icon

The power of the prompter


In the previous episode, to answer the question of who owns the copyright to the creations of Chat GPT and Dall-E, we left off with the instructor, the person who whispers the prompts to these creation robots. Instructions such as (to CGPT): “Write a short explanation about the copyright protection of texts created by ChatGPT“. Or (to DALL-E): “create a painting of artists that struggle about ownership of a work of art, in the style of Salvador Dali.”[1]

Is giving such an instruction, a prompt, after which the machine does the work, sufficient to acquire a copyright on the result?

Well… About the instructions mentioned above, we can at this point already state that they are inadequate . They are so general in nature that you can still follow many paths when it comes to the design (yes, even within the style of Dali). All that is prescribed are the subjects, and with the painting there is also a style indication. For our purpose this is still too general.

concrete design

As we saw in the previous episode, copyright is all about concrete design. To my students, I sometimes summarize: “Protected is not what you say, but how you say it.” If five journalists were to independently write a report on the Eurovision Song Festival finals, these would contain many of the same facts and sub-topics. Still, there would be no chance of plagiarism. Everyone would have used their own descriptions, emphasized certain events, used a certain structure, used their own tone of voice. It is these elements that individually shape a text, that make it original, that put a personal stamp on it. The above mentioned instructions to the machine do none of that.

highly detailed, creative prompts

But exceptions can be imagined. I don’t dare to rule out that certain instructions could be so detailed, and make such clear creative choices within this detailing, that thinking up the instruction itself already meets the work-test (see part 1 for an explanation of this concept). Be aware that in such a case, it could still be in question whether that would immediately mean that the picture created on that basis would also yield a copyright for the instructor, but it is certainly a step in that direction.

A step, incidentally, for which it is difficult to determine where the lower limit lies. Prof Dirk Visser, who holds the chair of IP at Leiden University, has been working for several months to shed more light on this, among other things, with the help of a number of interesting propositions submitted to the vox populi (download pdf – in Dutch). The first of these is whether a picture created from the AI prompt “blue horse, purple dog & yellow hippopotamus” is copyrighted. Well, I think not. Even this instruction, although original, is still too little detailed to my taste.

the prompt as brush?

One might even wonder whether, at the current state of the art, AI programs are capable of expressing (or rather, in this example, “conveying“) such a level of detail, that the outcome of the question (using a necessarily more detailed prompt) could be different. This is also where my aforementioned second question comes in.

That this will one day be different, is nonetheless almost a certainty by now – developments are happening so fast. Once we get to that point, the instruction, the prompt, will have become the creator’s instrument, as it were, like a brush is for a painter and a camera (or even better: Photoshop) for a photographer. But the creator would then always need to place a firm personal stamp in his prompt (which remains a tall order to fill).

adding ones own touch

We have probably not yet reached that point at the moment. It is a lucky thing for the instructor however, that such a highly detailed prompt is not at all necessary to obtain copyright. Note that at this point in our search for a possible human copyright holder to an AI-creation, we have not found anyone yet. In fact, we may safely conclude that no copyright rests on the average AI-output. This is highly convenient for the instructor in this context. After all, this entails not only that anyone may freely reproduce and publish the AI-creation. But also that anyone may freely “build on” to it. In other words, anyone may edit the AI-creation to their heart’s content, give it their own (possibly only minor) spin. Post-editing (in text or image) does not require very much to add its own original character to the creation.

Under Dutch copyright law, we speak here of a so-called “reproduction in modified form”. Section 10(2) of the Copyright Act teaches us that such a reproduction is protected as an independent work, albeit without prejudice to the copyright on the original work. But if, due to the absence of human creative choices, the original creation is not copyrighted, that limiting conclusion (“without prejudice…”) does not come into play. The instructor is the first person to observe the still unprotected AI work – and, if he wants, the only one. This creates a unique opportunity to remain the only one to obtain copyright by means of a simple post-processing operation if, as at present, the “prompt” by itself will almost always be insufficient to reach that goal. This is now happening on an ongoing basis.

we have a winner

In short: We have a winner – even if they have to perform a little additional action themselves. More on modfified reproduction and the position of the modifier in the next instalment.

[1] This was the instruction for the picture posted with this blog: “Create a painting of artists that struggle about ownership of a work of art, in the style of Salvador Dali“. It was no coincidence that “the style of Salvador Dali” was chosen. In fact, the programme’s name DALL-E is, according to its creators, an amalgamation of the name of the Pixar robot WALL-E and Salvador Dali’s surname.


The power of the prompter