut restrictive terms attached to other uncopyrightable public data, making it harder for platforms to use contract law to control scraping of facts and user-generated content.
- The paper implies that the debate over whether restrictive open-weight licenses are "truly open source" is somewhat beside the point: if the weights are uncopyrightable, the licenses are legally hollow regardless of their classification.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper argues that the restrictive terms of use attached by AI model creators to model weights and model outputs are largely unenforceable. It develops three claims: first, model weights and model outputs are generally not copyrightable, so there is no copyright-based licensing hook; second, state-law contract and tort claims built on those restrictions face serious copyright preemption problems under recent cases like Genius and X Corp. v. Bright Data, while DMCA, CFAA, and trespass-to-chattels theories offer little recourse; and third, anti-competitive restrictions are especially vulnerable to preemption, misuse, and antitrust doctrine, whereas narrow responsible-use restrictions have a better, though still uncertain, chance of survival. The paper recommends that policymakers rely on statutory regulation rather than private licensing terms as the primary mechanism for controlling harmful uses of AI.
Significance. If the paper's central conclusion is correct, it would substantially narrow the enforceable core of AI terms of use, with direct implications for the California AI Transparency Act, NTIA policy discussions, open-weight model licensing, and the broader debate about private ordering as a substitute for AI regulation. The paper is valuable for its systematic survey of current model-provider terms, its careful doctrinal mapping of the preemption landscape, and its willingness to engage with contrary authority such as ProCD and the Seventh Circuit's approach. Its strengths are institutional rather than formal: it is a well-hedged, internally consistent legal analysis that identifies concrete test cases and policy levers. It does not claim machine-checked proofs or quantitative results; its contribution is doctrinal synthesis and policy argument.
major comments (3)
- [Part III.C.1 and III.C.3] The article's headline conclusion that AI terms of use are a "mirage" is stated without a geographic qualifier in the abstract, the introduction, and the conclusion, but the paper's own doctrinal analysis shows that the outcome depends on an unresolved circuit split. In ProCD v. Zeidenberg jurisdictions, a two-party clickwrap agreement supplies an "extra element" that saves a state contract claim from section 301 preemption even when the underlying data is uncopyrightable, and the paper concedes at Part III.C.3 that "if Judge Alsup's conflict preemption analysis takes hold, even responsible AI terms may face strong preemption challenges." Because the article does not argue that ProCD is wrongly decided or inapplicable to AI terms of use, the enforceable core of AI terms is substantially larger if ProCD controls. This is load-bearing: the title, abstract, and policy recommendations should be conditioned on the preemption-friendly circuits, or the article should make an affirmative doctrinal argument for why AI terms should follow Genius and X Corp. rather than ProCD.
- [Introduction and Part I.D] The paper repeatedly relies on the observation that "no model creator has actually tried to enforce these terms with monetary penalties or injunctive relief" as circumstantial support for its legal conclusion (Introduction; Part I.D). This inference is equivocal. The absence of litigation is equally consistent with companies avoiding adverse precedent, settling quietly, relying on account suspensions and technical access control, or deciding that the reputational costs of suing researchers and users outweigh the benefits. The paper should either present the non-enforcement observation as a neutral motivating fact or identify what additional evidence would distinguish these explanations. As written, the passage risks treating the very doctrinal uncertainty the article identifies as if it were already a settled judicial rejection of enforcement.
- [Part II.D] The discussion of "system code" dismisses the copyright hook too quickly. The paper acknowledges that inference-system code is likely copyrightable, but argues that users can simply switch to interchangeable software and that reverse engineering is protected by fair use. That response does not address the common distribution channel in which the model creator distributes the weights and the inference code together in a single package, and the license is presented as a condition on downloading that package. A user who copies the creator's code, rather than independently written compatible code, may be bound by copyright conditions associated with that code even if the weights themselves are uncopyrightable. The article should explicitly state whether, and why, this route is unavailable for the specific terms surveyed in Part I, or should narrow the "no licensing hook" conclusion accordingly.
minor comments (5)
- [Abstract] The phrase "systematically assesses of the enforceability" contains a typo; it should be "systematically assesses the enforceability."
