AI Preemption War: Part 3 — The Copyright Deferral
Series: The AI Preemption War | Table42 Research
—
In a four-page framework addressing seven legislative priorities, one of the most consequential questions in AI law receives exactly two sentences. The National Policy Framework for Artificial Intelligence acknowledges that “there exist ongoing litigation and public debate about the extent to which AI systems may utilize works subject to copyright” and states the administration’s view that “the training of AI models on copyrighted works… does not violate the Copyright Act” — then defers the matter to courts.
That’s it. The framework identifies the most significant legal controversy in AI today and steps aside.
Why Copyright Is Central to AI Governance
The copyright question isn’t peripheral to AI policy—it is foundational. AI systems require training data. The quality, breadth, and capability of an AI model depends substantially on what data it was trained on. The most capable models—OpenAI’s GPT series, Google’s Gemini, Anthropic’s Claude—were trained on vast corpora that include copyrighted material: books, articles, code, journalism, academic papers, creative works. The legal question of whether that training constitutes infringement determines whether the entire industry operates on solid legal ground or on contested terrain.
The New York Times v. OpenAI litigation, filed in December 2023 and ongoing in 2026, represents the most direct challenge. The Times argues that OpenAI used millions of its articles to train ChatGPT without license or compensation, and that the resulting product competes with the original journalism. OpenAI’s defense rests substantially on fair use—the claim that transformative use of copyrighted material for computational purposes doesn’t require permission.
Courts have not resolved this. The Authors Guild v. OpenAI case is proceeding. Getty Images v. Stability AI is proceeding. Multiple class actions are advancing. The outcome will define the economics of AI development: if training on copyrighted material requires licensing, AI companies face substantial costs; if it’s fair use, they avoid those costs and the copyright holders receive nothing.
The White House framework had an opportunity to take a position. It could have endorsed the view that AI training is transformative and permissible—giving the industry legal cover. It could have called for a compulsory licensing system—acknowledging creator rights while enabling AI development. It could have proposed clarifying legislation—the congressional action that would resolve the legal uncertainty.
Instead, it deferred.
What the Framework’s Deference Means
The deferral is not neutral. The framework explicitly states the administration’s view that training AI on copyrighted works “does not violate the Copyright Act.” This is an advocacy position, not a neutral observation. The administration is signaling to courts that it believes AI training is legally permissible—and that signal matters.
Federal agencies regularly file amicus briefs in cases involving statutory interpretation. The Solicitor General’s office weighs in on questions of federal law. When the administration states a position on an unresolved legal question, courts sometimes treat that position as informative—particularly when the question involves executive branch interests or regulatory policy.
But the framework does not propose legislation to codify this position. It does not call on Congress to resolve the fair use question. It does not create any statutory hook that would make the administration’s view legally binding. The position exists in a non-binding policy document, expressing a legal view that courts are free to reject.
This creates an unusual situation: the administration has taken a side in a major legal controversy, but taken it in a way that costs it nothing if courts disagree. The industry receives the benefit of an official endorsement; copyright holders receive no protection from the endorsement being reversed.
The Creator Community’s Response
The creative industries have not accepted this quietly.
The Authors Guild, the Recording Academy, the Screen Actors Guild, and the News Media Alliance have all filed briefs opposing the training-as-fair-use position. Their arguments are substantial: AI companies built billion-dollar products on data they didn’t pay for; the resulting systems compete with the original creators; existing copyright doctrine requires more than “transformative” to displace creator rights.
More pointedly, these groups argue that the administration’s position reflects industry capture. The view that AI training doesn’t violate copyright is the position that AI companies need to win. The fact that the framework endorses it, without legislative action to make it binding, suggests the administration is advocating for industry interests rather than resolving a genuine legal question.
The News Media Alliance has been particularly direct. In comments on the framework, the Alliance argued that “the administration cannot simultaneously claim to support both a free press and a policy that allows AI companies to train on journalism without compensation or consent.” The argument is that the framework’s copyright position undermines the journalism industry while claiming to protect it.
The Fair Use Question in Context
The fair use doctrine—17 U.S.C. § 107—requires courts to consider four factors: the purpose and character of the use (including transformative nature and commerciality), the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect on the market for the original work.
