Defining AI Image Generation Copyright

At the top of the “Ten Most Ridiculous Lawsuits of 2015” was Naruto v. Slater. In Naruto, the 9th Circuit court dismissed claims that a monkey who had seized the camera of a photographer owned copyright over the selfie that it subsequently took [1]. The court held that the monkey lacked statutory standing because the Copyright Act did not explicitly state that animals can file copyright infringement lawsuits [2]. At the time, Naruto seemed a trivial case, but the Monkey Selfie Trial would later dictate the future of copyright law. Naruto set an important precedent: a nonhuman could not gain copyright protection for its creations [3]. With the recent development of artificial intelligence (AI) enabling nonhuman identities to create thousands of unique human-like images in a few minutes, Naruto can be used to refuse copyright protection for AI-generated works. While the current consensus does not consider AI image generations eligible for copyrightability [4], AI-generated images should qualify for copyright protection based on a principle that I propose called: “beyond mere reproduction.”

To understand how to regulate AI, one must understand that AI is inherently human and mirrors human problem-solving and pattern recognition skills [5]. Most AI models are built with neural networks drawing inspiration from the human brain. Neural networks consist of layers of artificial neurons that communicate and carry information [6]. There are three types of modern AI image generators: diffusion models, generative adversarial networks, and transformer-based models. Diffusion models are defined as “a family of probabilistic generative models that progressively destruct data by injecting noise, then learn to reverse this process for sample generation.”[7] Provided with thousands of pieces of art as training data, the diffusion model learns how to replicate the techniques of artists [8]. Generative adversarial networks take advantage of two neural networks called the generator and the discriminator [9]. The generator, which is trained on art found online, attempts to replicate the training data after being trained and the discriminator provides feedback on how close the generated image is to the real image in the training set. Transformer-based generators divide the training data art into different chunks and determine the relationships between the chunks.

As artist Pablo Picasso once said: "good artists copy, great artists steal."[10] Artists, whether human or not, will inevitably draw inspiration from their predecessors. Progress in art is built upon the foundations laid by those who came before. Michelangelo was influenced by Greco-Roman art. A diffusion model drawing a Greek warrior takes similar inspiration. The diffusion model is an artist [11].

Modern U.S. copyright law is characterized by finding a balance between technological advances and the preservation of the founding principles set forward by the Founding Fathers. Article 1, Section 8 of the Constitution states that: “The congress shall have the 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.”[12] To receive copyright protection, a work must have an author. There is no official definition of author but historically legal constructs have identified “author” as a human capable of intent, creativity, and expression. This linking of authorship to personhood makes copyright anthropocentric.

One group of experts argued that “one could conceptualize alternate models of agency and accountability in the copyright domain that could possibly distance itself from the existing anthropocentric models.”[13] Although this may be a good idea, it is not necessary to enable AI-generated images to be copyrighted. To consider current AI models’ works as qualifying for copyright protection, one does not need to change copyright law to include the works of nonhumans. To prove copyrightability, an image must satisfy two requirements: fair use and sufficient originality.

There is a concern that AI image generators taking copyrighted images from online sources such as training data without permission is a violation of copyright law and that therefore outputs from AI image generators themselves cannot receive copyright protection. To prove copyright infringement, one must demonstrate “ownership of a valid copyright and copying of constituent elements of the work that are original.”[14,15] The “fair use” doctrine in Section 107 of the Copyright Act of 1976 allows the limited use of copyrighted material without permission. Its application considers four conditions: the purpose and character, the nature of the copyrighted work, the amount and substantiality of the work used, and the effect of the use upon the potential market [16].

In the landmark case Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994), the Supreme Court emphasized the importance of whether a new work supplants the original or if it adds something new and “transformative.” AI image generators do not reproduce exact copies; instead, they generate entirely new images. As Judge Souter, writing for the majority in Campbell, commented, “The more transformative the new work, the less will be the significance of other factors, like commercialism, that may weigh against a finding of fair use.”[17]

While the datasets of image generators may contain creative art, the AI’s objective is to learn the artistic designs and then create new art. As evidenced in Feist Publications Inc., v. Rural Telephone Service Co., only the creative selection, arrangement, or coordination of facts can be copyrighted, not the facts themselves [18]. AI image generators can be viewed as art students who memorize the relationships between different objects and images. The art students then apply their knowledge to create new art. Inputting a specific prompt into an image generator, known as prompting, will change the organization of facts. Sufficiently prompting an image generator would therefore satisfy the requirement of creative selection and arrangement.

The substantiality of the art used can outweigh the quantity of the copyrighted material used in the output. In Harper & Row v. Nation Enterprises, the Court found that even though a small portion of the copyrighted work was used, it was the essential part. This thus infringed upon copyright [19]. AI does not use copyrighted work as its essential parts, but instead, its models learn patterns from an extensive collection of different art pieces. Therefore, AI image generators pass the amount and sustainability test. Claiming that AI generators replicate an essential part of certain art pieces is unreasonable. This would mean that the artistic techniques used by the original artist that the AI generator replicates is copyrighted, thus implying that no new art pieces containing that specific art feature or technique, whether AI-generated or not, could be copyrighted.