- [Introduction] The sentence ending "illegality amount machine learning practitioners" appears to be a typographical error; it should likely read "among machine learning practitioners."
- [Part III.C.3] There is a duplicated word: "If enforced, these terms would would:" should be "would:".
- [Part IV.A] Footnote 308 contains an incomplete citation: "cite statute" appears where the Defend Trade Secrets Act reverse-engineering provision should be cited, and the following sentence "Contra Bowers 320 F.3d at 1317" is too cryptic to be useful.
- [Appendix] The appendix references four figures with captions, but the images themselves are not visible in the manuscript text; please confirm that the figures render in the submission format.
Circularity Check
No circularity: the legal claims are anchored in external statutes, Copyright Office guidance, and litigated cases; the authors' self-citations are supporting authorities, not definitional inputs.
full rationale
This is a doctrinal legal analysis rather than a derivation with fitted parameters, equations, or a purported theorem, so the circularity patterns involving self-definition or fitted prediction do not apply. The central claim—that AI model weights and outputs are largely uncopyrightable and that copyright preemption may therefore invalidate many terms-of-use restrictions—rests on external sources: the Copyright Office's AI registration guidance, Thaler v. Perlmutter, Feist, 17 U.S.C. §§ 102(b) and 301(a), and litigated preemption cases (ProCD, Genius, X Corp. v. Bright Data). The authors cite their own prior articles (e.g., Lemley's 'How Generative AI Turns Copyright Law Upside Down' and 'Beyond Preemption,' and Henderson et al.'s empirical memorization studies), but those citations are not the load-bearing proof of any premise; each proposition they support is independently grounded in the cited regulations, judicial opinions, or empirical literature, and no 'uniqueness' theorem or author-specific construct is invoked to foreclose alternatives. The paper also expressly builds its conclusion on an unresolved circuit split rather than concealing it, stating that the enforceability outcome varies by circuit and that 'if Judge Alsup's conflict preemption analysis takes hold, even responsible AI terms may face strong preemption challenges.' That caveat shows the conclusion is not forced by the paper's own assumptions. A critic might argue the headline overgeneralizes because ProCD-following circuits enforce clickwrap restrictions over uncopyrightable data, but that is a vulnerability about doctrinal contingency and scope, not circularity: the argument does not reduce to its inputs by construction. No manuscript passage was found asserting a hidden limitation or omitted proof that would change this assessment.
Assumptions & free parameters
assumptions (5)
- domain assumption Copyright law's human authorship requirement and the Copyright Office's AI guidance govern model outputs and weights.
- domain assumption Copyright preemption under 17 U.S.C. Section 301 can reach state-law claims over uncopyrightable material within the subject matter of copyright.
- domain assumption Recent narrowing of the CFAA in Van Buren and rejection of trespass-to-chattels for public website access are stable and extend to AI services.
- domain assumption Contracts do not run with model outputs to third parties absent privity or a distinct property right.
- domain assumption Antitrust and copyright-misuse doctrines will police anticompetitive license terms.
