Regulatory Comments Before the U.S. Copyright Office Library of Congress In the Matter of Artificial Intelligence and Copyright

Before the
U.S. Copyright Office
Library of Congress

 October 30, 2023

Submitted by
Wayne T. Brough and Ahmad Nazeri

On Behalf of the
R Street Institute

The R Street Institute (R Street) is pleased to submit these comments on the Artificial Intelligence and Copyright Notice and Request for Public Comment to the U.S. Copyright Office. R Street is a nonprofit, nonpartisan public policy research organization (“think tank”) whose mission is to engage in policy research and outreach to promote free markets and limited, effective government. Our organization regularly advocates on copyright policy in a digital world. In this capacity, we welcome the opportunity to provide comments on the disruptive innovation brought on by the introduction of artificial intelligence (AI) and machine learning.

 Understanding the inherent tension between the principles of copyright and freedom of expression is paramount. The First Amendment declares, “Congress shall make no law… abridging the freedom of speech.” This stands in contrast to copyright article, U.S. Const. Art. 1, § 8, cl. 8, which mandates, “[The Congress shall have Power] To… promote the Progress of Science and useful Arts, by securing for limited Times to Authors… the exclusive Right to their respective Writings…”

 To navigate this dichotomy, we must adopt a narrow interpretation of copyright, ensuring it imposes minimal constraints on the freedom of expression. To this end, we believe that existing copyright law is suitably flexible to address ongoing concerns with respect to copyright and the implications of AI, and we urge prudence in the introduction of new copyright rules or regulations.

 Artificial Intelligence and Creative Arts

AI promises to bolster the American economy, amplify the capabilities of creatives and catalyze advancements in science and the arts.[1] Recognizing that contemporary copyright doctrine has evolved to embrace a myriad of transformative technologies is crucial. Existing copyright law is adeptly poised to address the genuine concerns of creators. Within the existing legal framework, the judiciary is ideally positioned to contextualize and apply these principles in the face of the multifaceted challenges presented by AI.

AI has great potential to solve complex math calculations, detect cancer early, create new art and much more. In light of these capabilities, some advocate for stricter copyright rules that could hinder AI’s potential and contradict the original purpose of copyright law—promoting “progress in science and useful arts”—by restricting freedom of expression.[2] Maintaining a balanced copyright framework is vital to keep the United States at the forefront of art and technology.

In the dynamic realm of AI, increasing copyright regulations could undermine the technology’s prospective advantages significantly. Tighter restrictions might impede AI’s growth by restricting access to necessary training data.[3] Such limitations risk diminishing AI’s capacity to address complex challenges, innovate across sectors and contribute to artistic endeavors. By constricting the data from which AI learns, we inadvertently limit the spectrum of expression and creativity it can manifest. Our comments are anchored in the belief that while safeguarding creators’ rights is paramount, it should not stifle innovation or encroach upon our foundational liberties. While the request for comments is extensive, R Street has focused on the questions we believe to be most significant with respect to our work on copyright and AI.

Question 8. “Under what circumstances would the unauthorized use of copyrighted works to train AI models constitute fair use?”

The copyright dichotomy between idea and expression is a foundational principle in copyright law.[4] This distinction clarifies that while the unique way something is expressed (its expression) can be copyrighted, the general idea or concept behind it cannot be. When it comes to training AI models, this principle becomes pivotal. If an AI model is trained using copyrighted materials, determining the purpose of the use becomes crucial. If the intention is merely to understand the foundational concepts or ideas presented by a copyrighted work (e.g., general themes, topics, knowledge), then it might be seen as fair use.[5] However, if the goal is to make the AI replicate or reproduce the specific way those ideas are articulated or expressed in the copyrighted work, then it could be a copyright violation. For example, when a machine learning (ML) system processes images of stop signs, its primary objective is to recognize the sign’s universal features, not to replicate the unique artistic nuances of each photograph. Similarly, a natural language processing system analyzing written content aims to understand general linguistic structures and patterns rather than to reproduce or capitalize on the distinct expressive qualities of the prose.[6]

The fair use doctrine should be interpreted in a manner that supports AI technology advancement. Allowing ML systems to train on comprehensive datasets, even if they include copyrighted works, serves the broader societal interest. Such access ensures that AI systems are safer, more accurate and unbiased. Conversely, restricting AI training to smaller, proprietary datasets due to copyright constraints can result in suboptimal, potentially biased AI outputs.

