Art created by a generative AI platform (like Stability or Midjourney) is potentially copyrightable. Art made entirely by a machine is not copyrightable. But art created by a machine, at the direction of a human, can be copyrightable. Such art is only copyrightable if the human artist’s input contained sufficient expressive authorship. Copyright requires some human authorship and some minimal creativity. Some AI art will have it, and some will not.
Generative AI models are trained on massive sets of image data. And the training images are often scraped from the internet without permission. Once the AI is up and running, a human user can type in a short prompt, and the software will return incredible images.
Does the person who entered the prompt own the copyright to the resulting image? I’m not aware of any US caselaw (as of Jan 2023) that directly answers this question, so here’s an analysis from first principals.
Background: Only Humans Can Own A Copyright.
The US Copyright Act protects “original works of authorship.” 17 U.S.C. § 102(a). To qualify as “authorship” the work must be created by a human being. Burrow-Giles Lithographic Co. v. Sarony (1884). There is no copyright for works produced by nature, animals, plants or machines. The crucial question is whether the “work” is basically one of human authorship, with the software merely assisting, or whether the software actually conceived and executed the traditional elements of authorship (i.e., the artistic expression).”
Generative AI puts an interesting twist on this question because a very small amount of human input can result in incredible artistic results.
Background: “Expression” Is Copyrightable, Ideas Are Not.
An idea is never copyrightable. But the expression of that idea in a tangible medium is (generally) copyrightable. The Copyright Act expressly prohibits copyright protection for “any idea, procedure, process, system, method of operation, concept, principle, or discovery, regardless of the form in which it is described, explained, illustrated, or embodied in such work.” 17 USC §102(b). As the Supreme Court put it “protection is given only to the expression of the idea—not the idea itself”.
Generative AI art involves some input from a human (the “prompt”), and some output from a computer. If the human input is a mere idea, then the output is not copyrightable. The output is just the results of a robotic procedure, even if it looks like expressive art.
For example, the idea of a “cartoon mouse” is not copyrightable. If I tell Midjourney to draw a “cartoon mouse”, and it draws an incredible mouse, I don’t get any copyright protection for the output. There was no expressive authorship from a human.
However, if the image prompt is full of artistic expression, then the AI output image may very well be protected by copyright.
Overall, the question of whether AI generated art is copyrightable should ignore:
- any plain ideas entered by a human
- any creative output or expression that was solely generated by a computer
And should focus solely on:
- The creative authorship input by a human.
If there is sufficient creative input from the human, then the output should be copyrightable.
AI Art Created Without Any Human Input Is Not Copyrightable.
A guy named Steven Thaler has repeatedly tried register the copyright to his AI generated art. He disclaims any human input at all, and his copyright registration claims keep getting rejected. On February 14, 2022, the US Copyright Office confirmed that his AI generated work “lacked the required human authorship necessary to support a copyright claim.”
Thaler could make a much stronger copyright claim if he identified the artistic expression he added to the mix. But I think Thaler is filing his cases to try to make some point about AI being alive. His legal claims are bit like performance art. They are not going to win him a copyright registration, but I respect his commitment to the bit.
AI Art Created With Only Minimal Human Expressive Input Is Not Copyrightable.
Words and short phrases, such as names, titles, and slogans, are not copyrightable because they don’t contain sufficient expressive authorship. 37 CFR § 202.1(a). The U.S. Copyright Office will not register short combinations of words. Sara Lee v. Nifty Foods (2d Cir. 1959).
AI Art Created With Sufficient Human Expression Is Copyrightable.
The Copyright office has hinted that generative AI work is copyrightable if you provide some “evidence on sufficient creative input or intervention by a human author in the Work”. This seems right to me.
Basically, if you input a sufficiently creative prompt, then the AI output is copyrightable.
This raises a few questions. Exactly how creative does the input prompt need to be? I suspect this will vary. The more the input prompt resembles the mere “idea” of the output image, the less likely the output image is copyrightable. The more expressive the input prompt language is, the more likely that the output image is copyrightable.
If AI Art Is Copyrightable, The Scope Of Protection Is Narrow.
If someone uses an incredibly creative prompt to generate AI art, and the resulting art is copyrightable, I expect the scope of copyright protection will be narrow. When there is very little artistic expression in a work of art, it is entitled to only “thin” copyright.
That is, the resulting copyright protection may only prevent identical (or nearly identical) copies of the original. If someone adjusts the prompt, and creates even a slightly different output, they are likely not infringing.
Can You Copy Existing Art Made From Generative AI Software?
Can you make identical copies of existing art made from generative AI software? I wouldn’t recommend it. There’s still some small risk that the artwork was created with sufficient creative expression (from a human) and therefore has some thin level of copyright protection. Unfortunately, the human input is usually not visible along with the AI generated output. So there is no easy way to measure how much human creativity was used to generate the image.
But why bother copying the image exactly? You can just take the ideas embodied in the AI-generated artwork (the ideas are never copyrightable), describe them back to some AI software, and have it create a new artwork. The results may be similar to the original, but likely not infringing.
A Test Case: Zarya of the Dawn
Kristina Kashtanova used Midjourney to create a comic called “Zarya of the Dawn”. The Copyright Office issued her a registration, and then backtracked and rescinded it. Kashtanova hired Van Lindberg, who filed a reply to the Copyright Office in Nov 2022, and we are currently awaiting a decision from the Copyright Office.
Zarya’s reply argues that she wrote the text of her comic without AI, so the text is obviously copyrightable. This is true, but boring. She argues that even if the AI art is not copyrightable, her selection and arrangement of the art into a comic is copyrightable as a compilation. Also true, also boring. Zarya then gets to the good stuff: she argues that her individual images are copyrightable:
The visual structure of each image, the selection of the poses and points of view, and the juxtaposition of the various visual elements within each picture were consciously chosen. These creative selections are similar to a photographer’s selection of a subject, a time of day, and the angle and framing of an image. In this aspect, Kashtanova’s process in using the Midjourney tool to create the images in the Work was essentially similar to the artistic process of photographers - and, as detailed below, was more intensive and creative than the effort that goes into many photographs. Even a photographer’s most basic selection process has been found sufficient to make an image copyrightable. Bleistein v. Donaldson (Supreme Court 1903). The same reasoning and result should apply to the images in Kashtanova’s Work.
Zarya’s letter goes on to discuss her artistic process, the dozens of iterations of each image she creates, and the evolution of her prompts. This seems like the kind of human authorship that merits copyright protection.
Zarya also gives an example of one of her prompts:
/imagine sci-fi scene future empty New York, Zendaya leaving gates of Central Park and walking towards an empty city, no people, tall trees, New York Skyline forest punk, crepuscular rays, epic scene, hyper realistic, photo realistic, overgrowth, cinematic atmosphere, ethereal lighting.
This particular prompt language, on its own, strikes me as not very creative. Is that prompt just describing the idea of a particular type of sci-fi scene? In copyright terms, is that prompt just scènes à faire? Or does the prompt include some element of human creative expression? For this particular prompt, I don’t know. I do know that some prompts will have sufficient creative expression to be copyrightable, and some will not.
Regardless of the answer to this particular question, I expect Zarya’s reply to be successful, and I expect her copyright will register for the comic as a whole.
Please tweet questions at me, and I’ll try to update this post to clarify any confusing points. @ericladler.