INTRODUCTION

Each year, the United States Copyright Office (USCO) receives around half a million copyright claims, and typically, only a very small percentage of these claims are rejected.¹ Recently, controversial claims involving works created by artificial intelligence (AI) have comprised some of these rejections. AI raises an obvious but an interesting legal question: are AI-generated works copyrightable? I argue that the USCO’s application of existing precedent, namely that rooted in the Supreme Court case Burrow-Giles Lithographic Co. v Sarony, to copyright claims involving prompt-based, AI-generated works of visual art is misguided and even myopic insofar as it inappropriately applies the human authorship test and neglects the standards of originality that exist under other judicial precedents and may prove more apposite. 

To make this argument, I will first begin with an analysis of the USCO’s interpretation and application of Sarony in copyright claim applications involving AI-generated visual art. I will then point out the faults in this approach, arguing that AI-generated art produced from creative prompting satisfies the human authorship requirement under Sarony. I will then transition to a discussion of originality standards stipulated under Feist Publications, Inc. v. Rural Telephone Service Co. and Bleistein v. Donaldson Lithographing Co. Finally, I will discuss the failure of AI-generated works to meet these standards, concluding that AI-generated visual works, even when prompt-based, are not copyrightable.

LEGAL HISTORY

A brief discussion of the standards for copyrights is in order. Broadly speaking, there are two criteria by which copyright claims are adjudicated: authorship and originality. Both have roots in the Intellectual Property Clause of the Constitution: “[The Congress shall have Power …  ] To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.”² Here, the Constitution implies the authorship requirement through the phrase “Authors and Inventors” and the originality requirement through the granting of an “exclusive” right to “their respective” works, those being the works of said “Authors and Inventors.”

Case precedent has since reinforced and formalized these requirements, especially Burrow-Giles Lithographic Co. v. Sarony.³ In this case, the Supreme Court considered whether Napoleon Sarony’s photograph of Oscar Wilde was eligible for copyright protection. The defense argued that, since the photograph was not a work of original authorship, it was ineligible. In the majority opinion, the Supreme Court, citing Worcester v. Georgia,⁴ expanded the definition of author to “he to whom anything owes its origin; originator; maker.”⁵ They also concluded that photographs, in light of this clarified definition of “author,” are copyrightable “so far as they are representative of original intellectual conceptions of the author.”⁶ In regard to the specific facts of the case, they concluded that Sarony’s photo was “an original work of art, the product of plaintiff’s intellectual invention, of which plaintiff is the author.”⁷ Writing on behalf of the Court, Justice Miller concluded his opinion by stating “These views of the nature of authorship and of originality, intellectual creation, and right to protection, confirm what we have already said.”⁸ In essence, they applied a test of authorship and one of originality when deciding whether Sarony’s photo was copyrightable. The USCO, as will be discussed, heavily relies on this precedent when considering AI-generated works, including those that are based on human prompting. 

USCO’S STANDARD AND ITS APPLICATIONS

Though claims involving works of AI are novel and relatively infrequent, the USCO has apprised copyright seekers that works where authorship is not primarily human are not eligible for copyright protection. For example, in 2018, David Slater, a wildlife photographer, claimed copyright to selfies that were taken by monkeys on his camera equipment. In response, the nonprofit People for the Ethical Treatment of Animals (PETA) sued Slater, arguing that Naruto, one of the monkeys who took a photograph, was the rightful copyright owner. Judge Carlos Bea dismissed the case, ruling that only humans have the ability to claim copyrights. He explained that “this monkey—and all animals, since they are not human—lacks statutory standing under the Copyright Act.”⁹ Following this peculiar case, the USCO released an updated third edition to Compendium of U.S. Copyright Office Practices, which stated that they “will refuse to register a claim if it determines that a human being did not create the work,”¹⁰ citing Sarony and emphasizing the human authorship standard. After explicitly mentioning “photographs taken by monkeys,” an obvious reference to Naruto v. Slater, they add, “Similarly, the Office will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author,”¹¹ again invoking Sarony.