A pivotal consideration in this context is the concept of “fair learning.” This principle posits that if an AI’s primary objective in accessing a work is to learn from its non-copyrightable elements rather than to appropriate its copyrightable aspects, then such use should be deemed inherently fair under the first factor of the fair use analysis.

Question 8.5. “Under the fourth factor of the fair use analysis, how should the effect on the potential market for or value of a copyrighted work used to train an AI model be measured?”

In the realm of copyright law, the fourth factor of “fair use” analysis examines the impact of a copyrighted material’s use on its potential market value.[7] When applying this factor to the training of AI models, it is important to consider the original intent behind the creation of copyrighted works like books or photographs. Typically, authors or photographers do not create these works with the expectation of selling them for AI use. Therefore, training an AI model on such materials does not inherently detract from the work’s intended market or its value because AI learning models were never the target audience or consumer. In essence, the usage by AI does not replace the need for humans to purchase or access the original, nor does it directly compete with the core market intended by the copyright holder.[8] As such, even if the use aligns with the educational ethos of “fair learning,” it is essential to recognize that AI’s utilization is unlikely to undermine the creator’s ability to benefit from their work in its intended market.

Question 10 and Its Sub Questions. “If copyright owners’ consent is required to train generative AI models, how can or should licenses be obtained?”

The process of acquiring individual licenses for the vast amount of data required for AI training is not only burdensome, but also potentially unachievable. Navigating licensing agreements for each piece of data can give rise to various challenges, including disputes over rights and the logistical nightmare of ensuring compliance for extensive datasets.[9]

Moreover, the costs associated with obtaining these licenses could make AI projects excessively expensive, thus impeding innovation and hindering industry growth. This approach may render many AI-driven projects unattainable, particularly for smaller entities or researchers with limited resources.

Although direct voluntary licensing may appear feasible in certain creative sectors, implementing it across all sectors presents significant challenges. For example, while structures exist to protect works as unified entities in sectors like music, the complexities of obtaining licenses for each individual piece of content can give rise to numerous challenges. It would be impractical for a company to negotiate licenses with all copyright owners due to the large number of works. When contemplating the future of work, particularly in AI and data utilization, such a licensing approach could render many projects unachievable. Considering the vast amount of data required for AI training, obtaining licenses for every piece of copyrighted content can be burdensome and potentially unattainable.

While a compulsory licensing regime may appear to address the challenges posed by voluntary licensing, it is essential to consider the broader implications. Although there are existing compulsory licenses in certain areas of copyright, the evidence does not necessarily support the notion that these licenses promote creativity or fair use.[10] Establishing a compulsory licensing regime is neither desirable nor effective in promoting creative activity.

Question 13. “What would be the economic impacts of a licensing requirement on the development and adoption of generative AI systems?”

Introducing a licensing requirement for the development and adoption of generative AI systems would have profound economic implications.

• Barrier to Entry: Given the vast number of works an AI training dataset might need to use—and the fact that thousands or millions of individuals might own those works—obtaining licenses for all underlying content becomes a significant challenge. This could act as a barrier to entry for smaller companies or startups that lack the resources to negotiate and secure such licenses.
• Increased Costs: The process of identifying, negotiating and securing licenses for every individual piece of content in a dataset would be resource-intensive. These increased costs could be passed on to consumers or could deter companies from pursuing certain AI-driven projects altogether.
• Stifling Innovation: The sheer complexity and cost associated with obtaining licenses might discourage innovation. Companies might opt for safer, less ambitious projects to avoid potential copyright pitfalls, thereby limiting the advancement of AI technologies.
• Monopoly Concerns: Only large entities, like tech giants, that have the resources to navigate the licensing landscape or have already amassed vast amounts of data might be able to compete effectively in the AI space. This could lead to a monopolistic environment where only a few players dominate, thereby reducing competition and potentially stifling innovation.
• Economic Incentives for Litigation: Given the structure of copyright remedies, even small-value infringements can lead to lawsuits due to the potential for statutory damages. This could encourage opportunistic lawsuits, further increasing costs for AI developers.
• Potential Negative Outcomes: While broader access to data can help mitigate some of the negative outcomes associated with AI (e.g., biases), restricting access through licensing could exacerbate these issues. For instance, limited data access might hinder the ability of AI systems to be trained on diverse datasets and lead to biased outcomes.
• Impact on Broader Economy: The ripple effects of these challenges could extend beyond the AI industry. Reduced innovation in AI could slow advancements in sectors across the economy that rely on AI, such as health care, finance and transportation.

