Theatre D’opera Spatial should be copyrightable because it is creative, but current U.S. Copyright Law prevents it from being registered.
New tools such as generative artificial intelligence models combined with greater computer processing power have made Some examples include DALL-E, Stable Diffusion, and Midjourney. Check out this video by MKBHD, one of my favorite technologists, to get a glimpse into how these tools work and how powerful they are. (You typically have to pay to use any of these models on your own.)
What is Theatre D'opera Spatial?
Theatre D’opera Spatial is a digital, two-dimensional work of art that was created by Jason Allen. Mr. Allen used multiple tools to create this work. To begin he used a text to image generator called Midjourney. To get the base image he used at least 624 text prompts which Midjourney translated into a two-dimensional image. He then used Gigapixel AI to upscale the picture. Upscaling a picture allows the artist to make the image more crisp, smooth out edges, and create more detail in an image. (https://www.topazlabs.com/gigapixel-upscale-ai-generated-art) Mr. Allen also used Photoshop to remove flaws and create new visual content within the image. Theatre D’opera Spatial should be copyrightable because it is creative, but current U.S. Copyright Law prevents it from being registered.
Figure 1 - Output from Midjourney
Figure 2 - Theatre D'opera Spatial (finished work)
Jason Allen filed an application for copyright of his work in September of 2022. He did not disclose on his application for copyright that he used AI tools to create the work. The Copyright Office found out because the work had gained notoriety after it won first place in the Colorado State Fair Fine Arts Competition in the digital arts category. The Copyright Office rejected Allen’s application based on the grounds that it was lacking human authorship. (https://fingfx.thomsonreuters.com/gfx/legaldocs/byprrqkqxpe/AI%20COPYRIGHT%20REGISTRATION%20decision.pdf) However, I think this is the wrong argument to make. This work is composed of 624 text prompts, upscaling, and refinement of the image, all certainly rise to the same level of a modicum of creativity that Sarony exhibited in his photograph of Oscar Wilde. The tools have changed, but this is still human authorship and certainly is a creative work worthy of copyright protection.
Setting the Technological Playing Field
Before we dive into our analysis, let us start with a few key terms to ground our understanding of the technology.
“An algorithm is any well-defined computational procedure that takes some value, or set of values, as input and produces some value, or set of values as output.” Introduction to Algorithms, Second Edition, Cormen, p. 5, 2001. To put it simply, an algorithm is a sequence of computational steps that transform the input into the output. Artificial Intelligence is a field of computer science, that is attempting to replicate rational thinking, meaning we are trying to get the computer to “do the right thing.” Artificial Intelligence: A Modern Approach, Fourth Edition, Russell, p. 1. 2021. Machine learning is a subset of artificial intelligence. Machine learning can be characterized as a computer which builds a model based on input given to it. Then the computer used that model to create new outputs. Id. at 651. A model is a mathematical abstraction that has linked the concept of truth to a particular value or set of values. All inputs are transformed into outputs based on this truth value. Id. at 214. Deep learning is a subset of machine learning where there are multiple layers of neural networks through which inputs are fed and outputs are obtained. Each layer represents a different level of (hierarchical-) learning by the model. Deep Learning with Python, Second Edition, Chollet, p. 7. 2021. Convolutional Neural Networks (CNNs) are one kind of implementation of deep learning that utilize the mathematical concept of a convolution or cross-correlation (signal processing). CNNs are characterized by patterns of weights (kernals) which are replicated across each layer. Russell, p. 760. Traditionally, in image processing, kernals were hand-engineered. With CNNs, the kernel weights are learned, stochastically, from the training data. An Introduction to Statistical Learning,Gareth James, p. 409. 2023.
Midjourney, the tool used by Mr. Allen, is considered a generative adversarial network (GAN), a deep learning architecture that features a pair of CNNs which combine to generate an output. Russell, p. 780. They are trained at the same time and at some point the model has completed training and is defined. The final trained GAN is deterministic, meaning each input will correlate to a definite outcome. So even though the model is generating the image, with respect to just the models, the same image can be obtained by another user by using the exact same input or it can be mathematically re-created by stepping through each layer of the networks.
Now, for a bit of literary whiplash, let us discuss creativity within the US Copyright Law.
Creativity Standard –modicum of creativity
The Copyright Act contains the formula for what works can be considered eligible for copyright protection. These are “original works of authorship fixed in any tangible medium of expression, now known or later developed, from which they can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device” 17 U.S.C. § 102(a). There are two parts of this definition we will examine as part of our analysis for Theatre D’opera Spatial, “original works of authorship,” and “with the aid of a machine or device.”