Cite this review
Pith. "Pith review of The Mirage of Artificial Intelligence Terms of Use Restrictions." pith.science (2026). https://pith.science/paper/UKIG4O3Q
@misc{pith2026241207066,
author = {Pith},
title = {Pith review of: The Mirage of Artificial Intelligence Terms of Use Restrictions},
year = {2026},
howpublished = {\url{https://pith.science/paper/UKIG4O3Q}},
note = {Machine review of arXiv:2412.07066}
}
read the original abstract
Artificial intelligence (AI) model creators commonly attach restrictive terms of use to both their models and their outputs. These terms typically prohibit activities ranging from creating competing AI models to spreading disinformation. Often taken at face value, these terms are positioned by companies as key enforceable tools for preventing misuse, particularly in policy dialogs. But are these terms truly meaningful? There are myriad examples where these broad terms are regularly and repeatedly violated. Yet except for some account suspensions on platforms, no model creator has actually tried to enforce these terms with monetary penalties or injunctive relief. This is likely for good reason: we think that the legal enforceability of these licenses is questionable. This Article systematically assesses of the enforceability of AI model terms of use and offers three contributions. First, we pinpoint a key problem: the artifacts that they protect, namely model weights and model outputs, are largely not copyrightable, making it unclear whether there is even anything to be licensed. Second, we examine the problems this creates for other enforcement. Recent doctrinal trends in copyright preemption may further undermine state-law claims, while other legal frameworks like the DMCA and CFAA offer limited recourse. Anti-competitive provisions likely fare even worse than responsible use provisions. Third, we provide recommendations to policymakers. There are compelling reasons for many provisions to be unenforceable: they chill good faith research, constrain competition, and create quasi-copyright ownership where none should exist. There are, of course, downsides: model creators have fewer tools to prevent harmful misuse. But we think the better approach is for statutory provisions, not private fiat, to distinguish between good and bad uses of AI, restricting the latter.
Reference graph
Works this paper leans on
-
[1]
the owner of [a] copyright… the exclusive right[] to reproduce the copyrighted work
on 106 of the Copyright Act grants the “the owner of [a] copyright… the exclusive right[] to reproduce the copyrighted work.”182 Section 301(a) of the Copyright Act preempts state-law rights that “are equivalent to any of the exclusive rights within the general scope of copyright as specified by section 106.”183 Courts determining whether a state law clai...
work page 2024
-
[2]
197 Id. 198 Genius Media Group Inc. v. Google LLC, 19-CV-7279 (MKB), 2020 WL 5553639, at *8 (E.D.N.Y. Aug. 10, 2020), aff'd sub nom. ML Genius Holdings LLC v. Google LLC, 20-3113, 2022 WL 710744 (2d Cir. Mar. 10,
work page 2020
-
[3]
Preemption of Generative AI Terms Model creators looking to enforce their terms of use around models and their outputs will inevitably face both express and conflict preemption claims. The success of these claims will vary depending on the terms being enforced, so we will analyze each separately. Restrictions on competition. Contract-based constraints on ...
arXiv 1991
-
[7]
202 Harper & Row Publishers, Inc. v. Nation Enterprises, 723 F.2d 195, 200 (2d Cir. 1983), rev’d, on other grounds, 471 U.S. 539 (1985); see Guy A. Rub, A Less-Formalistic Copyright Preemption, supra note 180 at 337-38 (2017) (describing the case law in this area). 42 Terms of Service Woes for AI [9-Dec-24 Circuit, for instance, noted that “the scope of p...
work page 1985
-
[9]
(unfair competition and deceptive trade practices claims preempted as applied to an uncopyrightable database). 206 See, e.g., Forest Park Pictures v. USA Network, Inc., No. 11-2011 (2d Cir. 2012); Wrench LLC v. Taco Bell Corp., 256 F.3d 446 (6th Cir. 2001). 9-Dec-24] Terms of Service Woes for AI 43 stataute.207 A large group of legal scholars have emphasi...
work page 1973
-
[10]
little more than camouflage for an attempt to exercise control over the exploitation of a copyright,
(cleaned up). 219 Id. 220 Id. 221 Id. (quoting Sony Corp. of Am. v. Universal City Studios, Inc., 464 U.S. 417, 431 (1984) (emphasis added)). 9-Dec-24] Terms of Service Woes for AI 45 of creative works.222 He argues that X Corp.’s state law claims attempt to protect content that Congress intended to be free from restraint, including non-copyrightable mate...
