Thoughts on creativity in the age of AI
When modern explorers visited Antarctica for the first time, no one claimed that they had created Antarctica. It was there all along, it was just discovered.
When astronomers viewed new planets for the first time, the same sort of thinking was applied: they had made discoveries, not creations.
Now imagine the first person who drew a triangle on a cave wall with ash, or a smiley face, or an antelope. This drawing did not exist in its physical form before the artist had made their mark. What about in principle though? If you demarcated a boundary, fixed the amount of ash available, and began enumerating all of the drawings that could exist, the drawings above would be members of that set.
Obviously, the potential drawings cannot be easily enumerated in this way. There are a countless ways to modify an existing drawing to get a new one. On the other hand there are also countless shapes that would qualify as a triangle, or an antelope, or a smily face. We could consider the the space of the abstract entities just as easily as concrete drawings.
So again, did that first person discover the drawing of the triangle on the wall? Or create it? Here, we would typically use the word “creation”. We would probably also use that word for the second person to draw the triangle in the same way, and the third person, and your neighbor if they were to go do it tomorrow.
Imagine our artist had access to an oracle. The artist must perform an elaborate interpretive dance and the oracle responds by producing a drawing on the cave wall from a limited supply of ash and wall real estate. Slightly modifying the dance changes the drawing, but the level of control is quite crude. Call this the poorly-controlled oracle.
Say instead we have access to a different oracle, which instead of accepting communication in the form of interpretive dance, accepts the time-ordered collection of brain states that the artist would have gone through, had they drawn the image they wanted to. For instance, if we could perfectly capture the sequence of brain states of the artist while they went through the entire artistic process, but without the action. Call this the finely-controlled oracle.
Finally note that both oracles cannot observe what they have drawn, so it is up to the person to determine when they are satisfied and to stop the process. For the poorly-controlled oracle this may involve several trials but the result may be better than what they could have done on their own, for the finely-controlled oracle it will correspond exactly to their intention.
With the poorly-controlled oracle, we have some of the effort associated with craft, but none of the control we might associate with a creative act, rather this looks more like a "discovery" in that sense.
With the finely-controlled oracle we have none of the effort associated with craft but we have an exact level of control we associated with the creative act. Which one is creation, and which one is discovery?
This discussion also ignores completely the the purpose of the act, is it to find Antarctica, find something that matches your intention, generate according to your intention, a mix of all three, or something else completely. The “action” is also not just to what leads up to the drawings generation, but also in recognizing when to stop and accept the drawing.
As Generative AI systems become increasingly good at producing output like text, images, and video from simple text descriptions, the role of the human in the creative process might seem less clear.
These issues are not new though, directors and producers don't have complete control over the implementation of their vision, to some extant they are "discovering" the work. The same is true for large scale team efforts.
Technologies like the camera removed some of the craftsmanship associated with painting, but introduce other flavors of craftsmanship. A low-effort camera still requires the photographer to know that the picture they have taken is good, and that they have found something worthwhile, just like our user of the poorly-controlled oracle.
Once you pick a way to define a space over which you will search, the number of possible elements will be huge. Recall Borges’ library of Babel, which is a fictional library containing every possible combination of 410-page books using a set of 25 symbols. This includes gibberish, but also the coherent novels.
In the Library of Babel, the person who reads any book might be the only one who ever does. Sure someone might read the book on a shelf over that is similar, but it won't be the same. Surely the person who first reads this book is making a discovery though, not creating?
Say we modify the of the library of babel in the following way: It still contains all the books that could possibly be written with the page/symbol restrictions like the original library, but with some additional structure. The appearance of various themes, character types, story arcs, motifs and so on in the books is connected to the spatial organization of the books on their shelves and the shelves on the floors. For instance you may find a shelf millions of varieties of adventures of a single character, and move over to the next shelf to find the color of the hat of the character has changed with everything else the same. The shelf N rows over might have the adventure in a different setting, or with a different ending or with a different style.
Now imagine our reader, wandering the floors and reading the stories he finds. He picks one up, reads it, doesn't find that it expresses the story that he wished to express. He bases this verdict on his taste, and the degree of alignment of the novel to the novel he imagines he wants bring into existence based on his thoughts, emotions and life experiences. So he moves to the left, and then the right, reading books as he goes and exploring in this fashion. Finally after some time, maybe seconds, maybe a lifetime, he picks up a book and realizes that it’s perfect. Isn't this a creative act? It’s not the library that has created the book, the book has always existed in its shelves.
Exploring this modified Library of Babel is analogous to exploring the latent space of a generic Generative AI Model. Our library has all possible novels of a certain type in it though so we are guaranteed the novel we seek is somewhere in the shelves, there is no such guarantee of our generative models.
The point of all this is to say, the human role, and purpose, in the creative act is not so obvious as it might seem. There is also a close link between creating and discovering, searching and building. As these tools improve, and our notions of control, craft, and appreciation evolve with them, these types of questions will continue to elicit new answers.