On March 20, 2026, the Trump Administration released its National AI Legislative Framework, a seven-section policy document covering children’s safety, energy infrastructure, intellectual property, censorship, innovation, workforce development, and federal preemption of state laws. Guided by a vision of “permissionless innovation” and “minimally burdensome” regulation, the framework’s most consequential provision is in Section III, where the White House states its belief that AI training on copyrighted material does not violate copyright law, but explicitly declines to ask Congress to codify that position, instead deferring the question to the courts and directing Congress not to take any action that would impact the judiciary’s resolution of the issue.
Generative AI is built on millions of copyrighted works used without permission, and the courts are increasingly signaling that this foundation is legally vulnerable, from the Supreme Court’s narrowing of transformative use in Andy Warhol Foundation v. Goldsmith (2023),[1] to Thomson Reuters v. ROSS Intelligence (2025),[2] where a federal court rejected an AI company’s fair use defense after finding it copied copyrighted legal headnotes to train a directly competing product, to Bartz v. Anthropic (2025),[3] where Judge Alsup held Anthropic liable for the unauthorized acquisition and storage of millions of pirated books used to train its Claude AI model, a case that produced a $1.5 billion settlement, the largest known copyright settlement in AI history. Notably, even in Bartz, where the court indicated that training on legally acquired materials could constitute fair use, the unauthorized means of obtaining the training data created separate and devastating liability. The White House framework is not proposing policy in a vacuum; it is placing a bet that the judiciary will ultimately rule in the industry’s favor, while building the surrounding legislative infrastructure to ensure the industry prevails regardless. This article explores that strategy: what the framework proposes, where the courts are heading, and what comes next for the legal foundations of generative AI.
Key takeaways from the framework include:
- The administration states its belief that AI training does not violate copyright law but explicitly defers the question to the courts—and tells Congress not to legislate on it. This prevents Congress from potentially ruling against fair use while betting that the judiciary will ultimately side with the industry.
- It proposes a voluntary collective licensing system modeled on ASCAP/BMI with antitrust immunity, but the framework deliberately leaves unresolved when or whether such licensing is required—and if courts agree with the administration’s stated belief that training is fair use, there would be no legal pressure to license at all.
- It requests a federal digital replica law, directing Congress to create a national standard for the unauthorized use or disclosure of AI-generated likenesses and voices that would preempt state laws like Tennessee’s ELVIS Act and provide a protect shield for AI model developers.
- It builds the policy architecture to favor the industry regardless of how courts rule—by constructing voluntary licensing infrastructure, preempting state regulation, and establishing federal standards on digital replicas, the framework ensures that even an adverse judicial ruling on fair use would land in a regulatory environment designed to soften its impact.
- It federally preempts state AI laws, blocking states from passing their own regulations on AI development, including state-level copyright or training data requirements.
- It says nothing about transparency or disclosure—no requirement for AI companies to reveal what copyrighted material was used in training, making it harder for rights holders to know whether their works were used or to pursue claims.
- The courts are moving faster than Congress—Warhol narrowed transformative use, Thomson Reuters rejected fair use for AI training, Bartz produced a $1.5 billion settlement even where training on legally obtained materials was deemed potentially fair, and pending cases like NYT v. OpenAI could deliver rulings before any legislation passes.
Generative AI models like ChatGPT, Claude, and Gemini are trained by ingesting vast quantities of text, images, code, and other content, much of it copyrighted, scraped from the open internet, digitized book collections, news archives, and proprietary databases. During training, the model processes these works to identify statistical patterns in language, style, and structure, encoding those patterns into billions of numerical parameters. The original works are not stored verbatim inside the model, but they are the essential raw material without which the model could not function. Copyright holders argue that this process constitutes mass infringement on an unprecedented scale; their works are copied, computationally analyzed, and used to build commercial products that compete with and often replace the originals, all without permission or compensation.
The AI industry’s primary defense is fair use, a doctrine that permits unauthorized use of copyrighted material when it serves a sufficiently different purpose than the original. This defense draws on precedents like Authors Guild v. Google, Inc. (2015),[4] where tech companies successfully argued that their use of data was transformative and thus fair use. Companies argue that training is “transformative” because the model does not reproduce the original works but instead learns generalized patterns to produce entirely new outputs. But the Supreme Court significantly narrowed this argument in Andy Warhol (2023), holding that a use is not transformative merely because it alters the original; it must serve a fundamentally different purpose and not act as a commercial substitute. That narrowing has already shaped the lower courts. In Thomson Reuters (2025), a federal judge rejected an AI company’s fair use defense where the trained model directly competed with the copyrighted source material. In Bartz (2025), Judge Alsup issued a split ruling that has become the most significant judicial statement on AI training to date: while indicating that training on legally acquired copyrighted materials could constitute fair use, he held Anthropic separately liable for the unauthorized acquisition and storage of pirated works obtained from datasets like Library Genesis. The resulting $1.5 billion settlement, the largest known copyright settlement in AI history, demonstrated that even where the training process itself may survive fair use scrutiny, the means of obtaining training data carries independent and potentially catastrophic legal exposure. Meanwhile, the highest-stakes cases like The NYT v. OpenAI[5] remain pending, with potential statutory damages in the billions. The legal terrain is evolving rapidly, and the outcome of these cases will determine whether the technical foundations of generative AI rest on solid legal ground or require fundamental restructuring.
