Joint Authorship as a Basis for AI Copyrightability

An interesting but maybe unsatisfactory answer to a complex problem

Generative AI presents a problem when it comes to IP protection of outputs. So far the copyright office has decided that AI generated aspects of works cannot be registered for copyright protection. That is probably correct. After all, the IP Clause of the constitution has a public interest mandate: “to promote the progress of science and useful arts.” The “authors and inventors” in the clause clearly mean humans, and the intended effect is to incentivize the creation works. Machines do not need or respond to a copyright incentive. So all around this looks bad for AI copyrightability.

BUT generative AI is also a tool, one that can be used by artists to assist in creating works that are nonetheless a product of human creativity. So presumably, there is some point in which a mere AI output, which is uncopyrightable, has enough human input to become a work that is copyrightable. So where is that line drawn, and does current law provide any help in finding the line? I think the answer is yes. But first let’s scrape the surface of the eternal question: What is art?

Creativity vs. Craft

All art has a creative component and a craft component. Creativity in that there is an artistic vision that is expressed through a large number of choices made by the artist and results in a final work that hopefully others will appreciate. The craft is quite simply how the artist gets from point A (thought) to point B (a work, fixed in time). The craft can be quite creative, but it is not the thing we usually care about or protect (although sometimes some elements of an artist’s craft may be patented).

The elements of craft are important to the artist, and the skill it takes to use them often serves as a gatekeeping function in art. I can’t draw a straight line to save my life, so I will probably never sketch. But I have converted a picture into CNC path that in turn was used to cut a stamp for pottery (meaning I could imprint a picture onto clay). The knowledge of how and when to use tools is a useful skill in creating a work of art. Certain tools can also complement or even lessen the need for fine motor skills that may take years to develop. I can’t paint a portrait but I know the elements of lighting, depth of field, and composition to take a decent portrait with a camera. In the end, art is the product of a creative mind and everything else is just process.

Technologies that make craft substantially easier have always been met with pushback from those that think it’s use cheapens the output. That has always been the case. For example, when photography was invented it was dismissed as not art. Yet today many appreciate photography as expressed in several different styles from photojournalism to heavily manipulated creative photography.

Generative AI may not be that different, but under the law we are treating it different. The art in Zarya of the Dawn was treated as uncopyrightable, and there is a strong public interest case for treating certain AI outputs as uncopyrightable. Otherwise, low effort generic works would drown out the work of living artists because of how quickly and cheaply AI works can be generated. But treating all AI output as uncopyrightable will make artists afraid of using AI tools to advance their craft, as Pixar did in it’s upcoming movie. So there has to be a way to divide uncopyrightable AI works and copyrightable AI assisted works.

Solving the AI Craft Problem

Using generative AI has a lot in common with joint works like comic books. A comic book writer will write the dialogue and describe the scene, including the characters as they’ve imagined them. The comic book illustrator will then take those notes and interpret the scenes as the illustrator imagines them. Depending on the relationship of these joint authors, there might be some back and forth. The writer might reject certain design choices as being out of step with what the writer thinks is important to the story. The illustrator may also make choices that are more interesting than the writer had in mind in ways that ultimately change how the writer develops the story. This process may seem familiar to anyone that has tried to get Midjourney to output something specific that they have in mind. Unpredictability is part of the process in AI generation in a way that it simply is not with craft.

AI is more than mere craft, and it makes logical sense for that difference to result in different treatment under copyright law. It exists in the space between craft and a true co-creator. That may seem to put AI beyond the outer edges of copyright law where dragons dwell, but joint authorship law seems flexible enough to fill in the gaps. Basically, if a human contributes enough to be a joint author, then they should just receive the full share of the copyright since AIs cannot claim their portion. And adapting joint authorship law also leaves space for AIs to advance into things that look even more like joint authors than they do now. Obviously this area of law can’t be used off the shelf, but the ways it need to be adapted seem logical and flow from related areas of copyright law.

What Gaiman Can Tell Us About AI Copyright

Surprisingly, the most influential decision in joint authorship law happens to be a case involving a writer describing characters to an illustrator in ways seem pretty darn close to how we use generative AI. The case, Gaiman v. McFarlane, written by Judge Posner, is extremely helpful in thinking through how copyright could work. It is also famous for its in depth discussion of Spawn lore. The relevant discussion starts on page 657 (use the numbers on the left side) if you want to read it for yourself.

Posner applies what is known as the Nimmer Approach, named after a copyright treatise author, that says joint authorship exists when there is intent and more than de minimis contribution. Taken separately, intent means the parties need to work towards a common goal and the contribution must be one of creativity but need not necessarily be copyrightable on its own. This is important in the Gaiman case because Neil Gaiman’s descriptions of the characters that were at the heart of the copyright dispute were not especially detailed. For example, he described one as "a really old bum, a skinny, balding old man, with a grubby greyish-yellow beard, like a skinny santa claus. He calls himself Count Nicholas Cagliostro." McFarlane, apparently, did not like this description and made Cogliostro (spelling changed) look more like Moses and less like a wino.

