Copyright Fair Use Regulatory Approaches in AI Content Generation

Ariel Soiffer is a partner and Aric Jain is an associate at the law firm WilmerHale.

The impact of generative artificial intelligence (AI) has quickly caught the attention of technologists and policymakers around the world. Among others, policymakers in Washington are scrambling to apply intellectual property (IP) laws and concepts in response. Indeed, just this month, the Senate Subcommittee on Intellectual Property held its second hearing on AI and its implications for copyright law. Congressional attention to copyright and AI matches a growing public interest in understanding how AI – and generative AI, in particular – uniquely affects what it means to be an author and how ownership of expression of ideas is determined.

Fundamentally, this is about the relationship of works generated from generative AI models (Output Works) to works used to train generative AI models (Input Works) and how U.S. copyright law applies to that relationship. This article begins with an overview of generative AI and copyright law with a focus on fair use doctrine. It then examines four schools of thought that have emerged to address the novelty of generative AI under copyright law. It posits some of the implications for each of these approaches for innovation and the growth of the generative AI industry.

Generative AI and its Reliance on the Fair Use Defense

Generative AI refers to the use of artificial intelligence technology to produce various types of content, such as images, text, audio, video and synthetic data. The generative AI (GAI) models making headlines are text-based – they generate and modify works based on text prompts by human users.

By way of background, a GAI learns by “training” itself on Input Works. A typical GAI is fed vast amounts of such works to process initially. Machine learning algorithms then detect patterns in the compositions of Input Works and create probabilities and complex correlations that the GAI uses to predict a suitable response to a user’s prompt. For example, large language models are often trained on billions of human-authored documents that allow those models to eventually recognize differences in human expression between an email or a newspaper article and adjust their output accordingly.

Similarly, image diffusion GAIs analyze the correlation of pixel arrangements of images that are responses to search engine queries with the text queries that produced the results. After analyzing billions of data points, the image diffusion GAI can generate novel images from a user’s text prompt based on the calculated correlation with image outputs that one would expect to appear from a relevant Google search.

These generative processes expose gaps in some less-explored areas in copyright law. In general, copyright law gives the exclusive right to copy works, among other exclusive rights, to the applicable copyright holder. Any works produced from unauthorized copying constitute copyright infringement and should be considered derivative works (as defined by the Copyright Act). Copyright law, as interpreted by the US Supreme Court, does not permit anyone to copyright facts, ideas or styles of expression, but only the “expression” of ideas. Thus, a modification of a painting for reprint would be considered a derivative work, whereas a painting in the same style as another would not. Unlike patented inventions, there is no exclusive right to “use” a work protected by copyright.

This raises the question of whether the use of Input Works by GAIs is a permitted use (because there is no exclusive right to “use” works protected by copyright) or constitutes preparation of a derivative work and therefore infringement of the Input Work: a question that is currently being litigated. For example, plaintiffs in Getty Images (US), Inc. v. Stability AI, Inc. argue that the unauthorized use of Input Works to train a GAI constitutes copyright infringement by the GAI because the GAI copies (at least) the data associated with the Input Works to process and “train” itself and then effectively prepares derivative works based on the Input Works or training data to generate Output Works. However, it is not entirely clear whether certain GAIs even copy the Input Works when analyzing them to identify patterns. Moreover, even if copyright-protected works are ingested in the training process, GAI providers may seek to rely on the fair use doctrine to use portions of copyright-protected works without agreeing to any license or other limitations. Whether or not fair use applies depends on a balancing of four statutory factors provided by the Copyright Act: the purpose of the use of the original work, the nature of the original work, how much of the original work is used, and potential harm to the market for the original work. In one case that may be relevant to the analysis for GAIs, fair use served as the basis for the Second Circuit’s holding in Authors Guild v. Google Inc., upholding Google’s copying of books to create a searchable database under the four factor analysis.