Question 18. “Under copyright law, are there circumstances when a human using a generative AI system should be considered the ‘‘author’’ of material produced by the system?”

As AI continues to evolve and integrate more deeply into our creative and professional processes, the lines between human-led creations and those generated by AI blur. Recognizing and understanding these different scenarios can provide clarity on how to approach this intricate issue.

• The Role of the Software Programmer: The original designer or programmer of an AI system provides it with foundational algorithms, logic and structures.[11] If the final output predominantly aligns with this initial framework without significant external influence, it can be argued that the programmer, having set the AI’s course, should be considered the author. Think of it as setting the initial conditions for a complex chain reaction—even if the subsequent events are unpredictable, they all stem from those initial conditions.
• Direct Influence by the User: In many scenarios, the end user or operator of an AI system plays a pivotal role in shaping its output. This can be achieved by selecting specific training datasets or by guiding the AI through a series of iterative commands or prompts. Here, the AI system is more akin to a musical instrument—while it has the potential to create, the direction, tone and final output depend heavily on the user. Thus, when the user’s influence is significant, the user can be considered the primary author.
• Collaborative Creation: Often, the final output is neither solely a result of the AI’s initial programming nor the user’s guidance. It is a symbiotic relationship in which the AI’s design and the user’s real-time inputs collaboratively shape the end product. This can be seen as a partnership or collaboration between the AI’s inherent capabilities and the user’s creative intent, resulting in both parties potentially being regarded as joint authors.
• AI Operates Autonomously: Conversely, there are times when AI, armed with its algorithms and learned behaviors, produces content with little to no human intervention.[12] In these cases, the output can be so unpredictable and detached from human input that claiming human authorship becomes challenging.

In the first three scenarios, we can consider a human as the author of an outcome work. There are instances in which AI serves purely as a tool, much like a canvas and brush for a painter. In these cases, the AI merely aids in translating the human’s specific vision into reality without adding its unique flair or direction. Even if the AI speeds up the process or simplifies complex tasks, the creative essence and primary direction come from the human. Thus, in such scenarios, the human remains the sole author.

In the final scenario, with AI operating autonomously, the concept of human authorship does not apply. Similarly, the use of copyright law to spur innovation does not apply in this case, making it difficult to justify a role for copyright law in such an instance. Indeed, it may be that such works fall within the public domain as stipulated by existing copyright law.

Question 19. “Are any revisions to the Copyright Act necessary to clarify the human authorship requirement or to provide additional standards to determine when content including AI-generated material is subject to copyright protection?”

There is no pressing need for revisions to the Copyright Act to address AI-generated content. The current framework is equipped to handle the challenges posed by AI and ML. As previously discussed, when there is significant human intervention and the output from the ML system aligns with the human’s intended creative vision, it can be protected under existing copyright provisions. Historically, humans have employed various technological tools to aid in their creative processes, and AI can be viewed as a continuation of this trend. Just as the use of these tools did not necessitate a change in copyright laws, the advent of AI does not inherently demand a revision. The key lies in discerning genuine human creative contributions from purely machine-generated outputs.

Question 20. “Is legal protection for AI-generated material desirable as a policy matter? Is legal protection for AI-generated material necessary to encourage development of generative AI technologies and systems?”

From our discussions, it is evident that while legal protection for creative works is essential, extending this protection to AI-generated material requires careful consideration. The primary concern is reconciling copyright protection with the broader principle of freedom of expression. While copyright aims to protect original works, it should not stifle the free flow of ideas—especially in the realm of AI and ML. Existing copyright protection for computer code does offer some incentives for the development of generative AI technologies.[13] However, the focus should be on ensuring that these protections do not hinder the broader goals of innovation, education, and promoting science and art. In essence, AI and ML utilize the “idea” behind a work rather than replicating its unique “expression,” which is what copyright traditionally protects.

Question 20.1. “If you believe protection is desirable, should it be a form of copyright or a separate sui generis right?”

The current copyright act is sufficiently equipped to handle the challenges posed by AI-generated content. Introducing a separate sui generis right would only complicate the landscape further. The existing framework of the Copyright Act—if interpreted and applied judiciously—can address the nuances of AI creations. While it is essential to protect creators and innovators, it is equally crucial to uphold the broader goals of freedom of expression, education and innovation. When approached with this balanced perspective, the current copyright act can serve these dual purposes without additional rights or protections.

Question 21. “Does the Copyright Clause in the U.S. Constitution permit copyright protection for AI-generated material?”

