Can AI be creative?
In the not-too-distant past, when I thought about what jobs and skills AI would replace, my bet would have been around typical routine tasks like repetitive and well-defined things like sorting boxes or unloading a truck.
I never thought too much about the arts or human creative endeavors.
Even after getting destroyed in chess, people dismissed it as compute thing. Just computers "mathing" again. A chess computer is not creative. It is just searching a tree of possibilities.
AI could never make "art." It can't create anything new.
Art is what humans, and only humans, can do. Creativity was the last line of defense.
Well, the evidence is starting to pile up against this belief, and with the release of tools like Dall-E-2 and Stable Diffusion, the question of artificial creativity is back on the menu.
Anyone with just a text prompt can create art. Sure not much of it is great, and it takes some time to play with, but it can produce some amazing things. What I think is amazing about technology is everyone can now access things that were once only available to a few people.
In the past, you had to be in the nobility or rich in order to commission art. Even today, custom artwork is still expensive.
Now anyone has access to a personal artist, which to me is amazing.
But with the advancement of these kinds comes pushback from the people it disrupts. We are seeing it with some artists who are not too thrilled with these tools.
This brings up the question of whether the tools are really creative or are they just regurgitating and stealing all the amazing work from all the beautiful unique, hard-working artists of the world.
What creativity isn't
There is a lot of mysticism and romanticism around creativity. Many people believe that you either have it or you don't.
Or that it comes only in flashes. The lighting bulb moment.
This is a trope in literature and movies, the story of a tortured artist who has an extreme life to stay in touch with that creative spirit.
The searching for a muse.
It is not mystical, and it is not extraordinary. It is a skill that all of us have. Sure some people are more creative, especially kids, but like many things, it is a range, but everyone has it.
It is just that we tend to glorify artistic work and not the office worker that finds a new creative solution to their work.
So let's start with the creative genius. This idea is that someone pulls things from the creative ether and creates.
It is nonsense.
Mozart is often touted as one of these. But then, when you get past the surface-level details, you see something different.
Mozart's father was a relatively successful composer in his own right and was obsessed with teaching his children music. Mozart and his sister got training from a successful composer from a very young age.
Mozart is often hailed as a prodigy because he composed when he was a young teenager but it wasn't anything great. It took time for him to come into his own as a composer until he made the things we actually, remember him for.
Mihaly Csikszentmihalyi's take is that "Children can show tremendous talent, but they cannot be creative because creativity involves changing a way of doing things, or a way of thinking, and that
in turn requires having mastered the old ways of doing or thinking." [1]
How about Newton, one of the most intelligent humans probably to have ever lived, who basically invented Calculus on a dare. Which seems incredible, and is, but he was not alone.
Leibniz also invented Calculus independently and whose notation we use today.
While they are both exceptional humans, Calculus was in the air because the need for it was in the air. Problems beget problems. Every new thing solves a problem, but you can't just leap ahead. You can not see the problem two steps ahead. You need to solve them in order. So you solve basic geometry and algebra but then still have questions about the universe and realize you need new tools to understand the motions of planets.
This is the spot that both of them found themselves in. The problem was sitting there waiting to be solved.
There are no leaps. History is replete with multiple people working on the same thing at the same exact time.
People in a time and place steal ideas from other people in that same time in place by stealing ideas from people that just came before them.
No one in art or science hopscotches over time. Everything has a lineage that can be clearly traced.
Science has paradigms.
There are no leaps or, as Kirby Ferguson says, "Everything is a remix." [2]
What is creativity?
"Creating is taking steps, not making leaps: find a problem, solve it, and repeat." - Kevin Ashton [3]
So if creativity is not magical thinking or divine inspiration, then what is it?
There are many theories put forth on what creativity is, and what we call creativity is probably a spectrum of traits.
I would put creativity into two somewhat overlapping categories.
Divergent thinking/Originality/Novelty
We often say that kids are creative, and that is true. What we mean by this is that they often see things in new, unexpected ways. It is easy for them. They make connections between random things.
Why? Because they haven't learned enough about the world to have any type of functional fixedness.
Functional fixedness is when you can only see a thing in a fixed, known way. You see a thing as only the thing it is "made" for.
A famous problem called the candle stick problem explores this very problem. [4]
So one type of creative person, especially kids, see new interesting connections between things that many adults tend not to.