work page 1984
-
[15]
promote violence, hatred or the suffering of others
(prohibiting use of ChatGPT to “promote violence, hatred or the suffering of others” or “Generating or promoting disinformation, misinformation, or false online engagement”). 246 As of October 2024 OpenAI still provides a version of ChatGPT without requiring a login. See Appendix for screenshot. 247 Foley v. Luster, 249 F.3d 1281, 1285 (11th Cir. 2001). 2...
work page 2024
-
[18]
(unpublished manuscript), https://arxiv.org/abs/2407.11969 273 Xiangyu Qi et al., Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!, in Proceedings of the Twelfth International Conference on Learning Representations (2024). 274 Yi Zeng et al., How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to ...
arXiv 2024
Show all 28 references
-
[19]
the ordinary course of its operation, prevents, restricts, or otherwise limits the exercise of a right of a copyright owner under this title
(unpublished manuscript), https://arxiv.org/abs/2401.06373 56 Terms of Service Woes for AI [9-Dec-24 data would copy whatever content was contained in the model without running afoul of the output restriction. In all, such a technical measure hardly seems like an access contro...
2024 arXiv
-
[21]
301 17 U.S.C. sec. 301; Goldstein v. California, 412 U.S. 546 (1973). 302 Id.; Sybersound Recs., Inc. v. UAV Corp., 517 F.3d 1137, 1152 (9th Cir
1973
-
[23]
No case has directly confronted the issue whether a literary work produced by a machine falls within the zone of copyright
(unfair competition and deceptive trade practices claims preempted as applied to an uncopyrightable database). No case has directly confronted the issue whether a literary work produced by a machine falls within the zone of copyright. As we saw above, the Copyright Office has ...
1974
-
[26]
We advocate for ‘standardized customization’ that can meet users’ needs and can be supported via tooling
(unpublished manuscript), https://arxiv.org/abs/2402.05979 (“We advocate for ‘standardized customization’ that can meet users’ needs and can be supported via tooling.”) 319 See, e.g., Arnav Gudibande et al., The False Promise of Imitating Proprietary LLMs 1 (May 25,
-
[27]
320 See, e.g., Stefano Maffulli supra note
(unpublished manuscript), https://arxiv.org/abs/2305.15717. 320 See, e.g., Stefano Maffulli supra note
-
[48]
responsible use
239 See, e.g., Paolo Confino, OpenAI Could Be in a 'Clear Violation' of YouTube's Terms of Service, CEO Says—Depending on How It Trains Its Sora Video Tool, Fortune (Apr. 4, 2024, 5:04 PM EDT), https://fortune.com/2024/04/04/openai-youtube-clear-violation-terms-service-ai-sora...
2024
-
[180]
an internet platform on which music fans transcribe song lyrics
193 See Rub supra note 180 at 1179-85 (providing a recap of the case law and history of adoption across different circuits). 9-Dec-24] Terms of Service Woes for AI 41 The Second, Sixth, and perhaps the Ninth Circuits, however, reject the ProCD approach. The Second Circuit, in ...
2022
-
[260]
But if we are correct, and there will be no legal enforcement of this term, then policymakers should be clear-eyed about the effectiveness of the mandate
66 Terms of Service Woes for AI [9-Dec-24 ensure that their licenses require the preservation of this watermarking mechanism. But if we are correct, and there will be no legal enforcement of this term, then policymakers should be clear-eyed about the effectiveness of the manda...
2024
-
[301]
displayed the requested website page with a ‘frame’ at the bottom of the page stating, for example, ‘VIEW 15 RELATED PAGES.’
The legislative history is clear about the broad sweep of preemption: H.R. Rep. No. 94-1476, at 129-33 (1976) (“The intention of § 301 is to preempt and abolish any rights under the common law or statutes of a State that are equivalent to copyright and that extend to works com...
1976
-
[1030]
exceed” that access merely by doing something with it that violated the rules.282 Instead, the Court read “exceed authorized access
279 Id. 280 For discussions of the CFAA and its abuses, see Jonathan Mayer Cybercrime Litigation, 164 U. PA. L. REV. 1453 (2016). 9-Dec-24] Terms of Service Woes for AI 57 CFAA in its 2021 decision in Van Buren v. United States.281 There, the Court held a user who had lawful a...