The Copyright Clause in the U.S. Constitution aims to “promote the progress of science and useful arts.” While it does not explicitly address AI-generated material, the spirit of the clause is to foster creativity and innovation. Protecting AI-generated content could arguably align with this goal, but only if done judiciously. As emphasized throughout our comments, while protection is essential, it is equally crucial to ensure that works remain accessible for future AI and ML endeavors under the Fair Use doctrine. Such accessibility ensures that AI and ML can continue to evolve without the threat of infringement. By interpreting copyright narrowly, in line with Article 8 of the Constitution, we can strike a balance that upholds freedom, education and innovation. This approach respects the essence of the Copyright Clause while ensuring that AI and ML can operate within a framework that recognizes their unique nature and potential.

Conclusion

How can we navigate the complex and ever-changing intersection of copyright law and AI in a way that protects creators, encourages innovation and upholds fundamental liberties like freedom of expression? While the current copyright system can handle many AI-related challenges, it is necessary to approach this issue with a nuanced perspective to ensure that protections do not hinder the broader objectives of freedom of expression, education and innovation. We must find a middle ground that safeguards original works while allowing for the exchange of ideas. It is important to interpret the Copyright Clause in the U.S. Constitution, which aims to promote the progress of science and arts, in a manner that acknowledges the transformative potential of AI. However, we should always interpret copyright narrowly to minimize its impact on freedom of expression—a fundamental right that predates any government and is superior to copyright.

As stated in Article 8 of the U.S. Constitution, copyright exists to serve a specific goal, and expanding its scope would undermine both freedom of expression and the philosophy behind copyright itself. Expanding the fair use doctrine can play a significant role in mitigating the damage caused by copyright to freedom of expression—especially in the context of AI and ML.

Respectfully submitted,

Wayne T. Brough
Policy Director, Technology and Innovation

The R Street Institute
1411 K Street NW, Suite 900
Washington, D.C. 20005
202-525-5717
[email protected]

[1] Jacques Bughin et al., “Skill shift: Automation and the future of the workforce,” McKinsey Global Institute, May 23, 2018, pp. 3-84. https://www.mckinsey.com/featured-insights/future-of-work/skill-shift-automation-and-the-future-of-the-workforce.

[2] Amanda Beshears Cook, “Copyright and Freedom of Expression: Saving Free Speech from Advancing Legislation,” Chicago-Kent Journal of Intellectual Property 12:1 (2013); see also Craig W. Dallon, “The Problem with Congress and Copyright Law: Forgetting the Past and Ignoring the Public Interest,” Santa Clara Law Review 44:2 (2003), pp. 365-455.

[3] Mark A. Lemley and Bryan Casey, “Fair learning,” Texas Law Review 99:4 (2020), pp. 743-785. 

[4] Embodied in 17 U.S. Code § 102(b) – Subject matter of copyright: In general; see also Edward Samuel, “The Idea-Expression Dichotomy in Copyright Law,” Tennessee Law Review 56 (1988), p. 321.

[5] CFR § 202.1 – Material Not Subject to Copyright; see also Lemley and Casey, supra, note 3.

[6] Lemley and Casey, supra, note 3, p. 749.

[7] Pierre N. Level, “Toward a Fair Use Standard,” Harvard Law Review 103:5 (1990), p. 1124.

[8] Lemley and Casey, supra, note 3, p. 776; See also Bill Graham Archives v. Dorling Kindersley Ltd., 448 F.3d 605, 614-15 (2d Cir. 2006).

[9] Id. at p. 770.

[10] Jason S. Rooks, “Constitutionality of Judicially-Imposed Compulsory Licenses in Copyright Infringement Cases,” Journal of Intellectual Property Law 3:1 (October 1995), pp. 255-276. https://digitalcommons.law.uga.edu/cgi/viewcontent.cgi?referer=&httpsredir=1&article=1112&context=jipl.

[11] Margot E. Kaminski, “Authorship, Disrupted: AI Authors in Copyright and First Amendment Law,” U.C. Davis Law Review 51 (2017), p. 597. https://scholar.law.colorado.edu/cgi/viewcontent.cgi?article=2193&context=faculty-articles.

[12] Id. at p. 598.

[13] Megan Svedman, “Artificial Creativity: A Case Against Copyright for AI-Created Visual Artwork,” IP Theory 9:1 (2020), p. 13. https://www.repository.law.indiana.edu/cgi/viewcontent.cgi?article=1053&context=ipt.