It is against this legal backdrop that the White House framework’s intellectual property provisions take on their full significance. Section III does not ask Congress to declare AI training fair use; it does something more calculated. The administration states its belief that training on copyrighted material does not violate copyright law, but it explicitly defers the legal question to the courts and directs Congress not to take any action that would impact the judiciary’s resolution of the issue. This posture accomplishes several things simultaneously: it puts the administration’s thumb on the scale by publicly declaring its position, it prevents Congress from potentially legislating against fair use for AI training, and it places the administration’s bet that the courts, particularly after Bartz, where Judge Alsup indicated training on legally acquired materials could be fair use, will ultimately rule in the industry’s favor. Rather than risk a legislative fight it might lose or a statute that could be challenged, the framework lets the judiciary carry the weight while the administration builds the surrounding infrastructure.
That infrastructure is significant. The framework resembles an ASCAP/BMI-style collective licensing regime, under which copyright holders could participate in a collective system to negotiate broad licenses for AI training and compensation terms, potentially with tailored antitrust protections for joint negotiations. But the framework deliberately specifies that the licensing legislation “should not address when or whether such licensing is required.” This is the framework’s central tension, not a contradiction between a fair use declaration and a licensing system, but something more subtle: a licensing architecture designed without any mandate to use it, whose viability depends entirely on whether courts create the legal pressure the framework deliberately declines to impose. If the courts agree with the administration’s stated belief that training is fair use, AI companies would face no legal obligation to license at all, and the voluntary system would exist as a hollow structure. Only if courts rule against fair use, the outcome the administration is betting against, would the licensing framework carry any practical force.
Notably, the U.S. Copyright Office had spent over a year studying these exact questions, publishing reports on digital replicas, copyrightability, and AI training that concluded existing law was sufficient with no immediate need for new legislation. While the framework’s explicit deferral to the courts on fair use might appear to align with the Copyright Office’s position that no new legislation is needed, the two diverge sharply on the broader question: the Copyright Office concluded that the existing legal framework, including the courts’ authority to resolve fair use, was adequate and that the surrounding policy architecture the framework proposes (licensing systems, federal preemption, digital replica standards) was unnecessary. The administration terminated Register of Copyrights Shira Perlmutter shortly after the final report’s pre-publication, a move widely seen as retaliation for conclusions that conflicted with the administration’s goal of ensuring broad, unlicensed access to training data. No further reports have been issued, and the office’s leadership and direction remain unclear.
The framework also introduces a federal digital replica law that would establish a national standard for AI-generated likenesses, voices, and personas, preempting a growing patchwork of state laws, including Tennessee’s ELVIS Act. Notably absent from the framework is a complete prohibition of any unauthorized generation and use of voice and likeness, instead mandating carve-outs for “expressive works” protected by the First Amendment. Nor is there a requirement that all such outputs be labeled as synthetic. This is a sharp turn away from protections currently trending at the state level, which can prohibit any unauthorized distribution of AI-generated voice and likeness, ban use of AI-generated voice and likeness for political messaging, or require clear labeling of such outputs as synthetic. The framework also seeks to prevent model developers from being held accountable for unlawful conduct of their users, creating a Section 230-style protective shield for these entities. These developers would not have skin in the game to affirmatively ensure their tools that can create realistic likenesses and voices are not used for criminal purposes, such as social engineering cyber attacks or the creation of defamatory and other illegal content (like CSAM).
Thus, the White House National AI Legislative Framework is not a neutral policy document; it is seen by critics as a sophisticated preemptive intervention designed to ensure the AI industry prevails on the question of copyrighted training data, whether through the courts or around them. It states its belief that training is fair use while deferring the question to courts that are increasingly skeptical of that position. It proposes a voluntary licensing system that, by design, carries no mandate, a system whose viability depends entirely on whether courts create the legal pressure the framework deliberately declines to impose. It prevents Congress from legislating against the industry’s position while building the surrounding policy architecture, federal preemption, licensing infrastructure, and digital replica standards, to soften the blow if the courts rule the other way. It sidelines the government’s own copyright authority when its expert conclusions prove inconvenient. The framework’s stated goal is to ensure American AI dominance through minimal regulation and permissionless innovation, and on its own terms, it is internally coherent. But the legal landscape it seeks to manage is not standing still. Courts are issuing rulings, settlements are reaching into the billions, and rights holders are pressing claims with increasing success. Whether the judiciary ultimately validates the administration’s bet, or forces the fundamental reckoning the framework is designed to forestall, remains the central question in AI policy today.
[1] Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023).
[2] Thomson Reuters Enter. Ctr. GmbH v. ROSS Intelligence Inc., 529 F. Supp. 3d 303 (D. Del. 2021).
[3] Bartz v. Anthropic PBC, 791 F. Supp. 3d 1038 (N.D. Cal. 2025).
[4] Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015).
[5] The New York Times Company v. Microsoft Corporation et al, No. 1:2023cv11195 (S.D.N.Y. 2025)