A common disparagement of copyright advocacy is that it is anti-technology. Despite overwhelming evidence that professional creators are early adopters of new technlogical developements, the talking point persists that enforcing the rights of creators can only “stifle innovation.” This “Luddite” critique of copyright rights was used to defend the predatory models of social and streaming platforms (and defend outright piracy), and now it is being used to defend the development of generative artificial intelligence (GAI).
Even if it were true that creators historically fear new technology until they discover ways in which it fosters new forms of expression, GAI does not easily fit analogies to tech revolutions of the past. Nevertheless, comparisons to the invention of photography do at least allude to the right starting point for a conversation about protected “authorship” using GAI as a tool of production. Before I proceed with that subject, however, let me clarify that a discussion about authorship with GAI does not necessarily justify the development of the products by means of “training” models with protected works. More on that below.
The Supreme Court Finds Authorship in a Photograph
The seminal case Burrow-Giles Lithographic Co. v. Napoleon Sarony (1884), which affirmed copyright protection for photographs under U.S. law, underlies the conversation about whether, and under what conditions, GAI may be used as a method of producing works that may be protected by copyright. As discussed in a recent post about Dr. Thaler’s lawsuit, the rule that human authorship is required for copyright rights to attach to a work should not be abolished or changed. But from there, the question remains as to how a human might use GAI as a tool to create a work of expression that is protected by copyright.
In Burrow-Giles, the defendant infringed the photographic portrait “Oscar Wilde No. 18,” taken in 1882 by the celebrity photographer Napoleon Sarony. The defendant argued that, as a constitutional matter, Congress should never have protected photographs as “writings” under a proper reading of the IP clause in Article I.[1] The heart of the defendant’s argument rested on the premise that the machine made the image, not the man. The Supreme Court, however, agreed with Sarony’s argument that his creative choices—posing Wilde, lighting, setting, arrangement of the furnishings, etc.—all constituted “authorship” of the resulting photograph, and thus, Congress had the authority to add the medium to the statute as a form of “writing.”[2]
From this precedent, then, it might seem like short work to argue that instructing a GAI to produce an image from one’s mental conception is easily analogized to Sarony, or any photographer, who selects and arranges certain expressive elements to produce the final work. And to an extent, this view is well-founded, at least as a starting point. But what frustrates the analogy, of course, is that a GAI is capable of producing something that is apparently “creative” with either no, or less than minimal, human direction.
Tell your camera to make an image of “Two mermaids and a walrus,” and the camera will do nothing whatsoever. Type that same prompt into Midjourney, and it will produce a few variations depicting the idea of what you described, but the authorship (i.e., how the idea is expressed) is not yours any more than if you were to present the same idea to a human artist who then created the image. A work generated by an AI is a probability outcome, a metaphorical statement roughly translated as, “Based on the information available, this is probably an image of two mermaids and a walrus.” So, how it initially represents the unprotectable idea of that image is a crapshoot and is in no way “authored” by the user who entered the purely instructional prompt.
But if the user of the GAI begins with the initial output and then proceeds to enter more complex and nuanced prompts, altering the image until it looks like what she has in mind (i.e., what she wants to express), this can reasonably be compared to the photographer in the studio. For instance, a photographer might begin by placing the subject against a simple backdrop, capture a few tests to get the creative process started, and then pose the subject, alter the lighting, employ makeup, wardrobe, set dressing, props, etc., until the idea of the subject becomes a distinctive expression or series of expressions.
I use this example on purpose because it emphasizes an important and vexing difference between photography and GAI from a copyright perspective. Whereas those first iterations the GAI produces of “two mermaids and a walrus” are not protected by copyright, those first test photos by the photographer (even if they represent steps toward the final images she has in mind) are protected by copyright. Further, the distinction between the two generative acts is not measured by the amount of labor performed. As a matter of doctrine, “sweat of the brow” neither vests copyright rights in an author nor gives an author any greater protection because she “worked harder” to achieve a particular result.
This is one reason Burrow-Giles is a starting point. One need not engage in the amount of creative control the Supreme Court inferred from “Oscar Wilde No. 18”—indeed, Sarony was more meticulous than the Court recognized—to claim copyright in a photograph. The spontaneous or serendipitous photograph taken by the “amateur” is automatically protected along with the painstakingly crafted, expert photograph.
Some have described this doctrine as one that rests upon a “convenient fiction,” a tautology that copyright “protects ‘original works’ made by ‘authors’ because ‘authors’ are people who make ‘original works.’”[3] And, indeed, this critique emphasizes the important distinction that with GAI, the question of “authorship” demands that we identify the human spark separate from the machine’s unprecedented capacity to make something on its own.
The Sarony Paradox
The conditions of 1884 are ghosts in 2024. Inference is not evidence of authorship. It is only in the last few years that one must wonder whether volumes of “creative” works have authors at all. One might even discover a “19th century cabinet card” that was generated by a computer last month. The card is neither work of authorship nor historic artifact of the period it claims to represent. It is, like so much material in the “abundant” digital age, just a prank—more disposable flotsam. The internet is of course flooded with such works, and both professional creators and rational minds, clinging to that anachronism we call “reality,” scorn the AI junk heap. But this sentiment, which I generally share, does not resolve the AI authorship question in copyright law.