This may also be called originality or novelty and some people like Mihaly will argue that while originality is great but it is not truly creativity if it doesn't bring value.
I would argue that it is a helpful trait to have when you get to the category.
Domain creativity
Most valuable creative work requires the person to spend a certain amount of time steeped in something to create something new of value.
You need to learn the rules of the system to break them.
Bruce Lee was an innovative and creative martial artist. He created his own martial art called Jeet Kun Do, which in many ways, is not too far off from modern MMA.
Bruce Lee thought that traditional Kung Fu, with its focus on rigid forms, was limiting and locked in a fighter. He thought you should look at all martial arts, take what works, and throw away what doesn't. That fighting should be dynamic and loose.
Jeet Kun Do was in many ways ahead of its time, but it didn't come from nowhere. Bruce Lee had studied the rules of Kung Fu since he was a child. He thought about martial arts all the time.
While he wanted to throw that all away, the structure that he learned gave him a starting point on where to go.
This is the same for music and art. You can see the roots of a style in the previous generations art.
This format is baked into the Phd path for students, you need to study and truly understand a field until you can contribute anything new to it.
This is the exact reason that John Carreyrou of the Wall Street Journal thought Theranos must be a fraud. He understood this fact and though it was highly improbable that a college dropout could have invented a new method for taking a blood sample.
How to be creative.
So if you want to be creative you need to
- Study a domain up until the point that you have truly grasped state of the art, then
- Try to make unlikely connections between things or see things in new ways. Play with things until you see things in a new way or make interesting mistakes.
- Study other fields to find metaphors that can be applied to your current work.
Last note on style. We say someone has their own style and I just think that is the later stages of creativity when someone gets to the point where they start to consume their own creativity to find further paths.
Can it be replicated in AI
So if that is the case, then can AI be creative? Let's start with the current controversy about AI digital artwork.
To dispel some misconceptions about diffusion models, they are not just storing copies of all the work it has seen.
These models are trained on an insane amount of image data, but the models themselves are only a few gigabytes. That level of compression is not currently possible, or if it were would be a huge breakthrough on its own.
The model is changed by all the images it sees, but it is not storing a pixel-perfect memory of all the things.
To me, this is exactly how humans work. We see a bunch of things, they leave an impression, and then we try to recreate it, but because we, in general, can't recreate it pixel-perfect, we end up creating something new.
This is similar for musicians. They have influences and imitate their favorites. Many famous bands started out as cover bands. They learn their domain, they imitate and then, if luckily create something new of value.
The diffusion model starts with random noise, kind of like a blank page, and then step-wise adds a little bit based on its experience to try and get the image asked for. Then it does it again and again. The more steps, the more detailed the image. This is the same thing people do.
There is also a random component to this. Each time through, it can follow a different path.
An AI can experience more artwork than is possible for most or any human.
These AI's only have a very loose "understanding" of the concepts we ask it to create. They don't understand things like what a hand is. You can see this now in the nightmare fuel that gets
created when you ask it to make things like hands. It is way closer to a child's understanding of the world, albeit on a conceptual level, way lower.
Same with Alpha Go, Alpha Go would make a move, and the commentators couldn't tell if what it did was brilliant or stupid because it could try things that a human would never try.
So diffusion models meet both criteria.
It can make new weird, and unexpected connections with things, producing novelty.
"Creativity seems, among other things, to be "usefully" putting together things, which were not perceived to be related before, and it may be the initial psychological distance between things which counts the most" - Richard W. Hamming [5]
It also understands the domain across the span of human styles far beyond than any human.
The last thing is developing its own style. Most artists grab onto something they like, and they keep working it, effectively becoming their own state-of-the-art. You can see this with famous artists
as they progress over their career.
This is totally possible with diffusion models. You could have it generate versions of a thing now, but you could also feed the output back into the model just like a human artist.
Is this more in the novelty and originality state at the moment. For things like image generation I would say possibly. Though in the case of Alpha Go I would say that it truly generated valuable knowledge to the game of Go by creating new ways of successful play that no human had ever attempted in the games lengthy history.
The idea that AI would peel off creative work first is for sure unexpected for me, but now that it is here, I believe it will continue unabated.
I am looking forward to the first hit song created by AI or having endless new material made in the style of my favorite but long-gone bands.