By: Jacob Alhadeff
On February 14, 2022, the Copyright Review Board (CRB) rejected Physicist Stephen Thaler’s claim for a copyright of his algorithm’s “authorship” because a “human being did not create the work.” On September 15, 2022, Kris Kashtanova received a copyright for their comic book Zarya of the Dawn, in which all of the art was AI generated, but Kris created the other aspects of the book. The difference in treatment is likely down to questions of originality, authorship, and simply that one work required human creativity while the other was effectively the work of a computer. Though these legal arguments are compelling in themselves, a necessary and implicit policy rationale seldom explicitly recognized by the law deserves highlighting — the relationship between work and incentives. Here, copyright incentivizes Kashtanova’s creative human work while reasonably denying that incentive to Thaler’s exclusively AI generated art.
AI art, AKA generative art, uses machine learning (ML) algorithms that have been trained on billions of images frequently from licensed training sets and images publicly available on the internet. The images these algorithms use are frequently copyrighted or copyrightable. Users then type in a phrase, “carrot parrot,” for example, and a unique image is generated in seconds. Creating novel art can now be as simple as an image search on Google. This technology has been in the works for many years, but recently, platforms like DALL-E, Midjourney, and Stable Diffusion increased the volume of training data from millions to billions of parameters and the emergent result was an exponentially better output. In response, on October 17, 2022, Stable Diffusion announced the completion of a $101M seed round at a $1B valuation. Sequoia Capital then posted a blog suggesting that generative AI could create “trillions of dollars of economic value.” The future of Generative AI looms large, and at the very least promises to expose unexplored ambiguities in copyright.
Functionally, in generative art there are two primary entities that may be incentivized through copyright — the programmer or the user. The programmer may have spent many hours writing and training the algorithm so that the algorithm may quickly create novel works of art. The user of the algorithm, on the other hand, is “the person who provides the necessary arrangements,” basically the person who prompts the program with a phrase. Providing either of these entities a copyright to exclusively generated art ineffectively balances incentives and ignores the purpose of copyright.
Incentives and the Purpose of Copyright
Copyright’s purpose is to “promote the progress of Science and useful Arts.” The Constitutional basis for copyright is therefore explicitly utilitarian. The Supreme Court has expanded on this language, suggesting that copyright’s purpose is to (1) “motivate the creative activity of authors and inventors by the provision of special reward” and (2) “to stimulate artistic creativity for the general public good.” Justice Ginsburg found that copyright’s dual purposes are mutually reinforcing because the public is served through copyright’s individual incentive. This mirrors James Madison’s claim regarding copyright, that “the public good fully coincides in both cases [copyright and patent] with the claims of individuals.” At its core, copyright is a monopoly-based incentive to create art to further public welfare. This incentive is at least implicitly predicated on the notion that creating valuable creative works is not easy, and therefore requires or deserves incentivizing. If improper law and policy are adopted, then Generative AI has the possibility to throw a wrench in this balancing of incentives.
The now rightfully defunct “sweat of the brow” copyright standard awarded a copyright partially because of the amount of work that went into the effort. One reason “sweat of the brow” was flawed was because it meant that facts themselves could be copyrighted if it took substantial work to attain those facts. The ability to copyright a fact “did not lend itself to support[ing]  the public interest” and the standard was discarded. Though improper, the underlying concept was not entirely baseless. If the Constitutional purpose of copyright is to provide incentives to artists for public benefit, then copyright law must balance incentives, which implicitly balances work versus reward.
Incentives are not absolute but are contextual and must at least tacitly recognize the difficulty of the act the incentive intends to induce. ‘Energy in’ must be somewhat commensurate with ‘value out’ — otherwise, the incentive structure is misaligned. This balancing of incentives is one of the reasons why a perpetual copyright is unconstitutional. If a copyright holder holds this monopoly right too long after its initial creation, they are rent-seeking, and the incentive that copyright provides far overshadows the public benefit. Rent-seeking is growing one’s wealth without “creating new wealth,” which has pernicious societal effects. For this reason, courts have determined that no amount of creativity, originality, or work merits an infinite monopoly on a creative work.
Exclusively Generated Art Should Enter The Public Domain
Neither the user nor the programmer should receive a copyright for exclusively generated art, in part because doing so would misalign incentives. To be overly reductive, incentivizing someone to dedicate their life to an artistic craft requires a substantial incentive — a copyright for example. By contrast, if the effort required to create the art is effectively null (typing a prompt into generative AI), then the incentive required to promote the useful art is effectively null. As such, the law should not be reticent to reduce or eliminate the incentive for someone to type five words into a generative AI and provide a public benefit by creating exclusively generated art. Importantly, this reasoning excludes an artist’s creations that use generative AI as a tool or a component of their work – these artist’s works deserve copyright’s protection. Given that without any guarantee of copyright protection, over 1.5 million users are creating 2 million images a day using Dall-E, current evidence suggests that generative art users are not concerned about a monopoly on the economic returns for their creations. Lawmakers should not be concerned either.
The owners of the generative AI algorithm should not receive a copyright for every work generated by their algorithm. Some in intellectual property suggest that AI generated art should be copyrightable because without protection, there will be a “chilling effect on investment in automated systems.” The argument is basically that if the owner of a generative art algorithm cannot hold a monopoly on the generated art, then there will be insufficient incentive to continue investing in automated systems. This ignores the concept of Software as a Service and the present reality that machine learning algorithms are currently effectively contributing to lucrative business models without guarantees of copyright protection. Relevantly, Stable Diffusion is valued at $1B.
Further, a world where the algorithm’s owners automatically have a valid copyright claim could completely undermine the market for art. Similar to how no amount of work can justify a perpetual copyright, no amount of work could justify a handful of entities with machine learning algorithms copyrighting a substantial proportion of modern artistic creation. While generative art may simply become another tool for artistry, it is conceivable that someday the world’s human artists would not compare to the volume of work accomplished by ML algorithms. Lawmakers should not reduce artistic markets to whoever can create or purchase the most effective machine-learning algorithms.