Potential Exceptions to the Non-Copyrightability Doctrine
Elsewhere, I described the “non-copyrightability doctrine” — that creative content produced by generative AI should not be protected as intellectual property that could be privately owned. However, some situations might merit exceptions. Below, I consider several classes of potential exceptions, each of which is associated with a difficult threshold question. I don’t espouse all of the exceptions, but I list the major classes that seem worth consideration. I consider these issues in the context of copyright, but analogous arguments could be made for other forms of IP.
Human-AI collaboration: The most obvious type of exception is when a work includes meaningful creative human contribution beyond generative-AI output. It would seem a shame to wholly disallow copyright of an illustration drafted largely by a human artist, merely because the artist used generative AI for minor aspects of their work. In fact, technologies like Photoshop’s “spot removal” feature are arguably generative AI, and undoubtedly, they have been used many times by human artists in work that has already been granted copyright registration. So, these cases are not so much an exception, but a continuation of existing practice.
There remains a crucial question about the threshold: How much human effort and of what kind must a work involve, in order to be eligible for copyright? In the case of the comic book, “Zarya of the Dawn,” the U.S. Copyright Office did not believe that a person’s minor Photoshopping of images qualified as a sufficient human contribution. (I believe they made the right call, for reasons similar to my original argument for the non-copyrightability doctrine.) The Copyright Office appears to have some clarity about what constitutes human authorship, but AI will test the threshold, as it enables a whole new range of human-machine collaboration. We will see an increasing number of cases where a creative work is produced through rich interaction between a human artist and an AI system, with each contributing to different degrees.
Human involvement in generative AI development: Another class of exception could go to instances where human creators built the generative AI to further their own creative goals. The U.S. Copyright Office’s current stance appears to be to deny these requests, but I think there are good reasons to allow some cases, as I discuss in a separate article. The main argument is that human creators should be allowed to build and use arbitrarily complex tools — including generative AI — to enable their creative vision without fearing loss of copyrightability.
If allowed, however, these exceptions need their own limits so as to avoid undoing the intended effect of the non-copyrightability doctrine. Without limits, the original creators of, say, Midjourney, could claim copyright over all of their system’s output (at least as long they were the ones doing the prompting), and that would be nearly the same as enabling copyright for all gen-AI output. There need to be two kinds of thresholds — one, which is probably a black-or-white yes/no question about whether the human creator(s) were sufficiently deliberate and effective in encoding their creative vision in an AI system; another about how much of a system’s output they can copyright. The first question would require, above and beyond testimonials by the creators, a technical investigation, such as code reviews; it would be difficult and put an additional burden on decision-makers such as the Copyright Office, but probably not insurmountably. I could see David Cope qualifying for copyright over the compositions of his hand-coded EMI system, but probably not Stephen Thaler for DABUS, which Thaler himself insists is “autonomous.”
The second question is, Should there be a limit on how much of a system can be copyrighted by its human authors, and if so, what’s the limit? One crude possibility would be to specify a bound on the number of copyrightable works at N per AI system (or family of AI systems[i]). Just to put a number out there, how about 11,000 distinct works, in honor of Cope, who unplugged his EMI music composition program after curating that number of pieces as its final output? The choice of N is a bit arbitrary, but no more or less arbitrary than other numbers associated with copyright, such as the lifespan + 70-year rule of copyright validity.
This “exception” is effectively a compromise solution between enabling gen-AI output copyrightability and banning it altogether.
Work imbued with special meaning. It might be worth considering exceptions for cases where a work has special meaning, such as a work generated for a historical event, or the first work produced by a specific gen-AI system. Jason Allen’s “Théâtre D’opéra Spatial” was denied copyright registration on the basis that it wasn’t human-generated, but perhaps now that it seems likely to occupy a unique place in history as the first AI-generated work to win an art award among human contestants, should it be granted copyright? In these cases, the context that the work was produced in or for is the determining factor, not the work itself or the process used to arrive at it.
The threshold question in this case is, What counts as sufficiently special? For example, any content generated for, say, the inauguration of a U.S. president might seem to be of clear historical value and worthy of copyright, even if generated entirely by AI. But, what about any content generated for any purpose by any presidential administration? What about the analogues for corporate leadership? Or, for the founders of a neighborhood mom-and-pop shop? And, should copyrightability be retroactively pursuable, because what has historical value might not be known in advance?
I include this possibility for consideration, but I am deeply skeptical about it. Relatively little is gained by allowing it, but it introduces a perilously slippery slope that would likely be exploited by rich, powerful entities that could effectively contrive prominent or history-worthy circumstances merely to copyright gen-AI work.