Finally, the most contentious aspect of fair use is the market impact if everyone were to be able to copyright AI-generated art. In Sony Corp. of America v. Universal City Studios, Inc., the Court considered the market impact of copying television shows with VCRs, which caused potential hardship to the market [20]. There have been many artists who claim that AI image generators will render their jobs useless. A case filed by Kelly McKernan and other artists against Stability AI, a maker of the text-to-image generator Stable Diffusion, attests to the fact that “industry will be diminished to such a point that very few of us can make a living.”[21] I would argue the contrary for two reasons. AI image generators online cannot compete with tangible art pieces which are still in huge demand. Granting copyright status to AI image creations does not mean AI will replace human creators. Instead, AI and human creations could coexist, which could lead to collaborations where AI could enable human creativity. However, it should be noted that copyright should only be attributed to art if the art is sufficiently unique.

In the recent case Thaler v. Perlmutter, a man who owned an AI image generator was denied registration of an image by the U.S. Copyright Office stating that his AI-generated work “lack[ed] traditional human authorship” and that “copyright law is limited to ‘original intellectual conceptions of the author.’” This case was interpreted to deny copyrightability to AI image generators, but it left a big gap [22]. The court ruled that the Register acted correctly, but within the case, it is mentioned that the plaintiff “elaborate[d] on his development, use, ownership, and prompting of the AI generating software … implying a level of human involvement in this case entirely absent in the administrative record.” The plaintiff represented to the Register that the AI system had generated the work “autonomously” and that he had played no role in the creation. This was later proved untrue, but the decision of the Register was sufficient based on those facts [23].

AI image generators rarely generate images autonomously. The common use of AI image generators occurs via human prompting. Since generation via human prompting is the most commonly used aspect of an image generator, I will limit my argument to this type of usage. The resulting image of a unique human-prompted image should be considered unique [24].

Many websites, for example, zapier.com, focus specifically on prompting AI image generators. They provide a general structure for prompting with specific elements that when modified, can create different unique images. Courts have historically set the amount of originality very low. As evidenced in Feist Publications, Inc. v. Rural Telephone Service Co., 4999 U.S. 340, 346 (1991), “a work can be copyrightable even if it only expresses independent creation with a “modicum of creativity.”[25] As long as the prompting is unique, then an AI-generated image would qualify as a unique creation.

Thus, the case of Naruto v. Slater, which ruled that nonhumans cannot qualify for copyright, does not necessarily mean that AI image generators cannot receive copyright. It instead adds more nuance; AI image generators that create images without prompting cannot receive copyright; however, users who insert a sufficient prompt into a website should earn copyright protection. Suppose the situation in Naruto was slightly different: the monkey, instead of taking the picture itself, starts peeling its banana strangely and the photographer takes a picture of the banana and publishes it. In this case, the photographer would be allowed to copyright the photo.

We have seen debates on copyright throughout history with new technologies. In Burrow-Giles Lithographic Company v. Sarony, 111 U.S. 53 (1884), the Supreme Court addressed the issue of whether a photograph could be considered an “original work of art” and thus be eligible for copyrightability. The court ruled in favor of copyrightability for three reasons. First, photography was not included in the original act of 1802 due to the fact that photography did not yet exist, so adding in new technology for copyrightability status made sense. Secondly, photography demonstrated a sufficient amount of originality by a photographer. Third, the Copyright Clause of the U.S. Constitution allowed photographs to be classified as “writings” of an author [26]. AI-image-generated copyright evidently is not in any previous copyright laws because it did not exist until recently; the first large-scale AI image generators were designed in 2014. If used with prompting, an AI image generation shows the “existence… of thought, and conception on the part of the author.”[27] Finally, the photographer giving “visible form by posing [a person] in front of the camera, selecting and arranging the costume, draperies, and other various accessories…,”[28] is similar to a user writing prompts for an AI to generate.

Since I have proven that AI-generated images qualify for copyright protection, I will now define what is beyond mere reproduction. A generic prompt can lead an AI to create art that resembles its training data too closely, whereas a nuanced prompt can produce original results. The nuanced prompt will push the AI to generate unique images distanced from its training data. To be beyond mere reproduction, a prompt should not be generic and ordinary but must introduce combinations of elements that cannot be found in one specific piece of the training data. For example, when provided the following two prompts: “generate an Impressionist painting” and “generate an Impressionist painting with a man chasing a goat into the sunset.” AI produces a more unique output for the latter prompt which satisfies the condition of beyond mere reproduction.

We must accept the role of AI in our lives. When considering copyright for AI images, the U.S. Copyright Office must request both the image and the prompt to determine if an image qualifies for copyright. If the prompt is beyond mere reproduction, it should be copyrightable.

Bibliography

1. “The Top Ten Most Ridiculous Lawsuits of 2015,” Faces of Lawsuit Abuse, December 21, 2015, https://www.facesoflawsuitabuse.org/2015/12/the-top-ten-most-ridiculous-lawsuits-of-2015/.

