Is AI Creative?

This essay is expanded from a comment reply on reddit to someone who suggested that AI models (especially image generation models) cannot demonstrate creative originality.

Is AI Creative? or, Everything Is a Remix, Remixed

You're probably aware of the video series, Everything Is a Remix, and if you are not, you should definitely watch it. The idea is simple: creativity is what we label something that a) combines existing ideas in ways we are not familiar with and b) strikes us as meaningful in its own right, not just as a sum of its parts. One of the formative examples of creative remixing in my generation was Star Wars. This movie brought dozens of concepts together which had never been combined. It had elements of classic Japanese samurai films, World War II movies, swords and sorcery, the lasers and aliens genre of 1950s science fiction, etc. But it brought these elements together in a way that had never been seen before, and it resonated with audiences in a way that was breathtaking. No one had seen a reaction to a movie on the scale of Star Wars. It literally saved a film studio and created new categories of box office records, spawning an entire family of franchises along the way.

But was Star Wars creative? If your bar for creativity is that something truly original is introduced, then probably the correct answer is, "no." Very few people would accept that answer, however. Star Wars clearly did bring something new to the table. Its impact alone shows that the public perceived that to be true, and on a technical level, it's hard to argue that the genre slot that Star Wars existed within was present before the movie was made.

So what is creativity? That's a question that's probably not possible to address in a single essay. It's arguably unanswerable, since it is often subjectively assessed. But what we can say is that a thing is creative when it is capable of blending ideas and concepts into forms that we had not yet imagined, rather than the naïve view that creativity must involve something truly original—something we can argue might not even exist.

This brings us to the original question posed in the title: Is AI creative? First, I think we need to parse out what we mean by, "AI."

In 2017, something changed in the world of computer science, and very few people noticed. A paper was published by Google researchers titled, "Attention Is All You Need." In this paper, the researchers introduced a new AI mechanism called a "transformer." This mechanism could take text or other "tokens" and convert it into a form that allowed the system to discover deep, semantic connections between the elements it was comprised of. In the case of text, this might be words. In the case of images it could be pixels. But the key feature here was the ability to build conceptual frameworks that the input data fit into, for the first time allowing AI to work on the truly semantic level and "understand" the flow and context of its input.

After transformers were introduced, AI training took on a new shape. It used to be that you had to carefully craft your inputs for a neural network (the most basic building block of most modern AI) to learn the right lessons, and after a certain amount of training you reached plateaus of diminishing returns on how much the AI could gain from new training data. With transformers, these barriers were substantially lowered and training on truly massive sets of data became a practical reality for the first time.

Image generation AIs, which studied as their input data, pairs of text and images, were fed large subsets of the entire internet and the result was a model capable of things we'd never seen before. Images that it generated could literally combine semantic concepts in ways that were "original." How it did this is probably a dip into vector math that goes beyond the scope of this essay, but at a very high level, it was now possible to "add" and "subtract" semantic concepts. You could ask an AI what a robot unicorn would look like, and it would not just paste robot parts onto a unicorn from images it had seen. Rather, it will take the conceptual notion of a robot and of a unicorn, and find the semantic "space" that exists where they meet. This was the technological breakthrough in AI that cracked originality as a skill.

So you might think that my answer to, "Is AI creative?"  would be, "yes." But it's not quite. Neither is it, "no." There's something else that we bundle into creativity which every human being (that I'm aware of) takes for granted: intentionality. We actualize our desires through our intent, and that is something we have not yet cracked when it comes to AI. AI does not "desire" to create a robot unicorn, it simply has the inert capability to do so. Until a human (or presumably some other highly evolved creature) comes along and provides it with that push of intent, it cannot and will not do anything creative.

This leaves us in a strange place where there is the potential to be creative, but not actual creativity exists. When a human says, "make a picture of a robot unicorn," or, "write a short story about a robot unicorn," it becomes clear that a creative interpretation of that directive is executed. So is it the human that is bringing creativity to the table? Partially, and partially it is the AI. The two are clearly collaborating, and so together they are definitely creative.

AI is a creative engine that must be initiated by an entity that possesses creative desires and impulses. It is a paintbrush that understands the work it is painting far more than any other human-created tool in history. It can do a great deal of the work involved in creative expression, but it must be given something to express.

As with many problems in AI, the answers are not terribly satisfying. This is because we enter this realm with preconceptions about how simple or complex our own intelligence and creativity might be. Those preconceptions are almost always wrong. But these are the answers we come to when assessing the observed capabilities of the systems. It is a strange new world we're embarking into and one that I'm sure will challenge many other aspects of our understanding of what makes us special (or not...)