U.S. Copyright Office Issues Guidance on Generative AI Training

The Office found that training and deploying a GenAI system using copyright-protected material involves multiple acts that could establish prima facie infringement. For example, Part Three states that "[t]he steps required to produce a training dataset containing copyrighted works clearly implicate the right of reproduction[]" and describes actions such as downloading, transferring, converting, and copying of the works. The Report further explains that the training process itself implicates the right of reproduction by downloading the dataset and copying it to high-performance storage prior to training and also when works are "shown" to the model (and thus temporarily copied) in batches.  

Additionally, model weights that demonstrate "memorization" of protected expression in works used during training could themselves be considered copies or derivative works. GenAI outputs may infringe the reproduction right, the right to prepare derivative works, and, depending on the content type, the public display and public performance rights.  

The Office then addressed whether training GenAI models on copyrighted material without permission may be considered fair use. In summary, the Office offered the following guideposts: 

• Purpose and character: Uses that are highly transformative weigh in favor of fair use. Models that perform a general-purpose function will often be transformative, but the degree of transformativeness will depend on the model's functionality and deployment. By contrast, training AI to generate expressive content that competes with the originals, especially for commercial ends, is less likely to be deemed fair. The commerciality inquiry depends on whether the user stands to profit without paying the customary price—not the motive of the user. In the Office's view, knowingly using pirated or unlawfully accessed works should weigh against fair use. 
• Nature of the work: Using highly creative or unpublished works weighs against fair use, while using factual or functional material may favor fair use.
• Amount and substantiality: GenAI training typically involves copying entire works, which often weighs against fair use. The Office acknowledges that copying whole works can be necessary to train effective models, but it distinguishes AI training from cases like Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015) where full copying was essential for a transformative search tool. Copying extensive material is not automatically excused by scale or necessity; the analysis looks at whether the scope of copying is reasonable in light of the transformative purpose. 
• Market effect: The potential for AI-generated content to substitute for original works, dilute their market value, or undermine licensing opportunities is a significant factor weighing against fair use, but the public benefits of GenAI provide compelling arguments in its favor. The analysis of just how much a particular GenAI tool disrupts the market for the works used in its training requires a fact-intensive inquiry. 

The Office concludes that the ultimate determination of fair use requires a case-by-case assessment that "balance[es] multiple statutory factors in light of all relevant circumstances." Although the Office notes the weighing of the four factors is left to the courts, the purpose-and-character and market-effect factors "can be expected to assume considerable weight in the analysis." 

Looking beyond fair use, the Office analyzes some of the challenges faced by voluntary licensing markets for AI training in sectors like music, news, and stock imagery. The Office discusses the role of collective licensing (e.g., through collecting societies or industry consortia) to streamline permissions, but notes potential hurdles such as antitrust concerns and administrative overhead. The Office found little support for imposing compulsory licensing or extended collective licensing schemes for AI at this time, noting that stakeholders generally prefer an "opt-in" approach where they can choose when, how, and to whom they license their works. Thus, the Office recommends allowing the licensing market for AI training to develop naturally, reserving any statutory interventions to address clear market failures. 

Overall, Part Three reinforces a case-by-case, market-driven strategy and advises all stakeholders to remain engaged as the intersection of AI and copyright law evolves. The Office acknowledged that its analysis is limited to "current circumstances and publicly available information" in a rapidly evolving field and that it is committed to "monitor developments to determine whether any conclusions should be revisited." Interested parties should watch for the final version of Part Three to publish, as well as Part 4 of the Report which will address infringement liability and transparency requirements.  

Also of note, the Office is currently in a state of flux, with the Trump administration's recent dismissal of the Register of Copyrights and Librarian of Congress and its appointment of new individuals to hold these positions.