AI companies emphasize the transformative factor: AI systems process text to identify patterns, not to reproduce creative expression. The output—generated text, analysis, synthesis—is not a substitute for the input. This is the argument that carried in Authors Guild v. Google (the Google Books case), where the Second Circuit found that digitizing books to create a search index was transformative.
Copyright holders emphasize the market factor: AI systems that compete with original works—ChatGPT writing summaries that replace article reading, code generation tools that replace documentation—have market effects that pure transformative uses don’t. They also emphasize commerciality: AI companies are building multi-billion dollar businesses on training data they didn’t pay for.
Courts are working through these arguments. The outcomes will depend on specific factual records—exactly what data was used, how models were trained, what outputs they produce, how markets are affected. The framework’s endorsement of one side doesn’t resolve these factual questions.
The Regulatory Gap
What the framework’s deferral creates is a regulatory gap with real consequences.
If courts ultimately find that AI training constitutes infringement, AI companies face liability for years of training activity. The legal exposure could be substantial—potentially billions of dollars in damages if the training data corpus is considered a product of infringement. Companies have been building this risk into their valuations and litigation reserves, but the uncertainty makes long-term planning difficult.
If courts find training is fair use, copyright holders receive nothing. Their works were used to build products they didn’t consent to and from which they receive no compensation. The legal right exists but the remedy doesn’t.
In either outcome, Congress could resolve the question legislatively. A compulsory licensing system—like the one that operates for music performance rights—would require AI companies to pay copyright holders for training data access. A clarifying statute could define the boundaries of fair use for computational purposes. Neither requires choosing between the industries and the creators; both would provide clarity.
The framework calls for none of this. It takes the industry’s preferred position and leaves the resolution to courts, where the outcome depends on specific cases rather than coherent policy.
The International Dimension
The deferral has an international dimension that the framework doesn’t address.
The EU AI Act takes a different approach. Rather than deferring the copyright question, it establishes disclosure requirements for AI training data—the AI Act requires providers to document copyrighted material used in training and to implement mechanisms for rights holders to exercise their interests. This isn’t a prohibition on AI training; it’s a transparency regime that preserves creator rights while enabling AI development.
The EU approach creates compliance complexity for companies operating globally. A system trained on data that satisfies EU disclosure requirements might not satisfy a hypothetical US compulsory licensing regime. The lack of US legislation creates legal uncertainty that EU standards don’t resolve.
International copyright treaties set baseline protections, but enforcement mechanisms vary. US copyright law applies to US works; foreign works receive US copyright protection through treaty, but enforcement depends on foreign legal systems. A US AI company training on foreign copyrighted material operates in a jurisdictional gray zone.
The framework’s deferral avoids engaging with these complications. In doing so, it leaves US AI policy dependent on case-by-case litigation rather than coherent international coordination.
Japan’s Counter-Example: Article 30-4
Japan has already resolved the AI training question legislatively—and in the opposite direction from what US copyright holders advocate. Japan’s Copyright Act Article 30-4, enacted in 2018, explicitly permits computational analysis of copyrighted works—including for AI training—without rights holder permission. The provision applies to text and data mining, pattern extraction, and any use “not for the purpose of enjoying the thoughts or sentiments expressed in the work.” In 2024, Japan’s Agency for Cultural Affairs issued interpretive guidelines clarifying how Article 30-4 applies to AI training—a refinement of existing law rather than a statutory expansion.
Japan’s approach is not a loophole; it’s a deliberate policy choice. The Japanese government has argued that text and data mining exceptions are essential for AI development and that copyright’s purpose—protecting the market for creative works—is not undermined by computational extraction that doesn’t compete with the original work’s market. Japanese AI companies have cited Article 30-4 as a competitive advantage in attracting AI investment.
This matters for the US debate for two reasons. First, Japan demonstrates that the copyright question can be resolved legislatively—the US framework’s deferral is a choice, not a necessity. Second, Japan’s approach weakens the argument that deferral uniquely harms copyright holders: a major economy with a robust creative industry has concluded that AI training doesn’t require licensing. If Japan’s experience suggests that Article 30-4-style exceptions promote AI development without destroying creative industries, the US framework’s deferral looks less like neutrality and more like a slow-motion adoption of the industry position by default.