However, there is no settled case law or legislation that outlines the scope of the fair use defense precisely as applied to GAIs. Legislators, the United States Patent and Trademark Office (USPTO) and the United States Copyright Office (USCO) have expressed interest in establishing a cross-agency commission on intellectual property issues related to GAIs. Separately, the USCO plans to review copyright law and issues related to generative AI and comment later this year. The USCO has already issued some guidance regarding whether Output Works may be protected by copyright, which we have previously reviewed. While the question of whether Output Works may be protected by copyright is separate from understanding whether an Output Work infringes on the copyrights of any Input Works, the two issues each turn on the application of the fair use defense. Absent applicable legislation or regulation, the evolution of common law through litigation will draw the contours of the fair use defense and copyright law as applied to GAIs, the use of Input Works and the generation of Output Works.

There are several approaches to navigating potential copyright infringement by Output Works which lie on the spectrum of what should be considered “fair use”. In addition, GAI providers/defendants may argue that there was no copying and that the AI system only “used” the Input Work. Because the use of an underlying work is not an exclusive right of a copyright holder, one may argue that no infringement by the GAI or the user occurred (assuming that no copying of Input Works, among other potential infringement, occurred). Interested parties may advocate for either a quite limited or an expansive view of fair use, but the regulatory approach that ultimately prevails will shape the generative AI industry and the IP law that surrounds it.

Given this technological and legal backdrop, we present four broad schools of thought regarding the scope of the fair use defense that consider whether Output Works infringe Input Works. This article discusses these general viewpoints, when each viewpoint would consider an Output Work to directly infringe on any copyrights of Input Works and the challenges in that viewpoint’s implementation. The discussion below does not address the potential for vicarious infringement or inducement of infringement actions against the developers or operators of GAIs, but such topics are also important considerations for any legislature or court contemplating intellectual property issues related to GAIs.

Fair Use Minimalism

At its extreme, the fair use minimalist view considers all Output Works to be unoriginal and therefore derivative of Input Works, even if, because of the intervention of GAI, no Output Work is an exact copy of any single Input Work. Fair use minimalists would argue that the amount and substance of the Input Works that are copied and used to train GAIs cannot be considered “fair use” and, in addition, that such training does not constitute a permitted use but instead necessarily involves the preparation of works that are (perhaps necessarily) derivative of the Input Works used to generate Output Works.

Fair use also considers “the amount and substantiality of the portion used in relation to the copyrighted work as a whole.” Copying the creative “heart” of a work, or its most expressive and creative components, weighs against fair use, especially when multiple complete works are copied. To generate quality Output Works, fair use minimalists would argue that GAIs must analyze and use as much as possible from the underlying Input Works, including the most expressive or creative components of the works. Fair use minimalists contend that GAIs will generally copy and use the “hearts” of Input Works to generate an Output Work that seems like it could have been generated by a human author or, even if copying is not involved, will have analyzed and reviewed the Input Works so thoroughly that any Output Work would necessarily be a derivative work of the Input Works.

Moreover, specific enough prompt engineering, the process of manipulating prompts to generate desired outputs, may generate an Output Work designed to mimic the character and creative expression of a single author or work. As a result, fair use minimalists contend that the Input Works and Output Work cannot truly be separated, and the Output Works should be considered infringing derivative works. There is some nuance in this, of course, as the specificity of prompts varies substantially. Therefore, along the spectrum of fair use minimalism, some might believe that only with greater specificity in prompts should less leeway be given as to whether Output Works are infringing derivative works.

