The U.S. Copyright Office concluded in May 2025 that copying copyrighted works to train generative AI implicates the owners' exclusive rights, and that whether it is excused as fair use depends on the facts of each case. Its report says many training uses are likely to serve a new purpose that differs from the original works, but that commercial use of vast troves of works to produce competing expressive content, especially through illegal access, goes beyond established fair use boundaries. The report is advisory, appeared only in pre-publication form, and does not bind any court.
Key takeaways
- Part 3 of the Office's report on copyright and AI was released in pre-publication form on May 9, 2025, and the Office's AI page still lists a final version as forthcoming.
- The Office said training a foundation model on a large, diverse data set will often serve a new purpose, while stressing that this is a matter of degree.
- Knowingly using pirated or illegally accessed works should weigh against fair use without settling the question alone.
- The Office read the fourth factor broadly enough to include market dilution, where AI output floods a market for works of the same kind.
- It judged government intervention on licensing premature and pointed to growing voluntary licensing.
What the report is and what it is not
The Copyright Office has published its work on copyright and artificial intelligence in parts. Part 1, on digital replicas, appeared on July 31, 2024. Part 2, on copyrightability, followed on January 29, 2025. Part 3, on generative AI training, was released on May 9, 2025 as a pre-publication version, responding to congressional inquiries and stakeholder interest. The Office states that a final version will be published and that no substantive changes to the analysis or conclusions are expected. As of October 27, 2026 its AI page still shows the final version as pending.
The report draws on more than 10,000 public comments on a notice of inquiry issued in August 2023. It offers an analytical framework and says it does not opine on specific cases. Courts decide disputes, and the report's reasoning is persuasive material, not law.
Where the report finds a copyright issue
The conclusion states that several stages in building generative AI use copyrighted works in ways that implicate the owners' exclusive rights. Most commenters agreed on what follows: the key question is whether those acts of apparent infringement can be excused as fair use. That doctrine weighs four statutory factors in light of all relevant circumstances, so the report declines to predict any single outcome.
The report also takes up whether a trained model itself contains copies. It notes that courts addressing claims about model weights have reached varying conclusions, citing one decision that called the argument "nonsensical" on the pleaded facts and another that let allegations proceed that protected elements remained in some form inside a model. The question stays open.
The four fair use factors as the report applies them
| Factor | Report's reasoning in brief |
|---|---|
| Purpose and character of the use | Training a foundation model on a large and diverse data set will often serve a new purpose, since the result can translate emails or answer questions rather than reproduce the works. How far the purpose differs is a matter of degree and depends on how the model is deployed. Knowing use of pirated or illegally accessed works weighs against fair use without being determinative. |
| Nature of the work | Using creative or expressive works such as novels, films, art, or music is less likely to be fair than using factual or functional works such as computer code. Because training sets usually include expressive works, commenters said this factor often weighs against fair use. |
| Amount used | Developers ordinarily copy entire works, which weighs against fair use, but copying whole works can be reasonable where the purpose is new and large volumes are needed. Effective guardrails that block infringing outputs make this factor weigh less heavily against fair use. |
| Effect on the market | Includes lost sales, lost licensing opportunities, and market dilution. The Office rejected a narrow reading limited to the specific work copied. |
Market dilution
The report's most discussed move is on the fourth factor. It acknowledges that this is uncharted territory but concludes that the statute's reference to any effect on the potential market covers dilution. The speed and scale of AI generation pose a serious risk of diluting markets for works of the same kind as the training data. The Office gives an illustration: if thousands of AI-generated romance novels reach the market, fewer of the human-written romance novels the system learned from are likely to sell. Royalty pools can be diluted too, and imitation of a creator's style, which copyright does not protect by itself, may still harm that creator's market.
Some commenters, including Meta, Google, and the Authors Alliance, argued the opposite, that fair use cases have never measured harm across a general class of works. The report records that disagreement and sides against the narrow view.
Licensing, and what the Office advised
The report concludes that licensing agreements for AI training, both individual and collective, are emerging in certain sectors, though availability is inconsistent. Given that growth and the lack of stakeholder support for statutory change, the Office considered government intervention premature. Where gaps are unlikely to be filled, it suggested that approaches such as extended collective licensing should be considered. It also wrote that outputs from models deployed for analysis or research are unlikely to substitute for the works used in training.
For readers in Canada, the report describes U.S. law. Canadian copyright uses a differently structured exception called fair dealing, so the four-factor reasoning above does not carry over directly.
This article gives general information and is not legal advice. Anyone with a specific dispute, license, or product decision should consult a qualified lawyer and the primary texts, since litigation on AI training is still developing.
FAQ
Is the Copyright Office's training report binding law?
No. It is a report of the Register of Copyrights to inform Congress, courts, and the public. Judges may find its reasoning persuasive, but they apply the statute and case law themselves.
Did the Office say AI training is fair use?
No, and it did not say the opposite. It said many uses are likely to serve a new purpose and that fairness depends on the works used, their source, the purpose, and the controls on outputs.
Does training on pirated books automatically defeat fair use?
According to the report, no. The Office said knowing use of pirated or illegally accessed works should weigh against fair use without being determinative.
Why does the report say a final version is coming?
The Office released the pre-publication version early because of congressional interest. It said a final version will follow and did not expect substantive changes. No final version was listed as of October 27, 2026.
Related reading
- International legal frameworks for AI
- What makes AI trustworthy
- The ethical dilemma around AI
- Explainable AI concepts
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version), U.S. Copyright Office, May 2025
- Copyright and Artificial Intelligence, U.S. Copyright Office, page checked October 27, 2026