Grey Room 94, Winter 2024, pp. 56-62. © 2024 Kate Crawford and Jason Schultz as the theory goes, some economic incentive is required to encourage human creativity in a capitalist economy. 3 The incentive is a time-limited monopoly on how the work of art is made, sold, and publicly displayed or performed and who controls and profits from those activities.
Yet algorithms do not have incentives. They have instructions, they have objectives to optimize, and they have outputs to produce. Generative AI tools present an extreme case. While some users do provide extensive, detailed, and highly specific prompts and parameters to these programs, many do not. The vast majority of creative outputs are provided by the system as a function of its operational parameters. Moreover, due to the complex neural net architecture that underlies generative AI transformer models, their outputs are inherently unpredictable and unknowable in advance. The prompter has far less control over the output than the model does.
This has led the U.S. Copyright Office to find that generative outputs are best defined as coming from "machine authors" instead of the human beings prompting them. And because machine authors do not respond to incentives, the office has stated that they are not entitled to copyrights. 4 It supported this position by pointing to one of the first U.S. Supreme Court cases concerning copyrights in photography and to two lower court cases involving nonhuman authors, one spiritual and one simian. 5 So where does that leave the question of authorship and artistry for generative works? For now, in limbo: billions of art works are currently being produced by a rapidly expanding field of generative AI systems in what may be the largest experiment in "art beyond copyright" to date. Regardless of whether it is a ChatGPT novella or a Stable Diffusion image promoting a new Disney film, such art now exists as unclaimable content in the commercial workings of copyright itself. This art is unowned and can be used anywhere, by anyone, for any purpose. This is a radical moment in artistic production: mainstream software tools around the world producing a stream of works with no legally recognizable author.
In conceptual art the idea or concept is the most important aspect of the work. When an artist uses a conceptual form of art, it means that all of the planning and decisions are made beforehand and the execution is a perfunctory affair. The idea becomes a machine that makes the art.
-Sol LeWitt, "Paragraphs on Conceptual Art" (1967) 6 In 1879, the U.S. Supreme Court first articulated one of the defining principles of copyright law: the idea-expression distinction. In Baker v. Selden, the wife of deceased accountant-author Charles Selden tried to sue a competing author for copying and publishing his "condensed ledger" book system. 7 The court rejected the claim, ruling that, while copyrights could protect against identical drawing or illustrations of a concept, the underlying concepts (especially when mathematical or methodological) were "ideas" that belonged in the public domain. 8 This was meant to protect common concepts, historical facts, and other basic building blocks of creativity from becoming part of copyright's monopoly system.
However, along the way, it quickly became apparent that this distinction was troublesome, if not impossible, to apply at scale. As Judge Learned Hand observed in 1930, Upon any work, and especially upon a play, a great number of patterns of increasing generality will fit equally well, as more and more of the incident is left out. The last may perhaps be no more than the most general statement of what the play is about, and at times might consist only of its title; but there is a point in this series of abstractions where they are no longer protected, since otherwise the playwright could prevent the use of his "ideas," to which, apart from their expression, his property is never extended.
Nobody has ever been able to fix that boundary, and nobody ever can. 9 Three decades and many copyright cases later, Hand concluded that "no principle can be stated as to when an imitator has gone beyond copying the 'idea,' and has borrowed its 'expression.' Decisions must therefore inevitably be ad hoc." 10 We see this play out in a range of cases, from those involving photographs of basketball stars wearing "bling-bling" to "maze-chase" videogames to films about the South Bronx police during the 1970s whose plots contain "Elements such as drunks, prostitutes, vermin and derelict cars." 11 In each case, courts struggled to articulate the exact line between unprotectable ideas and exclusively controlled expressions.
