On Could 9, 2025, the U.S. Copyright Workplace launched a “pre-publication” model of Half III of its extremely anticipated Report on Copyright and Synthetic Intelligence (AI) (Report). The Report gives a technical overview of how generative AI fashions are developed, educated, and deployed and the way U.S. copyright legislation, notably the honest use doctrine, ought to apply within the context of coaching generative AI fashions. The prepublication report states that it was launched “in response to congressional inquiries and expressions of curiosity from stakeholders” and that “[a] last model will probably be revealed within the close to future, with none substantive adjustments anticipated within the evaluation or conclusions.”
Copying Can Happen Throughout Coaching or Use of Generative AI, and the AI Mannequin’s “Weights” Could Additionally Infringe
The Report begins by discussing how curating, accumulating, downloading, reformatting, transferring, and incorporating copies into AI mannequin coaching datasets can contain creating a number of copies of protected works. The Report notes that constructing a coaching dataset utilizing copyrighted works “clearly implicate[s] the best of replica” and additional notes that if mannequin outputs are considerably much like coaching sources, the output may additionally implicate a protected proper. In these cases, the conduct is presumptively infringing except the honest use protection applies.
What occurs in the course of the coaching course of is a little more nuanced, and the extent to which fashions memorize coaching examples is disputed (and sure varies throughout fashions). Nevertheless, in accordance with the Report, if the mannequin can generate an equivalent or almost equivalent copy of the underlying work with out that expression being offered within the type of a immediate or enter, there’s a sturdy argument that the mannequin’s “weights” — numerical parameters that decide the significance of dataset options — might implicate the best of replica. Mannequin weights which have memorized protectable expression from coaching knowledge may additionally infringe the spinoff work proper.
The Report notes that whether or not a mannequin’s weights implicate the replica or spinoff work rights activates whether or not the mannequin has retained or memorized “substantial protectable expression” from the underlying works. In such an occasion, distributing, fine-tuning, or deploying a mannequin might expose builders and downstream customers to legal responsibility for infringement.
The Truthful Use Protection Should Be Evaluated Throughout the Context of General Use
The place copying constitutes prima facie infringement, the following query is whether or not the honest use protection applies. The honest use evaluation considers 4 nonexclusive elements: (1) the aim and character of the use, (2) the character of the copyrighted work, (3) the quantity and substantiality copied, and (4) the market impact.
Issue One: The Objective and Character of the Use Rely on How the AI Mannequin Is Used.
The Report’s evaluation of the primary issue — goal and character of the use — focuses on figuring out the use, transformativeness, commerciality, and lawful entry to the work, with transformativeness and commerciality being key parts. On the vital problem of transformative use, the Report depends on the Supreme Court docket’s reasoning in Warhol v. Goldsmith that transformative use is a matter of diploma. When utilized within the context of coaching AI fashions, the Report asserts that the evaluation relies upon not simply on the coaching course of but additionally on how the mannequin is used. At one finish of the spectrum are research-driven or closed-system makes use of. For instance, scanning books to create a full-text searchable database to help content material moderation could also be extremely transformative. Coaching fashions to generate considerably comparable copyrighted works, nevertheless, might not be. The Report notes that not like instances the place copying was merely a method to take away interoperability obstacles, utilizing photographs or sound recordings to generate considerably comparable expressive outputs is unlikely to be transformative except the work itself is being focused for remark or parody.
The Report additionally explains that retrieval-augmented technology (RAG) — which reinforces the efficiency of generative AI fashions by scraping data from exterior databases, paperwork, or the net — requires separate consideration. Not like pretraining with a big, various dataset, RAG retrieves focused works for the aim of enhancing the immediate output. RAG is much less prone to be transformative the place the aim is to generate outputs that summarize or abridge copyrighted works.
In making this distinction, the Report primarily disagrees with two frequent arguments that coaching AI fashions is inherently transformative. As to the argument that the aim will not be expressive, the Report causes that as a result of fashions do extra than simply statistical sample recognition (they study the choice and association of underlying phrases, photographs, and sounds), coaching the fashions encompasses the “essence” of inventive expression. As to the argument that the method is much like human studying, the Report causes that the honest use protection doesn’t defend all copying for the aim of studying and doesn’t distinguish between acts finished by a pc or a human.