The meaning of the phrase original works of authorship has been developed through case law since it was adopted in 1976. The precedent-setting case is Feist Publications v. Rural Telephone Service Co. from 1991. This case was about phonebooks. Rural published a phonebook for a particular geographical area. Feist published a phonebook for a larger geographical area. Rural sued Feist for using the content of Rural’s white pages in Feist’s publication. The Supreme Court ruled that although there was a selection of what to include in the phone book, listing out facts in alphabetical order does not rise to a level of originality that could elicit copyright protection. Originality requires independent creation plus a modicum of creativity.” Feist, 499 U.S. 346.
From Feist, we have the precedent that originality requires a modicum of creativity. Meriam Webster defines modicum as “a small portion; a limited quantity.” Therefore, the amount of creativity that is required is quite small. In fact, the Supreme Court found that there was modicum of creativity in the geometric shapes featured on cheerleader uniforms. Varsity Brands, Inc. v. Star Athletica, LLC, 580 U.S. 405. That case turned on the fact that the design elements of the uniforms, we severable from the article of clothing itself. I think this distinction is relevant here because it shows that even the smallest design elements, such as geometric shapes on cheerleading uniforms are seen as worthy of copyright protection.
As a note to the global nature of these topics. I think it is important to remember that this standard for creativity is unique to the United States. Neither the Berne Convention nor the TRIPS Agreement have a requirement for creativity.
Photography Precedent – Sarony
There is a long history which establishes photographs as copyright eligible. When the camera was first invented, many people thought that photographs were not a worthy art form. The precedential case in this area is Burrow-Giles Lithography Co. v. Sarony from 1884. Sarony was the photographer of a portrait of Oscar Wilde. Burrow-Giles Lithography Co. reproduced the portrait without Sarony’s permission. The Supreme Court ruled that when the author exercises control over the subject matter, rather than just performing a mechanical process, the photo is protected under the copyright statues. Sarony was credited with selecting the scene of the photograph, the pose, clothing, and expression of the subject, exhibiting creativity and purposeful selection in the composition of the photograph. All of these things added up to authorship and warranted copyright. One of the issues in this case was the fact that photographs were not originally included in the copyright statute. However, the court acknowledged that photographs were not included because they had not yet been invented.
Conclusion
In the rejection letter from the Copyright Office, they offered that if Mr. Allen disclaimed the elements which were generated using AI from the copyright, the remaining work could be registered. Under the current law, this could be acceptable. However, I think it is difficult to defend the position that Mr. Allen did not control the creative elements used to compose Theatre D’opera Spatial. The elements of generating the original image, upscaling it, and then adding completely new features sound very similar to Sarony composing the photograph of Oscar Wilde. I think it is also safe to say that Mr. Allen used a modicum of creativity when he was generating iterative prompts to enter into Midjourney, curate the sharpness of the image using Gigapixel AI, and then photoshop to add completely new features to the work.
The Copyright Office disagrees. They state that providing the text prompts to Midjourney does not form the generated image. Second Request for Reconsideration for Refusal to Register Theatre D’opera Spatial, SR # 1-11743923581; Correspondence ID: 1-5T5320R, p. 7. However, as we learned from above, it is only the input to a GAN that can generate an output, essentially forming an image. These text input prompts form the basis of a digital brushstroke. Let us compare this to using photoshop to generate a completely digital image using only their drawing tools. Each new shape, line, contour, color, diffusion, edge filter, or gaussian blur an artist affixes to the screen is put there through a mathematical formula. The software is translating your graphical input into a digital mathematical representation. Under current copyright law, this type of art is registerable. You can distill a neural network into one giant, non-linear mathematical formula, just as you can with any of the above tools. As such, the use of the GAN in Midjourney is as much a digital brushstroke as any other drawing tool available within a digital artist’s toolbox. When Mr. Allen input text into Midjourney, a mathematical calculation began. The output was a digital brushstroke, the mathematical representation translated into graphical form. Therefore, Theatre D’opera Spatial should be copyrightable because it is creative, but current U.S. Copyright Law prevents it from being registered.
Note: This paper jumps over the argument about human authorship (or lack thereof) to discuss a more interesting question, creativity. A more thorough paper would also consider human authorship in its analysis. Generally speaking, there is human authorship in both the creation of these AI tools as well as in the outputs which are generated.