2. “Naruto v. Slater, No. 16-15469 (9th Cir. 2018),” Justia Law, April 23, 2018, https://law.justia.com/cases/federal/appellate-courts/ca9/16-15469/16-15469-2018-04-23.html.

3. Ibid.

4. “Thaler v. Perlmutter et al, No. 1:2022CV01564 - Document 24 (D.D.C. 2023),” Justia Law, August 18, 2023, https://law.justia.com/cases/federal/district-courts/district-of-columbia/dcdce/1:2022cv01564/243956/24/.

5. See Alan Turing’s “Imitation Game”

6. I will spare the reader from a more in-depth explanation of neural networks. Please read https://arxiv.org/abs/1901.05639 for a comprehensive explanation.

7. Ling Yang et al., “Diffusion Models: A Comprehensive Survey of Methods and Applications,” arXiv, October 11, 2023, https://arxiv.org/abs/2209.00796.

8. O’Connor, Ryan. 2022. “Introduction to Diffusion Models for Machine Learning.” News, Tutorials, AI Research. May 12, 2022. https://www.assemblyai.com/blog/diffusion-models-for-machine-learning-introduction/.

9. Ian J. Goodfellow et al., “Generative Adversarial Networks,” arXiv, June 10, 2014, https://arxiv.org/abs/1406.2661.

10. Panasuk, Curtis. 2019. “Pablo Picasso on Creativity, ‘Good Artists Copy, Great Artists Steal.’” Creativity Workshops. March 6, 2019. https://creativityclasses.com/good-artists-copy-great-artists-steal/.

11. “Artist means the creator of a work of fine art.”

    “Artist Definition: 717 Samples,” Law Insider, accessed October 15, 2023, https://www.lawinsider.com/dictionary/artist#:~:text=Related%20Definitions&text=Artist%20means%20the%20creator%20of,creator%27s%20heirs%20or%20personal%20representatives.

12. “Origins and Scope of the Power,” Justia Law, accessed October 15, 2023, https://law.justia.com/constitution/us/article-1/50-copyrights-and-patents.html.

13. Taruna Jakhar and Dr. Hardik Parikh, “Artificial Intelligence Personhood: Beyond the Anthropocentric Approach,” Journal of Optoelectronics Laser, July 5, 2022, https://gdzjg.org/index.php/JOL/article/view/663.

14. Feist Publications, Inc. v. Rural Tel. Service Co., 499 U.S. 340, 111 S. Ct. 1282 (1991)

15. Arica Institute, Inc. v. Palmer, 970 F.2d 1067 (2d Cir. 1992)

16. U.S. Copyright Office, “U.S. Copyright Office Fair Use Index,” Copyright.gov, February 2023, https://www.copyright.gov/fair-use/#:~:text=Section%20107%20of%20the%20Copyright,may%20qualify%20as%20fair%20use.

17. “Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994),” Justia Law, March 7, 1994, https://supreme.justia.com/cases/federal/us/510/569/.

18. “Feist Publications, Inc. v. Rural Tel.. Serv. Co., 499 U.S. 340 (1991),” Justia Law, January 9, 1991, https://supreme.justia.com/cases/federal/us/499/340/.

19. “Harper & Row V. Nation Enterprises, 471 U.S. 539 (1985),” Justia Law, May 20, 1985, https://supreme.justia.com/cases/federal/us/471/539/

20. “Sony Corp. of America V. Universal City Studios, Inc., 464 U.S. 417 (1984),” Justia Law, January 17, 1984, https://supreme.justia.com/cases/federal/us/464/417/.

21. Jocelyn Noveck and Matt O’Brien, “Visual Artists Fight Back against AI Companies for Repurposing Their Work,” AP News, September 1, 2023, https://apnews.com/article/artists-ai-image-generators-stable-diffusion-midjourney-7ebcb6e6ddca3f165a3065c70ce85904.

22. https://www.natlawreview.com/article/judge-rules-content-generated-solely-ai-ineligible-copyright-ai-washington-report

23. “Thaler v. Perlmutter et al, No. 1:2022CV01564 - Document 24 (D.D.C. 2023),” Justia Law, August 18, 2023, https://law.justia.com/cases/federal/district-courts/district-of-columbia/dcdce/1:2022cv01564/243956/24/.

24. Provided is a mathematical explanation: let U(x) be a function that measures the uniqueness of an object x. The function will return a value greater than 0 if x has any uniqueness. Given an object a and its derived product b (b is some multiple h of a not equal to 0 - i.e. b = ha), then if U(a) > 0, then U(b) = U(ha) > 0

25. “Feist Publications, Inc. v. Rural Tel.. Serv. Co., 499 U.S. 340 (1991),” Justia Law, January 9, 1991, https://supreme.justia.com/cases/federal/us/499/340/.

26. “Burrow-Giles Lithographic Company v. Sarony, 111 U.S. 53 (1884),” Justia Law, March 17, 1884, https://supreme.justia.com/cases/federal/us/111/53/.

27. Ibid.

28. Ibid.