The rest of the descriptions are equally sparse, but they are good enough for an illustrator (or Midjourney as seen above). Posner rejects the idea that a co-author’s contribution must be independently copyrightable (the competing Goldstein Approach) and says something that seems rather prescient to today’s dilemma:

The decisions that say, rightly in the generality of cases, that each contributor to a joint work must make a contribution that if it stood alone would be copyrightable weren't thinking of the case in which it couldn't stand alone because of the nature of the particular creative process that had produced it.

Posner likens this to “peeling the onion until it disappeared.” Unfortunately, Posner then bases a lot of his decision on intent, which AI can’t share, and rejects a lot of limiting principles that would be helpful in the case of AI. That means applying Gaiman without modification would put us back into the land of everything produced with an AI is copyrightable.

Picking Up What Posner Puts Down

Posner rejects a lot of helpful limiting principles involving the idea-expression distinction in his haste to decide in favor of Neil Gaiman based on intent. To make any of this workable, these limitations would have to be added back in because intent is somewhat useless in solving the question of when a human is a co-author with an AI. My proposal is to modify the Nimmer Approach into something like the following: Does the human author intend to use AI to make an original work and is the human contribution, while maybe falling short of copyrightability on its own, a creative expression rather than a mere idea?

Both prongs of this modified test are closely related, but I think this framework remains useful. Take the intent side - most images and text generated by people using AI is not intended to be a work. The images I’ve generated here are largely to break up text and draw attention to certain things I’m talking about. I do not think they stand up to this first prong, because there was no intent to create a work. However, I’ve thought about creating a Zine about a punk rock parrot since I first started playing with generative AI. I’ve tried many prompts, and to translate that into a Zine I would have to select the best representation of the character as I imagine it and then fine-tune a model so that it will consistently reproduce the same character in different settings. All of those extra steps provide evidence of the intent to create an original work.

The second prong gets more into what Posner rejected - that generic contributions should cut against copyrightability. The idea-expression distinction in copyright law has many facets, but they all work towards a common purpose. One cannot copyright ideas, only expressions of those ideas. This extends to things that have become generic like a hardboiled detective or a femme fatale. Posner rejected these “Scènes à faire” as a limiting principle, but they are helpful in determining whether intent AND contribution align in a way that is enough to produce a copyrightable work. In resolving this prong, the text of the prompt is useful to the point that I’d recommend including it in metadata of all AI produced works. The prompt for the image above was “an anthropomorphic parrot singing at a punk rock show in the style of Love and Rockets.” I’d say this prompt lacks expressive elements and leans too hard on existing work to get where I want to go. While I think that the use of reference works should not bar copyrightability, I also don’t think that they should be given much weight (if any) as evidence of creative expression.

Herein lies a problem, AI generation is currently not sophisticated enough to necessarily create better outputs based on highly expressive prompts. For example, my prompt for the dragon picture above was “a dragon attacking a boat at the edge of a flat world with water falling over the edge, gothic illustration.” Current AI can be thought of more like a search engine, it recognizes certain words based on commonly used image-text pairs and tries to put them together in a way that makes sense. It doesn’t always work, and often simpler prompts produce better results. I had several outputs that were boats made out of dragon attacking an empty sea. But technology improvements, like ControlNet, will solve many of these problems in the future. In the meantime, I think that the web of creative decisions and alterations of the output should be instructive in weighing whether the human expression was sufficient to claim joint-author status and therefore be rewarded with the copyright.

If we apply this test to a work like Zarya of the Dawn I think we might end up exactly where the Copyright Office did. There was obvious intent by the human to create an original work, and there was a web of creative decisions and alterations in putting the images into a comic book that I think would meet the standard for copyrightability. However, it’s an open question of whether any particular image standing alone had sufficient human expression to be copyrighted. The answer is probably no, considering the author used Midjourney - a tool that neither has fine-tuning controls nor responds well to expressive prompts. The Copyright Office’s decision, that the text and arrangement is protected but the art is not, has the practical effect of allowing copyright on the comic book but not the images standing alone.

I think there are other examples where clearly the visual works are human copyrightable. The Pixar example, where AI was just used to help make the flames, shows clear intent and a high degree of expressive control by the artists. Likewise, Corridor Crew’s use of AI to create an anime easily meets both prongs. The human creators had the intent of creating a new work and the expressive elements they added went so far as costuming and reference shots for all of the AI generated frames. These artists should not be denied copyright protection simply for using AI in their creative process.

Conclusion

I don’t know whether this is the best solution to the problem of copyrightability of AI outputs, but I do think it makes sense and is a small lift in terms of adapting current law. That might be good enough, although the obvious problem is that copyright will go from something easy to determine to one that needs to be determined case by case. That is why this answer is unsatisfactory, but I struggle to think of a better approach.