The U.S. Copyright Office has denied an attempt to register copyright in images created using the Midjourney generative AI tool. The reasoning of its decision sharply limits the potential paths to receiving copyright in images created using generative AI tools, even where that output is generated in response to a prompt containing substantial textual authorship. The Copyright Office reasoned that because the person providing the textual prompts cannot predict how the generated image will look, that person is not an author of the image.
The decision is subject to appeal but provides the first look at what the Registration Division of the Copyright Office thinks can and cannot be registered.
Kashtanova’s Comic Book
In September 2022, “part-time prompt engineer” Kris Kashtanova secured and publicly announced what is believed to be the first U.S. copyright for a work created using a generative AI tool. The work, Zarya of the Dawn, is a comic book that combines text written by Kashtanova with images they generated using Midjourney and cropped and arranged into comic panels.
Although Kashtanova, who uses gender-neutral pronouns, said they were “open” with the Copyright Office about how the work was made in the Instagram post announcing the registration, the Copyright Office did not see it that way. And, in October 2022, the Copyright Office notified Kashtanova in a letter that it was initiating cancellation of the registration of the work. Kashtanova argued in response that the images were copyrightable because they contributed the “modicum of creativity” required under Feist Publications v. Rural Telephone Service Co.
The Copyright Office Decision
To be copyrightable, an image, like any work, must include at least a “modicum of creativity.” But, as the Copyright Office suggests in its recent letter, to be an author of that image, the putative author must not only provide that modicum of creativity but must also be a “master mind” of the work. Historically, to be a “master mind,” the individual must have significant creative control over the resulting work. However, in its letter to Kashtanova, the Registration Division appears to take the position that an individual does not control the resulting image if the individual cannot predict how it will look ahead of time. “[T]he process is not controlled by the user because it is not possible to predict what Midjourney will create ahead of time.” Copyright Office Letter, 8 (emphasis added).
To show their creative contribution, Kashtanova walked through the arduous iterative process they undertook to reach the final version of an image used in the comic book. They made hundreds, sometimes thousands, of iterations along the way, each modified via their prompts. However, the Copyright Office took this extensive iteration process to be necessary because Kashtanova was unable to predict the outputs that Midjourney would generate. But the Copyright Office analogized Kashtanova’s prompts to the suggestions and general direction a client may provide an artist they hire to create an image, which would not (in the decision’s view) lead to the client owning a copyright interest in the image.
Key Insights
1. No one should assume they own the copyright in any image they create using a generative AI tool.
It will be difficult to satisfy the Copyright Office’s predictability requirement in many cases.
2. Expressive modifications to an AI tool-generated image are likely to be copyrightable.
Although the Copyright Office denied copyright registration for the images that Kashtanova created solely using Midjourney, Kashtanova also sought registration for images that they created using Midjourney but altered post-production using Photoshop. The Copyright Office concluded that one of the Photoshopped images was still not copyrightable because the modifications did not contain sufficient expression. With respect to a second image, though, the Copyright Office withheld judgment because it did not have access to the pre-Photoshopped image and so could not determine whether copyrightable expression was added.
3. This decision’s “control entails ability to predict” position may not be the final word.
To be the “master mind” of a work, the Copyright Office and courts have long taken the position that the purported author of the work must have sufficient control over it. The Copyright Office’s position that Kashtanova could not be the “master mind” of the Midjourney outputs because they could not predict those outputs ahead of time arguably misunderstands the relationship between control and prediction.
We can accurately predict lots of things that we cannot control. And, in contrast to the Copyright Office’s suggestion, we can control things that we cannot predict. Art, including music, gives many examples of control—examples where we take the artist to be the “master mind”—where the artist would readily admit that they cannot predict the results that their control will yield. For example, Jackson Pollock could not predict with certainty what shapes his paint would make on the canvas. And even when artists do not inject that sort of intentional unpredictability, they frequently describe the inability to predict the outcome of their work ahead of time. Indeed, some aestheticians argue that if you can predict everything about the art ahead of time, you’ve made bad art. See, e.g., R.G. Collingwood The Principles of Art (Oxford University Press, 1938) at 122 (“But if his business is not amusement but art, the object at which he is aiming is not to produce a preconceived emotional effect on his audience but by means of a system of expressions . . . to explore his own emotions; to discover emotions in himself of which he was unaware, and, by permitting the audience to witness the discovery, enable them to make a similar discovery about themselves.”). The creation of art is, in real time, a process of discovery—not the rote execution of a preconceived plan.
Future Copyright Office or court reasoning may attempt to tease out a distinction between the unpredictability in Kashtanova’s outputs and that in Pollock’s works. Whether that distinction will and should be a meaningful one for purposes of copyrightability remains to be seen.
4. This decision does not directly address how to analyze a claim that an image created using a generative AI tool infringes the copyright in a preexisting work.
To be the author of a work, a person must have contributed their own copyrightable expression to it, and the “contribution” of copyrightable expression created by a different author is insufficient. There has been no claim that Kashtanova’s images contain human expression created by someone other than Kashtanova, and the decision does not provide direct guidance on how such a claim would be evaluated.