Copyright Should Apply Broadly to AI Outputs

I think the protectability of AI outputs is the biggest long-term issue in copyright and AI. Unlike the pitched battles currently being fought over the use of copyrighted works on the input side, applying copyright on the output side is likely to align the interests of creators and AI companies alike, all of whom benefit if AI outputs can be protected. However, there are significant legal questions to consider and many are taking what I think is too narrow a view.

The Constitution authorizes Congress to protect the works of “Authors,” which has been interpreted to mean human authors so that limits copyright to human creativity. The Copyright Act is necessarily consistent with that. The Copyright Office has issued a report on the copyrightability of AI outputs and I agree with it on this point. We all understand that the use of tools does not invalidate a claim to copyright in the resulting work: paint brushes, cameras, spelling and grammar checks, and many other commonly used tools are employed in between the human creator and the final product but the human still directly controls the output. Generative AI is also a tool, but there is little doubt it distances the output from human control. So, the question becomes very fact-specific: how much and how directly does the human author control the final expression in a given instance? The Copyright Office report also focuses on this. I see more room for nuance and space for protection than they seem to.

Prompts

Can AI prompts be copyrightable in and of themselves? Computer code is copyrightable as a set of instructions. The function it produces is not protected, but the instructions themselves are treated as a literary work. Couldn’t that be said for AI prompts, as well?

The difference between prompts and computer code is the AI model has a probability element; I will get a somewhat different output each time I use that same prompt, even from the same AI model. In contrast, computer code performs the exact same function every time it is executed. Indeed, the statutory definition of “computer program” is “a set of instructions…in order to bring about a certain result.” 17 U.S.C. §101 (emphasis added). Does the word “certain” mean that AI prompts are outside the statutory definition because there will be variances in the output? Even if prompts don’t fit the statutory definition of computer program (and I’m not conceding that is the case), as a broader policy matter, why shouldn’t they be capable of being protected as literary works?

Computer programs are unusual if not unique in copyright law. We don’t require other literary works to “bring about a certain result.” What would that even mean in the context of a novel? That every reader perceive it in the same way as a condition of protecting that literary work? Obviously not. If we consider plain-language prompts, a broader approach is called for, consistent with the type of analysis used for poetry or short stories.^[1] To put the question in a more pointed way, does a haiku lose its claim to copyright protection if it was written for the purpose of being a set of AI prompts? The answer must be “no.” That means AI prompts are not categorically uncopyrightable, it’s only a question of what is sufficiently creative expression.

Protecting AI Outputs

In considering the protectability of the outputs themselves, we need to consider the prompts and outputs together. Let’s say I type in prompts for “an image of a dense pine forest at night under clear sky with a full moon and a river winding through the forest.” The particulars of each element of the image are produced as result of the probability functions of the model, independent of the human user. How wide is the river? How tall are the trees? Where is the moon in the sky? I don’t think there is a strong argument of direct control over every aspect of the image.^[2] In this, I at least partially agree with the Copyright Office:

“While highly detailed prompts could contain the user’s desired expressive elements, at present they do not control how the AI system processes them in generating the output.”^[3]

What my prompt IS specifiying in the example above are the elements: dense pine forest, nighttime, clear sky, full moon, winding river. Copyright law has long included a category of protection where the individual elements may not be protectable, but the selection and arrangement is: a compilation. I can see a reasonable claim for copyright in the selection and arrangement of human-directed elements in a purely AI-generated image (or text, video, music, etc.). Of course, a compilation has “thin” protection, meaning it is only infringed by copying enough that one is replicating the selection and arrangement. Fair enough.

Iterative Uses of Generative AI

One of the increasingly common uses of generative AI is “vibe coding” – a plain language description of the desired function with an instruction for the AI to write the code that executes that function. In such an instance, the AI model is producing all the code itself, including the architecture. In the case of a single set of prompts and a single AI response, I would see too little direct human control to claim copyright in the output. But vibe coding is rarely a one-off process. Like many other uses of generative AI, vibe coding is commonly iterative. That is a much more interesting question.

With coding or any other creative process, as the human user/creator works back and forth with the AI model, inserting their own authorship, making editorial selections, tweaking the output and refining the prompts, the lines between what was created by the human and what was created by the AI blur into an amalgam that is impractical if not impossible to parse. That is even more so when the user/creator has trained a model on their own prior works.

THIS is increasingly going to be the creative process and we need to have a practical answers to the questions of how to establish copyrightability and how to determine the scope of protection in a given work. I certainly don’t have all the answers and no doubt there are further examples I haven’t even considered. What I will say is that as a policy choice, I believe we should favor copyrightability at least in these sorts of circumstances.

Conclusion

I firmly believe that copyright does and should continue to require human authorship. I also believe just as firmly that copyright is a key ingredient in promoting the creation and dissemination of original works, to the benefit of all. It is crystal clear that generative AI will be increasingly interwoven into the human creative process. If copyright is to continue to serve its role we must find principled and practical standards that afford copyright protection to the products of this new reality.

^[1] The short phrase doctrine still applies; a prompt would need to meet the minimum threshold of creativity like any other work.

^[2] The more refined and detailed the prompts are, the closer the facts move towards human control over the output. There is at least one case that presents this issue pending in federal court. I’m not going to comment directly on that case, but I will add here that I am categorically NOT trying to resurrect the sweat-of-the-brow doctrine. Copyrightability should be assessed by evaluating the creativity of the expressive content and the degree of human control over that expression.

^[3] “Copyright and Artificial Intelligence Part 2: Copyrightability,” U.S. Copyright Office, Jan., 2025 at 18.