Pre-registration for work created through a larger creative effort centrally involving human creators. Consider the following situation: An organization hires a design firm to design a graphic for an ad campaign. As part of the brainstorming process, the human designers at the firm use generative AI to generate ideas; and, human designers invest additional time to develop additional ideas, some of which riff off of the AI output. The firm ultimately conjures up thousands of possibilities which are then put through a thoughtful culling process, resulting in a shortlist of several dozen designs. The organization’s leadership considers the shortlisted options carefully, and images by human artists remain in contention all the way to the top three ideas. Ultimately, and after painstaking consultation with human designers, the leadership opts for the option which turns out to have been generated by AI and untouched by human artists. Should it be copyrightable?
There is both a practical and philosophical issue at stake here. The practical one: Not enabling copyright, as is current practice, would effectively ban the use of generative AI in any creative work for hire (which without AI is generally copyrightable and owned by the hiring entity), because of the danger of the above outcome. The philosophical question is, Can there be human contribution to an artistic work (e.g., through curation), even if the work itself did not involve direct human handiwork? The scenario demonstrates a case where human creative effort is both significant and essential to the final work, but not tangibly represented in it.
In some ways, this scenario represents the crux of the copyrightability question, because it pits the potential advantages of gen-AI copyrightability against its disadvantages. If we, as a society, want to prioritize the employability of human artists, then we should simply not enable this class of exception. Not much would be lost: Plenty of design firms would be happy to dispense with gen-AI use entirely if it means they would stay in business; and any corporation willing to hire consultants would probably not care about whether gen-AI was used, as long as the outcome was good. But, if we choose to prioritize the innovative capacity of human-AI collaboration, then we ought not to hobble it by disabling the copyrightability of work produced as in the scenario.
Assuming for the moment that the exception is desirable, one way to enable it would be to allow for a kind of copyright pre-registration, whereby an entity could file for copyright registration of a creative outcome that is not yet known, but for which the process would involve significant effort by human artists in addition to generative AI. The pre-registration application would involve a description detailing the intended process, and the Copyright Office would issue a conditional decision based on the apparent nature of human input. The decision would then be finalized upon submission of the final work and documentation of the process. The point of pre-registration is to disallow post-hoc claims of gen-AI output copyrightability.
Allowance for specific industries, with conditions. Thinking beyond coypright, it could be that entire classes of AI-generated ideas should be eligible for IP protection for the sake of some larger good. For example, if AI systems could crank out formulas for new medicines, could they be patented, with a new requirement for patentability? It is known that pharmaceutical companies rarely spend the significant costs to develop a new drug, unless they sense the opportunity for large future profits, which are enabled by patents. Thus, disabling patents on drugs developed by generative AI could stunt the development of life-saving medical treatment. Yet, simply providing IP protection for the discoverers of a new drug could be just as inhibiting, because owners of gen-AI systems could become patent trolls who churn out drugs, patent them, and effectively extort drug manufacturers for the right to produce and sell them. Perhaps patents for such pharmaceuticals should be enabled, but with the IP going not to the discovering agent, but to the entity that is the first to pass FDA requirements for going to market. A general principle here goes back to Lockean labor — it’s the entity that puts in the effort toward practical use that deserves the IP, not the entity that first generates the content which, with gen-AI, becomes cheap and effortless.
Unique cases. Finally, it is possible there should be a catchall exception for rare, unforeseeable cases, enabled through petitions. For example, an artist might wish to develop work about AI itself that uses the output of generative AI for the sake of very specific provocation or commentary; it could be argued that such work should be copyrightable, with adequate conditions. Over time, common classes of petitions that are decided in favor of copyrightability could be added to a permanent list of copyrightable exceptions.
One point that this inquiry raises is that, human creators will likely need to become far more rigorous in documenting their processes, in order to establish their creative contributions. Agencies that grant intellectual property registrations arguably should start requiring in-depth documentation of the creative process as a precondition for registering IP. The U.S. Copyright Office appears to have made a step along these lines with new guidance it issued in March, 2023, requiring applications for copyright registration to indicate the way in which any generative AI was used. (Over time, I expect such guidance will become more stringent.) Artists might push back against such requirements as arduous bureaucracy, but I expect that detailed process documentation will evolve as a standard across many creative arts, separate from regulatory requirements. In a world where creative output becomes infinite and infinitesimally cheap because of generative AI, some buyers will still value the human provenance of work, in the same way that people currently value handmade goods or the genuine provenance of fine art. And, I suspect artists themselves will want to be able to demonstrate the human origin of their own creations. Without such documentation, there may always be some doubt that a work was created or greatly influenced by generative AI.
Notes
[i] There is a very difficult question here, which I don’t address for now: When should two systems be considered sufficiently different for each to merit its own set of copyrightable output?