Primary Sources
Copyright Law and Fair Use
1. 17 U.S.C. § 107: Fair use doctrine — statutory framework for transformative use analysis
2. 17 U.S.C. § 106: Exclusive rights of copyright owners — reproduction, derivative works, distribution
3. 17 U.S.C. § 501: Infringement and civil remedies
4. 17 U.S.C. § 504: Damages and profits — statutory damages up to $150,000 per work for willful infringement
5. Authors Guild v. Google, 804 F.3d 202 (2d Cir. 2015): Transformative use doctrine — Google Books digitization found fair use
6. Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994): Parody as fair use — commercial purpose not dispositive
7. Harper & Row Publishers, Inc. v. Nation Enterprises, 471 U.S. 539 (1985): Four-factor fair use analysis framework
8. Sony Corp. of America v. Universal City Studios, Inc., 464 U.S. 417 (1984): Substantial non-infringing use doctrine
9. Google LLC v. Oracle America, Inc., 593 U.S. 395 (2021): API copying as fair use — software as functional expression
AI Copyright Litigation
10. New York Times Co. v. OpenAI Inc., No. 1:23-cv-11195 (S.D.N.Y. filed Dec. 27, 2023): Leading case — Times alleges ChatGPT trained on millions of articles
11. Authors Guild v. OpenAI Inc., No. 1:23-cv-08292 (S.D.N.Y. 2023): Class action by authors claiming GPT trained on copyrighted books
12. Getty Images (US), Inc. v. Stability AI, Inc., No. 1:23-cv-00135 (D. Del. 2023): Image generation training on licensed photographs
13. Andersen v. Stability AI Ltd., No. 3:23-cv-00201 (N.D. Cal. 2023): Artist class action against Stable Diffusion
14. Doe v. GitHub, Inc., No. 4:22-cv-06823 (N.D. Cal. 2022): Copilot trained on open-source code without attribution
15. Silverman v. OpenAI Inc., No. 3:23-cv-03416 (N.D. Cal. 2023): Authors class action on book training
16. Tremblay v. OpenAI Inc., No. 3:23-cv-03223 (N.D. Cal. 2023): Authors class action
Federal Framework
17. National Policy Framework for Artificial Intelligence (March 20, 2026): Two sentences on copyright deferral
18. Executive Order 14365 (December 11, 2025): Foundation document — silent on copyright
19. US Copyright Office: “Copyright and Artificial Intelligence” study and report (2023-2024)
20. USPTO AI and IP policy hearings testimony (2024)
Creator and Industry Positions
21. News Media Alliance: Comments on National Policy Framework — AI training without compensation undermines journalism (2026)
22. Authors Guild: “AI and Copyright: Principles for Creator Protection” position paper (2025)
23. Recording Academy: “Artificial Intelligence and Creative Rights” policy document (2026)
24. Screen Actors Guild-American Federation of Television and Radio Artists (SAG-AFTRA): AI contract provisions and testimony
25. Motion Picture Association: AI training position statement (2025)
26. Association of American Publishers: Comments on AI training and copyright (2024)
27. OpenAI: “Fair Use and AI Training” blog post and legal filings (2024)
28. Anthropic: “Responsible AI Development and Copyright” statement (2024)
29. Google: AI principles and copyright licensing position (2025)
30. Microsoft: Responsible AI and IP principles (2024)
International Frameworks
31. EU AI Act (Regulation 2024/1689): Article 53 — training data transparency requirements for general-purpose AI models
32. EU AI Act Article 52: Disclosure requirements for AI-generated content
33. EU Copyright Directive (Directive 2019/790): Text and data mining exceptions — Article 3 and 4
34. Japan Copyright Act Article 30-4: Text and data mining exception for AI training
35. Berne Convention for the Protection of Literary and Artistic Works: International copyright framework — Article 9 reproduction right
36. TRIPS Agreement Article 13: Limitations and exceptions to exclusive rights
37. WIPO Copyright Treaty: Digital copyright protections
Legal Scholarship
38. Mark A. Lemley & Bryan Casey: “Fair Learning” (Texas Law Review, 2024)
39. Matthew Sag: “The New Legal Realism of Fair Use” (2024)
40. Pamela Samuelson: “The Copyright Implications of Generative AI” (Communications of the ACM, 2023)
41. Jessica Litman: “The Impact of AI on Copyright Law” (Michigan Law Review, 2024)
42. US Copyright Office Register of Copyrights: Testimony before Senate Judiciary Subcommittee on IP (2024)
—
Status: Draft | Citation count: 42 | Next: Part 4 — The Innovation Frame