U.S. Copyright Office (2025). Copyright and Artificial Intelligence, Part 3: Generative AI Training. Pre-publication version, released 9 May 2025. A Report of the Register of Copyrights. Washington, DC: U.S. Copyright Office.
Key takeaways
From the reading, NOTE-1024.
- Training is copying, at the input. Downloading, converting, filtering and compiling a dataset “clearly implicate” the reproduction right, and so does copying the dataset to storage for training (pp. 26–27). Whether the batch copies made during training are fixed is left to the facts, and the Office says it matters little: “Such temporary copying may, however, be of little significance given the other acts of reproduction in AI training.” (p. 28 n. 160).
- The weights are a copy only of what they can give back. “Whether a model’s weights implicate the reproduction or derivative work rights turns on whether the model has retained or memorized substantial protectable expression from the work(s) at issue.” (p. 30). The test is fixation plus recoverability with a machine, as for a JPEG (pp. 28–29).
- Warhol makes the use the unit. Each act (curation, pre-training, fine-tuning, RAG) is weighed separately but judged by its ultimate purpose (pp. 36–37, 45). Training a foundation model “will often be transformative” (p. 45). It is least so when a model is trained to emit work like what it was trained on, and “at best, modestly transformative” when its outputs share the originals’ audience (p. 46).
- Two arguments rejected. Training is not non-expressive, because models learn selection and arrangement, “the essence of linguistic expression” (p. 47). It is not like human learning, because “Copyright law should not afford greater latitude for copying simply because it is done by a computer.” (p. 48).
- Market dilution. Outputs that copy no particular work can still harm the market for works of the same kind, including by imitating a style: “If thousands of AI-generated romance novels are put on the market, fewer of the human-authored romance novels that the AI was trained on are likely to be sold.” (p. 65). The Office calls this “uncharted territory”.
- Licensing, not legislation. Voluntary licensing is growing. The Office recommends no government intervention for now, an antitrust clarification for collective licensing, and extended collective licensing where markets are shown to fail. It rejects an opt-out default (pp. 103–106).
Standing in the record
Filed on 2026-10-20 at the owner’s request, as the second work on copyright after Lessig’s Remix (LIT-1298). It is read against Lessig’s copy trigger (NOTE-1023): the report confirms that training sets off the trigger at every input step, and then puts the whole question into fair use, as Lessig said the law does for every digital use. It applies the Supreme Court’s Warhol v. Goldsmith test, which is being filed in this record at the same time. Its bearing on CLAIM-002, CLAIM-210 and the translation line is set out in the reading. It is an anthology candidate (the anthology’s deployment-and-society topic could hold it) and is held here because the anthology holds nothing on copyright.