As important as Burrow-Giles remains, it was also relatively simple work for the Court. It merely had to infer Sarony’s authorship from the evidence plainly visible in the photograph, and Sarony’s counsel, a skilled copyright attorney, provided descriptive language which the Court even copied in its opinion.[4] At the same time, the Court did note that, perhaps, an “ordinary” photograph might not be protected, and I often wonder how the same arguments might have been weighed if the photo before the Court had been one of the many “ordinary” portraits made at the time.
Personality of the Artist
Beginning in 1903, the Court moved on to what some might describe as a metaphysical conclusion that the “personality” of the artist lives in works where protectable expression may be quite subtle. Justices Holmes and Hand both articulated this notion,[5] which informed the doctrine known today as the “modicum of originality” necessary for copyright rights to attach to a work.
So, while it is fun to play alternate history and imagine the Burrow-Giles Court considering, say, Lynn Goldsmith’s comparatively minimalist portrait of Prince, the bottom line is that her photograph is protected while an image of Prince requested from a GAI is not. And the task in the present is to decide what, if anything, the human using the GAI might do to create an image in which even a modicum of his “personality” can be said to live in the work.
One could argue that because the dataset of the GAI comprises millions of personalities of millions of artists that, in general, its probability outcomes will only express a randomly selected arrangement of those personalities but never the personality of the GAI user. Here, the developers attempt a sleight of hand, arguing that every author both consciously and unconsciously learns from, or is inspired by, other works they have experienced. This is true but irrelevant.
Of course, works experienced by the artist shape her personality, which may be reflected in her work. But that creative experience is sequestered from at least the initial output of the GAI. Even more dubious is the prompt which instructs a GAI to produce a work “in the style of’ a named artist, directly conjuring the personality of a specific author and, quite possibly, producing a result that infringes that author’s copyrights. But what about the artist who begins with his own, highly original, mental conception and then makes extensive and nuanced use of the GAI as a tool to express that conception?
Authorship in Allen v. Perlmutter
Artist Jason Allen, used Midjourney to generate an award-winning image called “Théâtre D’opéra Spatial.” He was denied a registration in the work by the Copyright Office and is now suing for relief. Unlike Dr. Thaler’s lawsuit, which asks the court to abandon the human authorship requirement altogether, Allen presents a compelling argument that his visual work clears the bar for “originality” with room to spare. Not only does his brief rely on Burrow-Giles, but the following description could almost be a contemporary version of Sarony’s brief:
Plaintiff initially envisioned a detailed image of women in Victorian dresses wearing space helmets. He wanted to bring that vision to life using the Generative AI tool called “Midjourney.” He selected the colors, the style, and the era of the artwork, and arranged the elements in the image to represent the women dressed in elegant Victorian dresses performing opera on stage, their attire presenting a juxtaposition between old-world charm and a futuristic twist. He selected and arranged the elements to depict each performer wearing a space helmet, creating a striking contrast between the classical and the sci-fi elements. Plaintiff set the stage in a grand theater, with an audience watching intently, overlooking a large circular window through which the vast expanse of the outer world is visible, adding an otherworldly ambiance to the performance.
According to Allen, he re-prompted Midjourney 624 times until he achieved the desired result—a process he reasonably analogizes to a motion picture director doing multiple takes, citing the world record at 148 takes for a scene with dialogue. While Allen’s descriptions of the process like “tedious, complicated, and frustrating” may be excluded by the court as “sweat of the brow,” Allen makes a plausible argument that he expressed his mental conception in a fixed visual work, applying the same principles that protected Sarony and all photographers ever since.
At present, the Copyright Office may be too intensely focused on “control” of every aspect of a work when, as discussed, copyright can attach to even the amateur photographer for whom even basic settings like shutter speed and aperture are automatically adjusted by software. Additionally, “selection and arrangement” of expressive elements constitutes authorship as a matter of doctrine. And importantly, the Copyright Office is not the right venue to decide how “broad” or “thin” copyright protection might be for a specific work. Registration is an administrative process, and it is sufficient to ask the applicant to disclaim whether AI was used (and perhaps describe that use), but the more nuanced consideration of what is or is not protected in a work (i.e., a substantial similarity test) has long been conducted by the courts.
Of course, Allen is one creator in one case, and even if he succeeds in proving authorship of his visual work, this does not resolve the question of whether the flood of authorless works produced by GAI might be generally harmful to copyright doctrine. One could argue, as a practical matter, that those works are as irrelevant as the millions of photographs taken daily by cellphones, none of which will ever be registered with the Copyright Office or have the attached rights enforced by their owners. On this basis, one might decide that the Allens of the world should be able to register their claims of copyright in contrast to the majority of authorless works that will never be claimed or adjudicated.