The implications of the fair use minimalism approach are expansive. First, all Output Works must necessarily infringe on the copyrights of authors of Input Works or unavoidably induce infringement of those copyrights. Thus, authors of Input Works would have some rights to remuneration from exploitation of any Output Works or otherwise to the exploitation of any Output Works. There is some case law that provides that copyright holders may use or exploit an unauthorized derivative work without compensation to the author of the derivative work (although the Supreme Court found no infringement or inducement of infringement when applied to companies manufacturing VHS systems). One possibility to avoid a tangled web of ownership rights is to develop a remittance scheme to compensate authors of Input Works for the training of GAIs and production of Output Works (perhaps in a similar manner to the remittance scheme for VHS systems that was ultimately struck down in the same case). Under this approach, legislators, agencies or the judiciary would also need to consider whether the user generating and/or exploiting an Output Work should be subject to liability for vicarious infringement and whether a GAI can induce copyright infringement.

Fair Use Maximalism

At the opposite extreme, fair use maximalism posits that the fair use defense should cover any (or almost any) Output Work because each is unique and created by a sufficiently transformative process. Unsurprisingly, this is the approach pushed by GAI providers, including OpenAI in a comment to the USPTO.

Fair use maximalists contend that each Output Work should be considered entirely distinct from any Input Work and relies very little on any single Input Work. Instead, fair use maximalists view AI models as tools, like a pencil or computer software, to create works. They argue that GAIs train on the expression of the Input Works just as human authors take inspiration from other works. In this view, general artistic ideas that can be derived from analyzing many Input Works and are embodied in an Output Work are no more based on Input Works than any author might generate by working on the same general idea or taking inspiration from the same general facts, ideas or styles. For example, the Impressionist painting school arose as one painter inspired another until eventually enough painters shared enough “inspiration” to create a common style. Thus, the fair use maximalists reason, because only the expression of the Input Works themselves can be protected by copyright and not the underlying idea, the generation process is sufficiently transformative. This means that the unique Output Work only embodies the idea of the Input Works and thus cannot infringe on any copyrights related to the Input Works.

The fair use maximalist approach reduces risks for GAI providers by effectively providing that Output Works do not infringe Input Works. As a result, the fair use maximalist approach would likely spur the development of GAIs trained on a wide array of Input Works – such as large language models and text-to-image models – and would be least likely to provide compensation for creators of Input Works. While this approach would limit the need for new copyright regulations, lawsuits that are in a similar vein to music copyright suits may develop to handle certain kinds of cases, such as where an Output Work copies too much of the underlying Input Works, where an Output Work is too similar to some Input Works that were used in the training or where prompts were too specific and result in an Output Work that is too similar to underlying Input Works.

Conditional Fair Use Maximalism

Sitting between the two extremes is what we call conditional fair use maximalism – an approach that evaluates an Output Work on a case-by-case basis to determine whether the fair use defense should apply. Under this approach, only works generated from sufficiently diverse Input Works that do not copy the “heart” of the expression contained in the Input Works should receive fair use protections. Thus, any infringement inquiry should focus on the sufficiency of the training set of Input Works and the specificity of the prompts made by the user used to refine or generate the Output Work.

The underlying idea behind this approach is that fair use protections should only apply to works that take no more than necessary to achieve a transformative purpose. For example, while each Output Work may be unique, the generation process can result in Output Works that are substantially similar to Input Works. As a training set gets larger and more diverse, the risk of the “heart” of any one Input Work being copied by the Output Work is mitigated. To ascertain the transformative purpose, the prompts used to generate an Output Work should be scrutinized to ensure the Output Work sufficiently transforms any components copied from any single Input Work. For example, fair use would not apply if the Output Work was generated from only minimal prompting and the prompts would clearly generate a work that substantially copies a single Input Work, a single author of multiple Input Works or a style that is representative of a very narrow set of Input Works. This directly implies looking for intent through evidence embodied in the prompts—one can more easily imagine the prompt “write a story about a teenage wizard prodigy in the writing style of J.K. Rowling” generating a derivative Output Work than the prompt “write a story about a teenage wizard prodigy” without the specific reference to an author.

In addition, a conditional fair use maximalist approach would be friendly toward an approach where an Output Work is generated based on Input Works that are mostly owned or controlled by the rightsholder with perhaps a small sampling of third-party Input Works. The underlying reasoning would be that the Output Work might benefit from some variety that derives from the third-party Input Works, but the overall style and content would be largely dictated by the contribution of the owned or controlled Input Works.