Indeed, in 2023, the USCO applied these sections from the compendium to a novel case involving an AI-generated work titled “A Recent Entrance to Paradise,” the creator of which was an AI software developed by Stephen Thaler. Although Thaler admitted that the piece was entirely generated by AI, he claimed that it was copyrightable under the works made for hire provision, which states that the party who hired an author is entitled to the copyright for the work that is created. In response to Thaler’s appeal to their rejection, the USCO, relying partly on Sarony yet again, insisted that his work was not copyrightable because it did not represent a work of human authorship.¹² Thaler then filed a lawsuit. Ultimately, the United States District Court rejected his claims and affirmed the USCO’s rejection of his copyright application on the same grounds. In fact, the majority opinion stated that the “defendants are correct that the Copyright Office acted properly in denying copyright registration for a work created absent any human involvement,” but also acknowledged that there are “new frontiers” that will need addressing as AI technology continues to grow. More specifically, “how much human input is necessary to qualify the user of an AI system as an ‘author’ of a generated work?”¹³

One such frontier-pushing complication is prompt-based generative AI software, which unlike in Thaler’s case, involves a degree of human input. A simplified version of the process of AI art creation is as follows: AI essentially learns to recognize patterns and associations from a database of pre-existing and often already copyrighted material, a process called “training.” This material is clouded by “noise,” essentially obscuring the original pieces. Once made available to the public, a user enters a prompt, and the software searches the database for relevant, “noisy” images. It then progressively attempts to remove the “noise” and achieve the original input image.¹⁴ Depending on the prompt, this might produce output that closely resembles the original input, or it might produce a derivative work that combines several “denoised” outputs. 

The question then becomes whether this prompting constitutes sufficient human involvement to be considered human authorship. Addressing this uncertainty, the USCO issued a statement of policy in March of 2023 titled “Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence,” which outlines their position on the copyrightability of AI-generated works. To contextualize their conclusion, the USCO relies largely on Sarony to outline the human authorship test before applying it to prompt-based AI.¹⁵ Then, they contend that even with prompt-driven software such as Midjourney, “users do not exercise ultimate creative control over how such systems interpret prompts and generate material.” Instead, “the machine determines how those instructions are implemented in its output.”¹⁶ As a result, they propose that there is not enough human input to meet the human authorship requirement as clarified under the Sarony decision, rendering all prompt-driven AI art non-copyrightable.

Subsequently, they applied this reasoning when evaluating the artist Jason Allen’s request to register “Register Théâtre D’opéra Spatial,” a digital painting generated by Midjourney based on his prompting. The USCO rejected his claim based on reasoning from the Thaler v. Perlmutter, Sarony, and Slater decisions.¹⁷ Allen challenged their decision, arguing that his prompting constituted sufficient human authorship, but USCO found these contentions “unconvincing” and maintained their position that the image was not copyrightable.¹⁸ They contend that “the features generated by Midjourney and Gigapixel AI must be excluded as non-human authorship,”¹⁹ concluding “that the Midjourney Image, which remains in substantial form in the final Work, is not the product of human authorship.”²⁰ Once again, their definition of “human authorship” is derived from Sarony. In effect, this decision and the statement of policy referenced above, when viewed in conjunction with their compendium, sufficiently elucidate the USCO’s stance on prompt-based AI– even with human prompting, generative AI works do not meet the human authorship requirement as described by Sarony.

PROBLEMS WITH USCO’S PROCESS: THE AUTHORSHIP STANDARD

This application of Sarony to prompt-based AI, however, is faulty. In Sarony, the Supreme Court ruled that photographs, provided that there is sufficient human authorship, are copyrightable. Even though the camera is the tool that is used to capture a moment, the photographer can still influence what is captured in a way that reflects their “original intellectual conceptions.” For Sarony’s photograph, this involved the following: 

…posing the said Oscar Wilde in front of the camera, selecting and arranging the costume, draperies, and other various accessories in said photograph, arranging the subject so as to present graceful outline, arranging and disposing the light and shade, suggesting and evoking the desired expression, and from such disposition, arrangement, or representation, made entirely by plaintiff, he produced the picture in suit.²¹

In the case of prompt-based AI, a human’s input can entail writing a prompt with details about what they wish to be contained in the final product. An AI prompt not only specifies the subject matter but may also include instructions about colors, lighting, background, and other artistic features. The user of AI has an intellectual conception of what they wish to portray and uses a set of phrases to creatively prompt the software to produce an image akin to their mental vision. Thus, when the prompt does include details representative of a creative vision, the process of AI prompting matches the extent of Sarony’s artistic input, qualifying these generative-AI works for copyright under the human authorship test.

Not only this, but the USCO’s argument that AI prompters do not have “ultimate creative control”²² over what the AI portrays and cannot therefore be authors is unsound. The Sarony majority opinion makes no mention of “ultimate creative control.” In fact, the very permittance of copyrights for photographs suggests that this is not a requirement for authorship. Consider the role of the photographer and their use of a camera. The photographer merely captures an image of something that already exists, and as a result, their control is limited to arrangement and alteration. Unlike a traditional painter, they are not creating the subject matter in the most literal sense. Thus, the photographer lacks “ultimate creative control” in the same way that an AI prompter does, yet photographs remain copyrightable under Sarony while the USCO claims that exact case prohibits AI-generated works from being granted protection. Thus, given the equivalent degree of creative human input between copyrightable photographs and AI-generated works, it seems that Sarony should permit the copyrighting of prompt-driven AI works under the human authorship test.

