Copyright Office says fair training turns on the works, their source, the purpose and the controls on outputs

2025-05-09Copyright

United States Copyright Office, "Copyright and Artificial Intelligence, Part 3: Generative AI Training," A Report of the Register of Copyrights, pre-publication version (May 2025), released May 9, 2025

The Copyright Office released a pre-publication version of the third part of its report on copyright and artificial intelligence, on training, and said it expects no substantive change in the final version. The Office concluded that several stages of building a generative system implicate the copyright owner's exclusive rights: assembling a training dataset reproduces works, training makes temporary copies, and model weights that have memorized protectable expression may themselves infringe the reproduction and derivative work rights, because memorized works can be generated from the fixed weights. On fair use, the Office stated that various training uses are likely to be transformative, but that whether they are fair "will depend on what works were used, from what source, for what purpose, and with what controls on the outputs." Deployment for analysis or research is unlikely to produce substitutes, while commercial use of vast troves of works to produce competing expressive content, "especially where this is accomplished through illegal access, goes beyond established fair use boundaries." The Office read the fourth factor to reach market dilution caused by the speed and scale of generated content of the same kind, and said that where licensing markets are available, unlicensed uses will be disfavored. It recommended letting the licensing market develop without government intervention, with targeted measures such as extended collective licensing only on a showing of market failure.

What it changesThe Office tied fair use to the controls on outputs, and those controls are engineering: filters that block reproduction of training text, deduplication that reduces memorization, retrieval systems that cite and limit what they return, and provenance tracking that shows where each work came from. They now carry legal weight as well as product value, which makes them worth patenting and worth documenting in the record a company would rely on in litigation. The statement that memorized expression in weights may infringe also changes diligence: a model acquired or licensed in should come with memorization testing, and an exclusive license to weights should carry indemnities that match. Keep the dataset provenance record, source and license status for each corpus, in the same confidential file that holds the trade secret inventory, and describe training data in patent specifications by type rather than by named source. The report is the Office's analysis, not law; courts decide fair use.