2016
-
[1201]
Read me Harry Potter,
They are extremely brittle and can be bypassed accidentally in the course of normal usage, including: switching to past tense,272 customizing a model for downstream tasks using model creators’ own customization tools,273 and arguing with a model.274 They are often hand-coded l...
2024
-
[1317]
don’t compete with us if you use our uncopyrightable information on a site open to the public
62 Terms of Service Woes for AI [9-Dec-24 • consider whether the law should change to permit the enforcement of AI terms of service (Section IV.C). A. Summarizing Legal Enforceability & Policy Implications Our analysis suggests that terms of use for model outputs and model wei...
2023
-
[1992]
An action will not be saved from preemption by [extra] elements such as awareness or intent, which alter ‘the action’s scope but not its nature
(“An action will not be saved from preemption by [extra] elements such as awareness or intent, which alter ‘the action’s scope but not its nature.”) (cleaned up). 188 Guy A. Rub, Copyright Survives: Rethinking the Copyright-Contract Conflict, 103 V A. L. REV. 1141, 1147 (2017)...
2017
-
[2001]
204 Dunlap v
(cleaned up). 204 Dunlap v. G&L Holding Grp., Inc., 381 F.3d 1285, 1297 (11th Cir. 2004). 205 See, e.g., Lipscher v. LPR Pubs., 266 F.3d 1305 (11th Cir
2004
-
[2011]
copyright law does not preempt an implied contractual claim to compensation for use of a submitted idea
(holding that “copyright law does not preempt an implied contractual claim to compensation for use of a submitted idea”) 243 X Corp. v. Bright Data Ltd., No. C 23-03698 WHA, slip op. at 25 (N.D. Cal. May 9, 2024). 244 Promises to pay for use are unlikely to be preempted under ...
2024
-
[2017]
covenants running with the software
250 265 F.3d 994, 1103 (9th Cir. 2001). 251 Id. 252 Laws v. Sony Music Entm’t, Inc., 448 F.3d 1134, 1141 (9th Cir. 2006). 253 Terms of Service, Midjourney (Dec. 22, 2023), https://web.archive.org/web/20240226144923/https://docs.midjourney.com/docs/terms-of-service The company ...
2023
-
[2022]
199 Id. at *7. 200 Id. at *2. 201 Genius, No. 20-3113, 2022 WL 710744 at 8 (quoting Harper & Row Publishers, Inc. v. Nation Enters., 723 F.2d 195, 201 (2d Cir. 1983), rev'd on other grounds, 471 U.S. 539 (1985)). Though, scholars have criticized this opinion, as well as simila...
1985
-
[2023]
no preemption
(unpublished manuscript) (available at https://arxiv.org/abs/2305.14233); Rohan Taori et al., Alpaca: A Strong, Replicable Instruction-Following Model, Stan. Ctr. for Rsch. on Found. Models (Mar. 13, 2023), https://crfm.stanford.edu/2023/03/13/alpaca.html. 231 Ding et al., sup...
1985 arXiv
-
[2024]
extra element
(describing these types of preemption). 178 Id. 179 Id. 180 See, e.g., Guy A. Rub, A Less-Formalistic Copyright Preemption, 24 J. Intell. Prop. L. 327, 340 (2017); Guy A. Rub, Moving from Express Preemption to Conflict Preemption in Scrutinizing Contracts over Copyrighted Good...
2017
-
[2025]
In the battle between freedom of contract and freedom of speech, contract almost always wins
(making this point). 9-Dec-24] Terms of Service Woes for AI 63 challenges. Vague restraints, or restraints more squarely centered in copying, will face more scrutiny. And terms are unlikely to run with outputs or models. But in some cases, highly specific terms furthest from c...
2016
Reviewed August 11, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.