When Slowing AI Stops Looking Luddite: Copyright and the Pace of Frontier AI

For several years, creators asking AI developers to obtain authorization before using protected works for training have faced a familiar objection: permission raises the cost of training and can slow technological development. Much of the policy debate has treated that consequence as a reason for making copyright more “accommodating” to AI. The discussions now taking place inside frontier laboratories make that assumption harder to sustain. Researchers and executives developing some of the most advanced systems are themselves warning that capability growth may be moving too quickly. The possibility of slowing AI development therefore cannot carry the argument against copyright enforcement on its own.

Jacob Coxon’s resignation from Anthropic brought the issue into unusually concrete terms. Coxon said he had spent the previous three years doing pre-training research at OpenAI and Anthropic before leaving after concluding that both companies were racing toward self-improving systems without adequate restraint. Days later, Anthropic CEO Dario Amodei published “We Must Pace the Frontier”. His proposal concerns the rate of capability advancement itself, and he writes that gaining “an extra year or two” before models reach critical capability levels could greatly reduce risk if that interval were used to improve alignment.

That argument matters to copyright at the level of regulatory presumption, and copyright must still be justified on its own doctrinal and economic grounds. Amodei’s intervention unsettles a narrower assumption that has repeatedly shaped the AI copyright debate, namely that a rule that slows AI development cannot be treated as defective simply because it creates delay. Copyright economics has long recognized that stronger rights can raise input costs and reduce subsequent innovation, as Landes and Posner emphasized. That cost remains real, and what changes is the assumption that technological speed always belongs on the benefit side of the calculation, so that the question becomes how the burden of justification should shift once the speed of AI development itself becomes contestable.

Within that debate, the case for a more accommodating copyright regime has typically been presented as a way of avoiding obstacles to an emerging technology, and rightholders who ask for ordinary rules to apply have found themselves expected to justify the friction that compliance would introduce. Those who invoke the speed of innovation as a reason for weakening copyright constraints have rarely been asked, in comparison, to explain why that speed should be counted as a benefit when the pace of development is itself a matter of policy choice.

American law makes this particularly visible because fair use permits technological investment to proceed under unresolved legal conditions surrounding copyrighted inputs. Bartz v. Anthropic provides the clearest example. Judge William Alsup held that the use of books to train Claude qualified as fair use and treated Anthropic’s acquisition of millions of pirated books for a permanent library as a separate infringement problem. The training holding strengthens the fair-use case under current law, while the provenance dispute illustrates a different point about timing. Litigation over those copies later produced a $1.5 billion settlement, and the provenance cost crystallized after substantial investment in model development had already occurred.

The wider American litigation remains unsettled, since Kadrey v. Meta produced another fair use ruling on a different evidentiary record, while the consolidated New York Times litigation against OpenAI and Microsoft is now before Judge Sidney Stein on cross-motions for summary judgment. The United States entered that case on September 1 with a Statement of Interest supporting OpenAI’s fair-use position. The government describes LLM training as “extraordinarily transformative” and links continued development to scientific progress and American national interests. Eleven days later, Amodei was publicly assigning substantial value to additional time before the next capability threshold. American copyright policy has elaborate ways of counting the benefits threatened by slower AI development and far fewer ways of counting the possible social cost of acceleration.

The government’s filing also states the strongest objection to a permissions regime. It warns that licensing requirements could favor the largest technology companies and create entry barriers for smaller firms. That risk is also one reason to confine the proposal to large-scale commercial training, with lighter obligations for research and smaller operators. The same filing expressly declines to take a position on whether a licensing regime would be financially or logistically feasible. That qualification is important because the competitive effect depends heavily on institutional design. Collective administration can reduce transaction costs, and competition law can address control over scarce inputs where it entrenches market power. Giving frontier developers uncompensated access to protected human production makes creators bear the cost of curing a market-structure problem generated elsewhere.

Congress should therefore bring the legal price of protected training inputs closer to the training decision for commercial frontier models. The White House’s March 2026 National Policy Framework for Artificial Intelligence already tells Congress to consider “licensing frameworks or collective rights systems” through which right holders can negotiate compensation with AI providers. It then adds that any such legislation “should not address when or whether such licensing is required.” The framework therefore contemplates the machinery of collective licensing while directing Congress to keep mandatory licensing outside the legislation.

That is exactly where I disagree. For large-scale commercial frontier training, permission should attach before protected works enter the training corpus, with collective licensing available where individual clearance is impracticable. Section 115, as redesigned by the Music Modernization Act, supplies an existing example of copyright law organizing mass permissions through a centralized licensing architecture. The point of the analogy is institutional, in that American copyright already knows how to construct scalable authorization before mass exploitation occurs.

Europe already requires more of the relevant decision to be taken before training begins. Article 4 of the DSM Directive permits text and data mining of lawfully accessible material subject to rights reservation. Its reach in the generative AI context remains judicially unsettled. Like Company v Google (C-250/25), a pending reference arising from the neighboring right of a press publisher, asks the Court of Justice, among other questions, whether training an LLM involves reproduction and whether the use of lawfully accessible material may fall within the Article 4 TDM exception.

In my 2025 study for the European Parliament, I argued that the structural weaknesses of the present opt-out architecture justifies moving commercial generative AI training towards an opt-in model supported by scalable licensing or remuneration mechanisms (). Such a system forces the developer to confront the legal status and price of protected inputs before training proceeds. Access to training data is routinely presented in Europe as a condition of technological competitiveness, as the Commission’s AI Continent Action Plan illustrates, and any future reconsideration of Article 4 for commercial generative AI should weigh that presumption alongside the temporal consequences of opt-in and opt-out architectures and their effects on authors and licensing markets.

I would resist any attempt to turn this observation into an additional purpose for copyright, since the institution is already expected to serve too many objectives. Training data constitute one part of a much larger technical process, and model capabilities can advance through compute and technical improvements even when copyrighted inputs become more expensive. What the current safety debate changes is the policy significance that can reasonably be assigned to the time licensing may add. The supposed Luddites were never making Amodei’s safety argument for him. They were asserting legal claims whose enforcement happened to create some of the friction that technology policy had learned to treat as inherently undesirable. The warnings now coming from frontier laboratories expose the weakness of that presumption and return a question to the center of the debate where it belonged all along: Who decided that faster was always better?