As to the function of the opposite vital ingredient of the primary issue — commerciality — the Report notes that commerciality activates whether or not the use “furthers industrial functions,” not on the for-profit or nonprofit standing of the entity concerned in use of the generative AI mannequin.
Issue Two: The Nature of the Copyrighted Work Is dependent upon the Kinds of Works within the Coaching Set.
The Report states that the second issue — the character of the copyrighted work — is dependent upon the “mannequin and work at problem.” Observers have commented that the second issue not often performs a considerable function in honest use balancing. The Report notes that the majority AI fashions are educated on a wide range of forms of works and concludes that if the works are extra expressive or beforehand unpublished, this issue would weigh towards honest use.
Issue Three: The Quantity and Substantiality Copied Ought to Think about Guardrails In opposition to Infringement and What Content material Is Made Public.
The third honest use issue examines how a lot of a copyrighted work was used and whether or not that quantity was affordable in gentle of the aim. Right here, the Report notes that AI mannequin coaching often entails full or almost full copying of whole works and makes use of their expressive content material for coaching, which weighs towards honest use. Nevertheless, the Report means that builders can mitigate the presumption towards honest use by displaying that the copying was functionally essential to a transformative goal and that efficient guardrails had been used to stop the output of protected expression. It’s value noting that the Report additionally finds that the presence of technical guardrails is related to the primary issue as a method of limiting a mannequin’s potential to breed copyrighted materials and the chance of market substitution.
Issue 4: Market Impact Is dependent upon Outputs That Could Influence the Market By means of Misplaced Gross sales, Dilution, and Licensing Charges.
As to the fourth honest use issue — impact available on the market — the Report evaluates other ways by which the usage of copyrighted works in coaching AI fashions can have an effect on the market worth of protected works and addresses broader claims that the general public advantages of unlicensed coaching may shift the honest use steadiness. Right here, the Report identifies three classes of potential hurt: misplaced licensing alternatives, misplaced gross sales, and market dilution.
Specifically, whereas the primary and second classes are sometimes thought of within the honest use evaluation, the Report notes that some commentators additionally advocated for consideration of the potential hurt brought on by market dilution (i.e., the place even these outputs that aren’t considerably much like a particular copyrighted work might however compete out there for that kind of labor). The Report authors seem to have been persuaded by this novel idea, noting that “stylistic imitation made potential by [the original work’s] use in coaching might impression the creator’s market,” and warn that “the pace and scale at which AI techniques generate content material pose a critical threat of diluting markets for works of the identical type as of their coaching knowledge.” Nevertheless, even the Report acknowledges that this place is “uncharted territory” and no court docket has but embraced it as a motive to disclaim honest use.
The Report Advocates for Voluntary Licensing, Collective Bargaining, and Coverage Reform.
The Report additionally discusses varied licensing options for AI mannequin coaching, together with voluntary, collective, and obligatory licenses, and a statutory “opt-out.” In giving its advice, the Report stress the significance of recognizing that coaching entails a large variation of works, which can have an effect on the feasibility of the licensing regime. Voluntary licensing could also be possible the place there are massive volumes of copyrightable materials or a restricted variety of copyright homeowners. Collective licensing could also be possible and will cut back transactional prices if acceptable safeguards towards anticompetitive habits are carried out and might be used to “protect some potential to dam undesirable makes use of or negotiate phrases” ought to Congress ever contemplate an exception or limitation for AI coaching. Nevertheless, the Report additional acknowledges that obligatory licenses might hamper versatile and artistic market-based options and are arguably inconsistent with the fundamental requirement of consent to make use of copyrighted works.
Taken as an entire, the prepublication report on AI mannequin coaching takes a measured method however seems to favor copyright homeowners — most notably in its endorsement of the novel market dilution idea of hurt.
