Generative AIs like MidJourney or DALL-E have become very popular, but they also have their detractors. We often hear stories of these systems cutting and pasting the artwork of human artists and parroting the unique styles of existing artists. Copyright law is attempting to adjust copyright law to exclude AI art. This is important for creators at all levels.
I would like to discuss AI art starting from the epistemology of creativity.
The selectionist manifesto states that “all useful novelty in the universe comes from processes of change and selection.”
I know this to be true. I hope you can see it too.
What are the best examples of useful novelty and variety? Earth’s ecosystem containing all living things. The reason why all living things on Earth, including us humans, exist is because of the natural evolution of species… Because it works.
human creativity is pathetic It is compared to creativity resulting from natural evolution. If we’ve never seen a gecko, giraffe, or octopus (or a picture of them), we probably can’t imagine them either. And while humans couldn’t create a platypus from scratch, Evolution did.
Selection theory is a generalization of Darwinism to all processes. Many everyday phenomena, such as the cars we can buy, are the result of selectivist processes. Car manufacturers don’t build cars that people won’t buy, so car design evolves toward a set of desirable standards.
This is also true for the processes in our brain. Creativity is the ability to take a few concepts, combine things in new ways, and mentally compare them to a desired pattern of results, such as the model of the world we have been building from birth and whether the results are aesthetically pleasing. We generate and evaluate many alternatives in our minds and then allow the best solutions to multiply with each other to create better solutions. We can do this incredibly quickly. Every sentence we speak is created in a selectivist (like Darwinian) battle of concepts, ideas, and sentence fragments. As we speak, the battle continues unabated.
The irony of this is that since the brain uses selective attention methods among ideas that breed with each other to create better ideas… “intelligent design” is actually an evolutionary process inside the skull. If God exists, he would be carrying out the evolution of ideas in his brain. This includes car designers and people who buy cars.
Throughout our lives, we humans build a ‘world model’ in our heads. Most of what we know we learn through direct experience. There is little to learn while enjoying art in various forms. What we take away from enjoying art as artists is that it can influence our own style and technique.
I would like to emphasize that our knowledge of art, even among artists, is only a small part of our total knowledge of the world.
For LLM, almost everything in the corpus is not art. Their learning corpus contains pictures of items and agents in the world. Cats, people, cars, buildings, nature, news. Posted photos of real-world objects are generally not classified as works of art.
And LLM reads about things that aren’t in pictures and somehow connects every image of a cat to the word “cat”. If you ask MidJourney or DALL-E for an image of a cat in a box, you won’t get a pixel from the Cat in a Box corpus image. Instead, the results are synthesized based on a pure understanding of what the cat looks like, what the box looks like, and what the cat looks like in the box.
This is also what we humans do when we create. to combine. Choose. repeat.
The “non-art” part of LLM’s (and brain’s) model of the world dominates the “art” by orders of magnitude. My thoughts on elephants are unlikely to be the same as Salvador Dali’s thoughts on elephants, nor even close to DALL-E’s thoughts on elephants.
Therefore, when a computer creates an AI artwork, most of its input is derived from its understanding of non-art images and the associated language, rather than actual art. They also make new connections between concepts that are usually separate from each other based on the prompts. These prompts are an important part of the creative process, but their impact on the resulting image depends on the system’s understanding of the prompt, which in turn depends on the LLM (both the language model and the partial world model).
If your prompt includes instructions for drawing a cat in the Rembrandt style, you can see the influence of Rembrandt’s style in every layer, from composition to lighting to brush strokes. What the AI gets by training on Rembrandt’s images is basically Rembrandt’s style, never (theoretically) using what he actually painted at the pixel level.
Because Rembrandt’s style is not his paintings. Style is an abstraction, an emergent property that we (or the LLM) can learn from studying Rembrandt’s work.
If a human being paints something that looks similar to what Rembrandt painted, Rembrandt may be entitled to some credit for both inspiration and skill, but this does not in any way give Rembrandt a copyright claim. And if you ask an LLM-based AI to do something in the Rembrandt style, it will adjust its output toward understanding that style. This is obviously from a rather small selection of actual Rembrandt art available — 300+
Drawings, over 2000 sketches. I don’t expect to find a cat in a box in a Rembrandt painting.
So when you review the art created by LLM, there are no lines or borders that can be copied or pasted anywhere, just like when a human would draw in the style of Rembrandt. It’s no different from PhotoShop.
LLM’s world model-based understanding of the world becomes art when the prompt calls for art.
And the result is original art.