AI copyright expansion would only magnify uncertainty

[Micaela, 4-22-24, The Yale Law Journal, “ARTificial: Why Copyright Is Not the Right Policy Tool to Deal with Generative AI,” Volume: 133, , accessed 7-9-24] nikoglenn

As discussed in the previous Section, numerous factors indicate an impending intense and lengthy judicial battle—particularly because the legal issues in terms of copyright are not as straightforward or binary as predominant media narratives have suggested. Indeed, several questions in the area of copyright and GAI remain open to judicial interpretation, with the most prominent being:

a) GAI using materials for training might not create a “copy” in the technical copyright definition of the term. A relevant element for analysis in the copyright context is whether GAI models create copies of works in a way that fits with the legal definition of “copy.” 17 U.S.C. § 101 provides that definition:

“Copies” are material objects, other than phonorecords, in which a work is fixed by any method now known or later developed, and from which the work can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device. The term “copies” includes the material object, other than a phonorecord, in which the work is first fixed.137

There are certain elements of this definition that raise questions about whether GAI models create a copy, particularly considering the fixation requirement. When describing the diffusion technique, briefing from one AI lawsuit characterizes Stable Diffusion as “a 21st-century collage tool.”138 Plaintiffs refer only to the second part of the process, inexplicably omitting the first part of the training that involves mapping the intersections between word-images pairs. In doing so, plaintiffs misrepresent the inner workings of these models as merely “a collage.”139 This mischaracterization is repeated when they argue, “Ultimately, a latent image is just another copy of an image from the training dataset,”140 trying to enforce the idea that there is a simplistic, traceable, and direct correlation between training data, latent representations, and the final image that results from a user’s prompt. Those inaccuracies were also pointed out extensively in an online rebuttal of the lawsuit.141

Contrasting the technical details outlined in Part I with the legal definition of “copy” under copyright law illuminates these inaccuracies. Take, for example, Stable Diffusion documentation, dataset and model cards. Diffusion models are trained to extract mathematical representations, linking text-image semantics and learning visual representations from natural language. Then, they use variations of diffusion techniques to reconstruct an image that correlates to the prompt input, starting from noise and adding layers of detail in each iteration. Extracting the underlying mathematical information from the correlation of a text with an image and the cosine distance is not a process that creates a reproduction of that work. There may not be fixation either, as some databases do not contain images per se, but rather URLs to images hosted on the Internet.142

The concept of “copy” should not be extended to cover any technical process that is applied to a work, particularly in the digital age. William Patry notes that the meaning of “copy” has been shifting—increasing in scope in copyright law to grapple with technological innovation such that it now includes “transitory acts such as buffering, caching, or non-consumable versions that are necessitated by the automatic operation of computers or other digital technologies.”143 These “ephemeral copies” (like the briefly stored copies of packets transmitted through the Internet that allow it to work) are subject to particular rules, like those stated in 17 U.S.C. § 112,144 and have been the subject of intense litigation as well.145

b) Training could be considered a transformative activity under fair-use analysis. A relevant precedent to consider is Authors Guild v. Google, Inc.146 The case stemmed from Google’s collaboration with several major research libraries to digitize their collections. The Google Books Project intended to index their contents to create a “virtual card catalog of all books in all languages” and make them searchable online.147 The Authors Guild initiated the suit, arguing that the digitization process amounted to copyright infringement. After a lengthy battle, the court ruled in favor of Google, finding that it was transformative and therefore a fair use to create digital copies of books from library collections (even in their entirety), without permission or payment, for the purpose of making the digital copies available for library collections and allowing the public to search them using a search engine. What AI models do in the learning process is so radically different from the creative use of a work that it is possible that courts could consider it transformative and hence a fair use, in the same way as the digitization of books by Google.

c) Styles are unprotectable under copyright law. Copyright does not protect ideas but rather their concrete expression. As a result, styles, genres, and other elements related to ideas are unprotectable under copyright law. There is, thus, a strong argument to be made that when AI models mimic an artist by extracting features of the artist’s work, those features amount to the artist’s style and are therefore beyond the scope of copyright’s protections.

d) AI-generated outputs are not necessarily derivative works. Some of the arguments usually tossed around in the ongoing debate about copyright and GAI rest on the assumptions that GAI models create a copy during training, and a work’s presence in the training dataset transforms GAI outputs into a derivative work. In fact, in one of the first suits regarding a GAI model, plaintiffs asserted that “[t]he resulting image [produced by AI] is necessarily a derivative work, because it is generated exclusively from a combination of the conditioning data and the latent images, all of which are copies of copyrighted images. [The AI model] is, in short, a 21st-century collage tool.”148

But these arguments rely on a causation fallacy and reductionism because, as discussed above, latent representations are not “copies” as defined by copyright law, and the outputs of a GAI model are not necessarily a derivative work in the sense of there being a direct correlation between the “original” work and the outputs (particularly in light of the traceability issues explained in Section IV.A.5 supra).

Moreover, both the copy and derivative work questions require factual determinations that need to be made on a case-by-case basis by courts, which need to consider the inputs, outputs, substantial similarity, and, even in the case of substantially similar and derivative works, if the new work is original enough to be protectable in its own right, and therefore neither derivative nor infringing. GAI models are developed in such a way that the output should not be a verbatim representation of a training element in the dataset. For example, OpenAI considered mitigation strategies to prevent what is called “image regurgitation,” overtraining, overfitting, and memorization.149

It is also worth noting that even if one were to accept the premise that GAI model outputs are effectively a collage, neither collages nor remixes are per se illegal; they must be subjected to the fair-use analysis in those jurisdictions that have established such legal provisions.150

5. Redistribution and Attribution Are Infeasible Through Copyright Rules

For the sake of this discussion, let’s assume that GAI ligation is successful. How would concepts of attribution and distribution work under existing copyright rules of compensation? Should every author whose work is present in the dataset have an equivalent claim over every single output? How would such an outcome work in practice? Here, consider again the Stable Diffusion example. The model’s training dataset, LAION 5B, is composed of “5.85 billion CLIP-filtered image-text pairs.”151 Given the massive size of the training set, it is difficult to imagine how one could trace the attribution and weight of a single work into the final end result. To do so would be like proposing that a given output image is attributable to 5.85 billion copyright interests.

To put that scale into perspective, recall that 6 billion was the entire world population in 1999.152 So, arguing that a Stable Diffusion output is derivative of the specs of those multiple individual works in the dataset, or that is results from the potential combination of all of the works, would be like saying that every person walking the Earth in 1999 would have a copyright interest in every single image that is produced by these systems. This sounds like subatomic-level copyright.

Also, this objective attribution of sorts erases the possibility of considering other instances of originality, like the creativity of a user’s prompt, and ignores the rich debate over computer creativity that has been sustained in the field of AI art.153

One could argue that some training works have a bigger presence in an AI model’s output. However, it would be difficult to trace that presence and quantify the weights and inner workings in order to measure its impact on each potential output. As the problems of interpretability and traceability in AI ethics have showcased,154 such tracing would be a near impossible proposition, especially if done at scale for each GAI output.