Ryan Abbott, who represented Stephen Thaler in Thaler v. Perlmutter, is now representing the artist Jason Allen in his case against Shira Perlmutter, the Register of Copyrights. Allen is seeking a declaratory judgment that the Copyright Office’s refusal to allow a copyright registration for Allen’s prompt-engineered visual work was wrong as a matter of law.
Allen has a much stronger case for authorship than Thaler did since Thaler didn’t even claim to be the author of the work that his AI created. By contrast, Allen was heavily involved in the creation of the work and engaged in 600 iterations of prompts and edits to render the final image.
As readers of this website know, I think the Copyright Office was wrong to deny Allen’s registration. The Copyright Office’s more recent position in its second report is step in the right direction, as are its recent registrations of AI-generated works based on post-generation editing and selection and arrangement. But the Office still clings to its overly rigid view that “prompts alone” can never result in a work of authorship under the current AI technology. At least it appears to have abandoned the justification for its rigid approach as somehow derived from the so-called “traditional elements of authorship,” a term that no federal court has ever used in a copyright decision in the history of U.S. copyright law, much less as a requirement to receive a copyright.
Once the Office’s report recognized that a human’s use of inpainting on AI generators can result in a work of authorship based on selection and arrangement of elements, that called into question the Office’s earlier analysis of Allen’s work. Inpainting is a form of prompt engineering that enables the editing or changing of discrete elements within an image. I think the distinction between (i) a series of iterative prompting to create and refine a work (aka “prompts alone”) and (ii) a prompt followed by inpainting is a distinction without a difference for the purposes of selection and arrangement. Either method can enable a human creator to make an original selection and arrangement of elements. Like a photographer, a person engaging in prompt engineering of an image can be an author based on their selection and arrangement of elements.
Related Stories
One response to “Jason Allen hires Ryan Abbott in case v. Copyright Office on prompt-engineered visual work”
I’ve written before about the U.S. Copyright Office’s stance being untenable, but if I were a die-hard copyright absolutist, I would argue that the difference that you believe is “a distinction without a difference” lies in who makes the choice of where to put what.
The ‘what’ is from a generative A.I. output in both cases, so that can be ignored as it is in “selection and arrangement”. The ‘where’ can differ.
Now here are three scenarios, all based on the same base image and the same desired effect. For illustrative purposes, let’s say it is a beach scene, with a seagull in the top right, and it should be replaced with a sandpiper.
Scenario A: Inpainting
The person operating things makes a selection around the seagull, and prompts the inpainting tool for a sandpiper instead.
Of note is that only the area within the selection is affected (barring any built-in feathering – no pun intended – features that would extend beyond it).
The person here had full control over the ‘where’ – for the most part. Did they make a broad selection, and the sandpiper get placed slightly further left, or slightly further right? What level of control did the person have?
For the purposes of “selection and arrangement”, that may not be a critical element.
Scenario B: adjusting an image based on prompts
The person operating things tells the generative AI / inpainting tool “replace the seagull in the top right with a sandpiper”.
Here it is the generative AI / inpainting tool that decides the ‘where’ to a much greater extent. Ideally of course it would still place a sandpiper in exactly the same place the seagull was, but without any direct input from the person operating, that’s relying on the generative AI understanding the task at hand.
But, again, for the purposes of “selection and arrangement”, that may not be a critical element.
Scenario C: refining prompts
The person operating things has a prompt. For brevity let’s say it was originally “a beach scene, with a bird in the top right”. They get the image described, with a seagull. So they change their prompt to “a beach scene, with a sandpiper in the top right”.
This is where things can start to differ wildly. Does the generative AI understand that they want to re-use the previous output and simply modify it? If not, out comes (hopefully) a new beach scene with a sandpiper in the top right – but it’s now also a sunset, whereas the previous one was the middle of the day. There’s a woman on a beach chair, where there was none before. The waves are calm, but they were cresting over in the previous result.
While the general task may have been accomplished, the person operating it had very little, if any, control over this output and I think it would be difficult to argue that they had control over “selection and arrangement”.
Of course this was with very basic prompts, and the more detailed the prompt is made, the less likely it is to vary to such an extent between outputs. But even if those refinements only lead to the seagull being replace with a sandpiper, and one way that was cresting now no longer doing so, I think would demonstrate that the person operating it did not, in fact, have control over “selection and arrangement”.
So which scenario out of the above applied to Jason Allen’s use? If it’s ‘C’, I would imagine the copyright office would stick to their existing arguments. If it’s ‘B’ I agree that there is a limited functional difference to ‘A’, and if they allow copyright to subsist in “selection and arrangement” for ‘A’, then that should extend to ‘B’.