ANOTHER AVENUE?: THE ORIGINALITY STANDARD

Though the human authorship test’s applicability to prompt-based, AI-generated works has been vitiated, the originality standard, though it remains largely underutilized to date, may prove more relevant. We should first clarify the criteria that the originality standard presents, especially given its tumultuous history. As the legal writer Howard Abrams puts it in “Originality and Creativity in Copyright Law,” “the question of originality, the threshold standard of qualification for copyright protection, is at the core of copyrightability.”²³ Abrams continues to summarize the foundations of the originality requirement, “Based on The Trade-Mark Cases and Burrow-Giles Lithographic Co. v Sarony, the Court held that originality is a constitutional requirement arising out of the use of the terms ‘Authors’ and ‘Writings’ in the Copyright Clause.”²⁴ He then summarizes the significance of the decision in Feist Publications, Inc. v. Rural Telephone Service Co.,²⁵ which creates a “two step analysis for issues of originality.”²⁶ First, there must be “the traditional originality inquiry into whether a work was independently originated,” and second, there “must also be a determination that whatever was independently originated was sufficiently creative to satisfy Feist.”²⁷

 In referencing Feist, he draws parallels to Bleistein v. Donaldson Lithographing Co., in the majority opinion of which Justice Holmes wrote:

The copy is the personal reaction of an individual upon nature. Personality always contains something unique. It expresses its singularity even in handwriting, and a very modest grade of art has in it something irreducible, which is one man’s alone. That something he may copyright unless there is a restriction in the words of the act.

Under this ruling, it is necessary that a work reflects some uniquely personal expression of self. This, in effect, adds a third criterion to the originality requirement: a degree of original personal expression.

As I will elucidate, AI-generated works are not independently originated, sufficiently creative, or personally expressive, thereby failing the originality standard for copyrightability. Since prompt-based AI-generated visual content possesses the strongest claim to copyright given the human input, any failure to meet originality standards would, a fortiori, render all AI visual content uncopyrightable. Thus, in what follows, I will look almost exclusively at works generated by prompt-based AI. 

The very processes involved in generative AI technology, particularly training and diffusion models, clearly demonstrates the failure to meet the multifarious originality requirement. Recall that generative AI works by combining and restructuring multiple different pre-existing images. Consequently, the AI-produced art is not independently originated as it is simply an attempt at reconstruction of original input material. In addition, though the prompting constitutes creative human input as discussed above, the image that is produced by the AI is not sufficiently creative to satisfy the “minimal degree of creativity” demanded by Feist. Given the mechanical nature of the process, moreover, AI-generated material lacks any meaningful expression of human personality, failing to meet the originality requirements set forth by Bleistein.

Recent court filings on the matter of generative AI image copyright infringement provide evidence supporting these claims. Getty Images, for example, recently initiated a lawsuit against Stability AI, Getty Images (US), Inc. v. Stability AI, Inc., alleging that the training process represented “brazen infringement of Getty Images’ intellectual property on a staggering scale.”²⁸ The ECF 1 complaint for this lawsuit outlines the extent to which major generative AI programs rely on copyrighted material. In this particular case, Getty Images uncovered several AI-generated images that contained the Getty Images watermark. Concretely, “Stability AI has copied at least 12 million copyrighted images from Getty Images’ websites, along with associated text and metadata, in order to train its Stable Diffusion Model.”²⁹ In many cases, as Getty Images proves with examples of Stability AI images that very closely resemble Getty Images’ works and contain the Getty Images watermark. In others, it generates new, derivative works, but “only because it was trained on proprietary content belonging to Getty Images and others.”³⁰ The particular example provided is that a user may ask for a cat wearing a scarf.³¹ In such cases, AI programs denoise and combine two related input images to create a synthesized product.

Here, a proponent of generative AI might posit that these images represent creative compilations, which under Feist are copyrightable. The relevant part of Justice O’Connor’s majority opinion reads:

The compilation author typically chooses which facts to include, in what order to place them, and how to arrange the collected data so that they may be used effectively by readers. These choices as to selection and arrangement, so long as they are made independently by the compiler and entail a minimal degree of creativity, are sufficiently original that Congress may protect such compilations through the copyright laws.³²

The process of diffusion, however, is highly automated and mechanical. The AI user inputs a prompt, which while it likely constitutes human authorship, does not involve direct and creative selection of works. The algorithm, not the actual author (in this case the user of AI), selects the works that are compiled into one. Thus, since the author is not choosing and arranging the inputs, Feist’s permittance of creative compilations does not apply.