While that may be true in principle, in practice, there will be chicanery. Parties will lie. And if there is a way to game a system (i.e., to automate claims in volumes of GAI material and enforce those claims for damage awards), somebody will at least try it. Michael Smith faces a criminal indictment for automating royalty payments to himself by flooding streaming services with AI musical dreck. Although not a copyright matter per se, the Smith example predicts the kind of shenanigans we can expect with mass GAI outputs, which is anathema to copyright’s purpose to incentivize humans to create works.
Authorship with GAI Does Not Justify Development by Mass Infringement
Weighing the matter of “authorship” in GAI outputs is separate from the question presented in every active lawsuit against the developers—namely, whether machine learning (ML) relied upon mass copyright infringement of existing works. Additionally, recognizing that authorship with GAI may be allowable under the right conditions should not be construed to morally or legally justify GAI development by means of ingesting protected works without permission, compensation, or transparency. On the contrary, I would say that simply because authorship with GAI may be plausibly argued as a doctrine, this does not in itself favor a finding that ML with unlicensed works is transformative fair use.
The developers will first argue that their products cause new works to exist in general and that this alone “fulfills the purpose of copyright.” Second, they will contend that their products are “tools” providing new methods for authors to create works of expression. The first argument is simply wrong, while the second argument (as discussed in this post) might support a finding that a given GAI is “slightly transformative,” a precedent held insufficient to carry the day on fair use.
Works Produced by Any Means?
Contrary to the PR messages coming from the Chamber of Progress, copyright’s purpose is not to cause works to exist by any means. From its inception, and specifically in the U.S., copyright was codified as an incentive to human beings, who have the talent and desire to create and distribute works of cultural value. Clearly, GAIs are no more incentivized to produce creative works than cameras or pianos, and in response to technological developments, Congress and the Courts have always sought to protect the author’s incentive.
In response to the invention of the player-piano Congress determined in 1909 that the scrolls were mechanical copies of music compositions created by human authors. Imagine if the piano makers had argued that each piano in each home performed a unique expression of its “personality,” and that this was, therefore, expanding the purpose of copyright. Absurd as that argument would have been, it is analogous to the contemporary GAI developer’s claim that their machines, capable of reading and playing back a random selection of the “scrolls” of millions of compositions, produce outputs that fulfill the purpose of copyright.
GAI as “Tools” for Creators
As a fair use argument that GAI are “tools” for creators, the scope of the business models of the major AI companies in suit, is instructive. If the major AI developers intended to build “tools” for creators (like Procreate for illustrators), their investors may be at even greater risk than some analysts have suggested. There simply are not enough professional creators in the world to subscribe to these products as “tools” and generate ROI for the billion-dollar cost of development and maintenance. (Adobe is used by approximately 25-million creators, but nobody is talking about refiring nuclear power plants to maintain Creative Suite.)
For example, the music making products Suno and Udio, presently being sued by the record labels, are anticipating that consumers will pay monthly subscriptions to use these GAIs as alternatives to listening to artists’ sound recordings. I predict this model will not be profitable, but from a fair use perspective, the ambition itself militates against a claim under factor one that the primary purpose of these products is to build “tools” for professional music creators.
Not only do Suno and Udio produce volumes of authorless works, but even where real musicians might use them as “tools,” the evolving doctrine on authorship implies considerable uncertainty as to how, and how often, use of these “tools” will result in protected works. Allowing for a minority of outputs to be plausibly considered works of authorship is the reason I would argue these products may be “slightly transformative,” but it seems clear that the primary purpose is to make toys for consumers, not tools for creators.
Further, if I’m wrong about the market, and the business models of Suno and Udio were to be successful as planned, these consumer toys become substitutes for music creators. And whatever else we might say about this as a cultural matter, the result is antithetical to the purpose of copyright law, which again, militates against transformative fair use.
Finally, the “tools” argument for fair use is intertwined with the authorship question. To the extent Burrow-Giles is instructive on the question of authorship of AI generated visual works, it may be muted when considering the use of GAI for works in other categories. Photography was seminal and remains paradigmatic in establishing a low originality threshold, but the considerations must naturally be articulated differently with literary, musical, and other works. As such, a “tools” based fair use defense may be stronger for one GAI product and weaker for another.
Thanks for indulging this long post. These are complex questions that, as mentioned, are difficult to analogize to technological developments of the past. Not only are GAIs uniquely productive compared to other technologies, but they are also still developing. The camera, player-piano, VCR et al. were all clearly defined in their functions at the time copyright confronted their potential effects on creators. This is worth noting when GAI developers lean too hard on those precedents.
[1] “…by securing for limited times to authors…the exclusive right to their respective writings…” Ar1, Sec 8, Cl 8 U.S. Const.
[2] Of note, photography means “light writing.”
[3] Newhoff, Who Invented Oscar Wilde?, citing Jane Gaines, Contested Culture.
[4] Moreover, Congress had added photography to the Copyright Act nearly 20 years prior to the case, and the Court might have been reluctant to overrule the Legislature without a damn good reason.
[5] Holmes in Bleistein (1903) & Hand in Jewelers’ Circular (1921).
