Copyright Office Part 3 Generative AI Training Pre-Publication Version 2025

[removed all footnotes and sections other than Analysis and Recommendations]

In assessing any form of licensing, it is important to recognize the wide variations in works and uses involved in AI training. Feasibility will depend on the types of works needed, the licensing practices of the relevant industries, the design of the AI system, and its intended uses. For instance, licensing a music model that can produce rudimentary jingles is different from licensing a state-of-the-art LLM that can compete on advanced reasoning benchmarks. And sophisticated commercial entities will be easier to find and negotiate with than individual non-professionals. As discussed above, a number of voluntary direct and collective licensing agreements for using copyrighted works in AI training have emerged over the past several years, with others in development. Some AI systems have now been trained exclusively on licensed or public domain works. These developments demonstrate that voluntary licensing may be workable, at least in certain contexts—particularly where training is focused on valuable content that can be licensed in relatively high volumes (e.g., popular music and stock photography), or in fields where the number of copyright owners is limited. The Office recognizes, however, that practical challenges remain in many areas. The growing licensing market does not itself establish that voluntary licensing is feasible at scale for all AI training needs. To the extent that the remaining gaps cannot reasonably be filled, alternative solutions may be needed. As to compensation, further market developments may provide more insight on the extent to which licensing agreements can effectively compensate copyright owners for the use of their works in AI training. The agreements that already exist indicate that mutually agreeable compensation terms can be negotiated in some situations, although it remains to be seen how they scale. Compensation structures based on a percentage of revenue or profits, without large up-front cash outlays, may be an attractive alternative for smaller developers looking to enter the market. As to concerns voiced by commenters about the affordability for academic researchers, we note that the research projects they identify may well qualify as fair use and therefore would not require licenses. And the amount of monetary compensation that some copyright owners will accept may depend on contractual conditions regarding control of the use of their works. As discussed above, there appears to be strong interest among those representing copyright owners and creators in developing voluntary collective licensing for the AI context. Collective licensing can play a significant role in facilitating AI training, reducing what might otherwise be thousands or even millions of transactions to a manageable number. The aggregation of rights could be mutually beneficial, such as where transaction costs might otherwise exceed the value of using a work or where copyright owners might be difficult to find. Although collective licensing presents its own logistical and organizational challenges, it affords copyright owners and licensees flexibility to tailor agreements to their needs. Multiple CMOs can each license different types of copyrighted works on terms that make sense for that particular creative industry and AI model. As to antitrust concerns, courts have found that there is nothing intrinsically anticompetitive about the collective, or even blanket, licensing of copyrighted works, as long as certain safeguards are incorporated—such as ensuring that licensees can still obtain direct licenses from copyright owners as an alternative. Although antitrust law is beyond the scope of the Office’s expertise, we believe that greater clarity would be valuable. We encourage the Department of Justice to provide guidance, including on the benefit of an antitrust exemption in this context. We agree with commenters that a compulsory licensing regime for AI training would have significant disadvantages. A compulsory license establishes fixed royalty rates and terms and can set practices in stone; they can become inextricably embedded in an industry and become difficult to undo. Premature adoption also risks stifling the development of flexible and creative market-based solutions. Moreover, compulsory licenses can take years to develop, often requiring painstaking negotiation of numerous operational details. For those sectors where voluntary licensing may prove unworkable or infeasible, ECL would be a less intrusive approach. It would permit copyright owners to choose to license separately, while enabling full coverage of the entire sector for AI training. Allowing authorized CMOs to negotiate rates and terms and establish policies and procedures, subject to government oversight would provide flexibility, rather than freezing rates in the statute or setting them through judicial or administrative proceedings.

As to the possibility of an opt-out mechanism, the Office agrees that requiring copyright owners to opt out is inconsistent with the basic principle that consent is required for uses within the scope of their statutory rights. But to the extent that Congress may consider an exception or limitation for AI training in the future, the ability to opt out could preserve some ability to block unwanted uses or negotiate terms. Nevertheless, significant concerns have been raised about the effectiveness and availability of opt-outs, which would need to be addressed.

Finally, we note that the law, technology, and markets for training are relatively nascent, and there is a dynamic interplay between them. To begin with, the current licensing market may be distorted by the unsettled legal questions about fair use. While some AI companies may have licensed works for training to avoid uncertainty or obtain access to high-quality or otherwise-unavailable materials, other licensing activities may be inhibited by reliance on fair use. As courts begin to resolve pending cases, greater legal clarity may lead to greater collaboration on technical and market-based solutions. Similarly, new model architectures and techniques may be developed to facilitate training using fewer unlicensed works without sacrificing quality. Whether companies devote resources toward such solutions may in turn be influenced by the shifting incentives created by legal and licensing developments. In light of the foregoing, at this point in time, the Office recommends allowing the licensing market to continue to develop without government intervention. If market failures are shown as to specific types of works in specific contexts, targeted intervention such as ECL should be considered.