CONCLUSION

Why bother restricting copyrights to non-AI generated images? This complex legal minefield of extending copyright law to AI created art might seem superfluous without a clear demonstration of the dangers AI pose for human content creators. Yet the answer is simply that AI runs amok of existing copyright claims, as was discussed with the Getty case and that AI art seems to lack an essential aspect of creativity that serves as the backbone of human creation and the industries that follow. USCO’s intentions are rightly placed, but their monomania with the human authorship test may prove problematic in the barring of AI generated images from copyright. As discussed, prompt-driven AI generated images involve creative human input and represent “original intellectual conceptions,” requirements stipulated under the Sarony precedent. That said, AI software, especially the underlying models of training and diffusion, do not produce independently originated, sufficiently creative, and personally expressive works. Instead, AI software utilizes copyrighted material and draws heavy influence from them in order to create images that are essentially just mechanically produced collages of pre-existing material. The application of originality standards can better protect the integrity of intellectual property in comparison to the human authorship test.

1. “Annual Report: FY 2021,” United States Copyright Office, 2021, 8, https://www.copyright.gov/reports/annual/2021/ar2021.pdf. ↩︎
2.  U.S. Const art. I, § 8, cl. 9 ↩︎
3. Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53, (United States Supreme Court 1884). ↩︎
4.  Worcester v. Georgia, 31 U.S. 515 (1832) ↩︎
5.  Burrow-Giles ↩︎
6.  Ibid. ↩︎
7.  Ibid. ↩︎
8.  Ibid. ↩︎
9. Naruto v. Slater, 888 F.3d 418, (9th Cir. 2018). ↩︎
10. United States Copyright Office, Compendium of U.S. Copyright Office Practices (2021), §306. ↩︎
11. United States Copyright Office, Compendium of U.S. Copyright Office Practices (2021), §313.2. ↩︎
12. U.S. Copyright Office Review Board, “Re: Second Request for Reconsideration for Refusal to Register A Recent Entrance to Paradise,” February 14, 2022, 4, https://www.copyright.gov/rulings-filings/review-board/docs/a-recent-entrance-to-paradise.pdf ↩︎
13. Thaler v. Perlmutter, F.Supp.3d, (U.S. District Court, District of Columbia, 2023). ↩︎
14.  See Jascha Sohl-Dickstein’s “Deep Unsupervised Learning Using Nonequilibrium Thermodynamics” for further reference. ↩︎
15. U.S. Copyright Office, “Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence,” Federal Register 88, no. 51 (March 16, 2023): 16191, https://www.govinfo.gov/content/pkg/FR-2023-03-16/pdf/2023-05321.pdf. ↩︎
16.  Ibid., 16192. ↩︎
17.  U.S. Copyright Office Review Board, “Re: Second Request for Reconsideration for Refusal to Register Théâtre D’opéra Spatial,” September 5, 2023, https://fingfx.thomsonreuters.com/gfx/legaldocs/byprrqkqxpe/AI%20COPYRIGHT%20REGISTRATION%20decision.pdf. ↩︎
18.  Ibid., p. 7. ↩︎
19. Ibid., p. 2. ↩︎
20.  Ibid., p. 5. ↩︎
21.  Burrow-Giles Lithographic Company v. Sarony, 111 U.S. 53 (1884) ↩︎
22. U.S. Copyright Office, “Copyright Registration Guidance,” 16191. ↩︎
23. Howard Abrams, “Originality and Creativity in Copyright Law,” Law and Contemporary Problems 55 (1992): 4, https://scholarship.law.duke.edu/cgi/viewcontent.cgi?article=4136&context=lcp#:~:text=The%20Supreme%20Court’s%20widely%20cited,53%2C%2058%20(1884). ↩︎
24. Ibid., 11. ↩︎
25. Feist Publications, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340 (1991) ↩︎
26.  Ibid., 14. ↩︎
27. Ibid. ↩︎
28.  Getty Images v. Stability AI, Case No. 23-135, (U.S. District Court for the District of Delaware) February 2, 2023, ECF Complaint, 1. ↩︎
29.  Ibid., 3. ↩︎
30.  Ibid., 17. ↩︎
31.  Ibid. ↩︎
32. Feist Publications, Inc. v. Rural Telephone Service Co., Inc., 499 U.S. 340, (United States Supreme Court 1991). ↩︎