In Satava v. Lowry, for example, the court attempted to analyze the work of two sculptors who made glass sculptures that resemble jellyfish-brightly colored glass inside a cylinder of clear glass. It found that, because jellyfish are physical creatures that already exist in the world, copyright protection did not extend to a glassin-glass jellyfish sculpture or to elements of expression that naturally follow from that idea, including any depiction of jellyfish physiology, such as tendril-like tentacles or rounded bells or the ability to swim vertically. Even though the two sculptures were nearly identical, the court found no infringement had occurred, because any similar expression was inextricably "merged" with the underlying ideas behind the work. 12 When we look at how generative AI produces art, we see an unpredictable reorienting of the idea-expression dichotomy, on the scale of anywhere from dozens to millions of works per prompt. Most prompts are essentially concepts or ideas, with each generated work manifesting as an expression of those ideas. If we agree with LeWitt, then generative AI art resonates in part with some of the systematic practices of conceptual art. The user conceives of an idea or concept, and the idea becomes a machine that makes the art via the prompts and other inputs that the user provides. Yet these expressions are entirely driven by probability-a statistical "guess" of sorts by the system as to which out of the millions or billions of possible expressions it can construct might map best to the user's idea, once again reinforcing that it is the machine, not the human, who "authors" the content theoretically eligible for copyright protection. This opens up the possibility that courts may find that the entire interface design of generative AI is essentially an ideaexpression translation machine, merging the two together into an epistemic skein that copyright law may be hopelessly unable to untangle, especially at scale.
Section 106 of the Copyright Act grants copyright owners exclusive rights to "do and to authorize" certain acts, such as reproduction, distribution, display, and performance. 13 Thus, when courts analyze copyright infringement claims, they must identify which specific acts of infringement were performed by which specific actors.
Historically, these actors have not been difficult to find. The copyist is the person who makes the copy. The tools of copying were merely extensions of their agency. But with the advent of algorithmic automation and decentralized computer networks, identifying a specific copyist became much harder. For example, when someone sends an email attaching an infringing photograph to a mailing list, courts have easily recognized the sender as a potential infringer, but what about the company that runs the email network? Or the software engineer who designed the email program and gave it the ability to attach photographs? The sender clicked the send button, but the actual copies were made on the company's servers by code that the engineer wrote specifically to ensure millions of copies were made and distributed to everyone on the list.
To confront this actor-network dilemma, courts began to develop what is known as the "volitional conduct" doctrine. Take, for example, the case of Cartoon Network v. Cablevision. Cablevision offered its customers the option of a "cloud-based" digital video recorderan online service to record television shows and watch them later. Many copyright owners objected, claiming the copies made in the cloud for this purpose infringed the copyrights in their shows. Cablevision responded that it could not be held liable because it had not made any of the copies. Instead, it argued, the user made the copy: they pressed the button on the remote of their cable box, telling the cloud server to store the show for later viewing. The court agreed, finding that the user was the one who engaged in the "volitional" act that caused the copy to be made and that, if the television companies wanted to sue someone, they needed to sue their own viewers, not the company providing the software that enabled their agency. 14 Generative AI systems are posed to push this concept of volitional conduct to its limit. When these systems produce infringing content, courts will struggle deeply with the question of who is acting as the copyist. Is it the AI researchers who gathered the training dataset? Is it the company that trained the model? Is it the user who prompted the model to produce the output? None of these actors specifically directed the model to make the infringing output or could have predicted the infringing outcome. Or does the machine itself now have the most agency and thus the most accountability for the potential harm caused? 15 These three central concerns of copyright law-authorship, expressiveness, and agency-face serious epistemological challenges from generative AI systems. Those challenges may be fatal or restorative, depending, as Walter Benjamin suggests, on another practice: politics. As he wrote in the epilogue to "The Work of Art," Fascism attempts to organize the newly proletarianized masses while leaving intact the property relations which they strive to abolish. It sees its salvation in granting expression to the masses-but on no account granting them rights. The masses have a right to changed property relations; fascism seeks to give them expression in keeping these relations unchanged. 16 Mistaking rights over expression for rights over one's livelihood is part of the politics that current copyright law perpetrates. The destabilization that generative AI offers us is an opportunity to rethink these politics and potentially change our relationship to it-within and beyond the art world.