Getting your AI art retouched by hand is not just about quality, it’s also about being able to copyright your AI art. Both the EU and the USA consider AI-assisted works on a case-by-case basis when granting copyrights. However, the copyright offices have made it clear that works generated solely by so-called prompt engineering cannot be copyrighted.
There is a human authorship requirement, which generally entails significant transformative actions taken by hand, which is what I did for this art sample. But even those significant changes are not enough!
The EU provides better guidance for how to exceed the amount of human authorship necessary for a copyright. As for the USA, it’s currently unclear where the boundaries are, but it’s clear that manual human work is required. Even 600+ prompts aren’t enough to qualify as seen with the case of Jason M. Allen’s award-winning AI artwork titled “Théâtre D’opéra Spatial”.
Keep in mind that human authorship only confers a partial copyright. The US Copyright Office will grant registration of “the human-authored aspects of the work,” but the AI-generated portions must be disclaimed and the use of AI tools disclosed.
For example, Allen explained that he “input numerous revisions and text prompts at least 624 times to arrive at the initial version of the image.” He also used Adobe Photoshop to “remove flaws and create new visual content and used Gigapixel AI to ‘upscale’ the image, increasing its resolution and size.”
The US Copyright Office asked Allen to disclaim the features of the work created by Midjourney but he refused. According to the decision, the examiner thus rejected the work because it “did not ‘fix only [Mr. Allen’s] alleged authorship’ but instead included ‘inextricably merged, inseparable contributions’ from both Mr. Allen and Midjourney.”
Recently, I helped author Joanna Scavo. She used AI-generated images from Midjourney as a starting point for creating her picture book. I fixed issues with the hands, eyes, teeth, faces, and more. She clearly documented this process so that she can attempt to secure a copyright on the important parts, which is the characters’ faces, main features, and the positioning of the text written by her.
Joanna lives in the EU so it’s possible her work will be copyrighted in that market, but it’s possible my art retouching is not enough to exceed the high threshold for human authorship set by the US Copyright Office.
The specific requirement is for transformative work to be done by a human. Here’s a quote from the US Copyright Office:
“Transformative uses are those that add something new, with a further purpose or different character, and do not substitute for the original use of the work. Nature of the copyrighted work: This factor analyzes the degree to which the work that was used relates to copyright’s purpose of encouraging creative expression.”
In other words, you MUST make significant transformative changes to secure a US copyright on AI work. For example, there is the past case of the Warhol Prince painting where the US Supreme Court ruling set the bar for transformative effort as fairly high. While the case did not directly involve AI artwork, the ruling set a precedent since Warhol painted a highly stylized version of a photo and, despite his many changes, it was deemed derivative, not transformative.
The changes made by a person must change the CHARACTER, NATURE, or PURPOSE of the speech conveyed by the image. In a nutshell, the emerging consensus seems to be as long as your work didn’t start life as a random noise pattern defined by a seed value then it can be copyrighted.
So, what counts as transformative work, you might ask? The easiest way is to combine elements of multiple AI-generated works in order to showcase your intention as a human.
For example, in Adobe Photoshop you can copy a character (or multiple characters) from one AI-generated image and then paste it into a background with NO characters that was also generated by AI tools like Midjourney, DALL-E, or Stable Diffusion. Merging multiple derived works is inherently transformative because the background image had no characters, and now, thanks to human effort, it does.
The legal copyright precedent for this scenario was set by fair use cases involving collages, which are considered “collective works”. They are made by collecting and assembling preexisting materials in a way that creates an original work of human authorship.
In general, collages are protected by copyright if they are original. However, unauthorized use of the human-made copyrighted materials in collages can be considered copyright infringement, but since individual AI-generated works inherently cannot be copyrighted it’s impossible for the specific collage of the AI works to be infringement even as the collage itself holds a valid copyright.
Another way to effectively gain copyright of AI works is to draw something first, then use your sketch in Adobe Photoshop (or Stable Diffusion image to image [img2img]) to generate an improved version of the original sketch. For example, sketch a scene, create a mask selection around the object in Photoshop and then with Generative Fill prompt it for an “illustration of [whatever is in the sketch]”. Alternatively, you can use Stable Diffusion ControlNet so your human-made sketch is the basis for the AI generation.
With effort, you can learn to use AI to generate an improved version of your sketch which is close enough to the original to be a clear derivative work. Now, the AI-generated art piece is still AI, and therefore cannot directly obtain a US copyright. However, since that art piece is a direct derivative of a human-made work – your sketch – then it can only be used with permission of the sketch’s creator. As such, beginning the work with human authorship automatically generates a de facto US copyright protection for the AI work.
The logic behind this method is essentially the Andy Warhol Prince painting case except in reverse. Warhol took a copyrighted photo and made an entirely new painting based on that photo. But because the painting was so similar to the original, it was deemed a derivative work and therefore Warhol violated the photo’s copyright in making and selling the painting. The AI image doesn’t have copyright. It’s still public domain. But if anyone ever uses it commercially, you provide your original sketch and defend your copyright on that basis.
Another example would be to create a black and white pencil sketch in the style of comic books. You would then use an AI tool to color in and shade your sketch. Because the images are the same, just colored in, it’s an automatic derivative similar to how the colored painting of Prince is derived from the black and white photograph.
The key is that the closer your final images are to the original sketch, the better. The AI image must be visibly close so manually retouching the generative output may be necessary.
To conclude, authors, when you’re planning out your next project, if copyright is important to you, please remember to hire an illustrator and/or create the initial canvas based on your human imagination and then retouch the important parts manually. You don’t want to wake up one day and find your artwork being unfairly commercially violated since it was 100% AI-generated.