Under the conditional fair use maximalist approach, GAI users would need to document their creative process substantially to receive the benefit of fair use protections, or, alternatively, GAIs would have to build in screening mechanisms to ensure that prompts would not result in an Output Work that would be derivative of the underlying Input Works. This documentation requirement will go far beyond any similar or proposed regulations today. Given the international nature of the Internet, there is some risk that documentation requirements will become de facto global requirements. In addition, an anomalous consequence of this approach is that, when considering two substantially similar Output Works, one of them may be considered non-infringing while another would be considered infringing.

This approach would also require GAIs to conduct data audits to ensure large and diverse datasets that include only the necessary elements of works to serve their purpose. GAIs may reject prompts that would generate Output Works that would not be protected by fair use, especially if the GAI may be liable for inducing infringement. This may require a redesign of current large language models, such as OpenAI’s ChatGPT.

Governmental entities should consider the effects on the nascent generative AI industry when designing any tests that scrutinize the generative process of GAIs, in part because of the burdens that may be imposed on generative AI providers. Any presumptions in favor of or against fair use when conducting conditional fair use inquiries could easily impact the size and scope of GAIs moving forward. For example, a presumption against fair use when analyzing prompts used to generate works may result in a shift towards very generalized, prompt-heavy GAIs. In addition, questions regarding remittance structures, vicarious liability and inducement of copyright infringement may impact the development of GAIs.

Limited Fair Use Maximalism

The fourth approach is an offshoot from fair use maximalism that distinguishes between Input Works protected by copyright and not subject to license restrictions and Input Works protected by copyright but subject to license restrictions. Under this approach, GAIs would need to comply with the license restrictions of applicable works protected by copyright before using them as Input Works. For example, Wikipedia licenses the majority of its text to the public under two open-source license schemes. These licenses include terms that dictate the public’s ability to use Wikipedia text, including “share alike” provisions that require works that alter, transform or build upon Wikipedia works be distributed under the same, similar or compatible license schemes. Under limited fair use maximalism, Output Works generated from GAIs trained on Wikipedia articles would be subject to the same “share alike” provisions.

Limited fair use maximalism would largely maintain the status quo under the current regulatory regime but would require GAIs to perform audits to ensure compliance with all license restrictions associated with any Input Works.

The Path Forward Around the World

Questions about copyright and GAIs are being grappled with around the world, with different countries taking different approaches. Japan has taken a broad, fair use maximalist approach. This will likely lead to more GAI development in Japan, and indeed some commenters believe the goal of this expansive approach is to facilitate more Japanese interaction with western-style works and to open up the vast, global array of Japanese media for AI generation. Closer to the middle of the spectrum, while China has not taken any legislative approach, recent court cases point towards an inquiry-based conditional fair use maximalist approach. Similarly, Singapore permits the use of Input Works protected by copyright and data mining without requiring permission or a license from the copyright owner, and the United Kingdom is considering a similar regulation. Towards the other end of the spectrum, the European Union is weighing the most restrictive approach; the EU is considering regulations that would allow authors of Input Works to “opt-out” and exclude their works from GAI training sets and attach restrictions to their works published online that GAIs must abide by if the works are used for training.

While expecting uniformity in international intellectual property laws is unrealistic and there may be some benefit to each country being a laboratory of copyright experimentation, it is nevertheless important that the United States take an informed approach that balances the rights of human authors with the needs of a burgeoning GAI ecosystem. The interpretation of the fair use defense will play an important role in the ability to exploit generative AI technology worldwide. If the United States can establish a consistent philosophical approach to guide intellectual property regulations now and moving forward, the generative AI industry will flourish within the established guardrails, and the world may well follow the approach taken by the United States.

[wpdatatable id=13 table_view=regular]