Artificial intelligence will take over a huge number of human occupations. With recent architectures such as DALL•E 2 and GPT-3 it has become increasingly clear that art will be one of those occupations. This can feel unnerving, as only a decade ago many agreed that art would be one of the last things to be automated.1
As a filmmaker, musician, portrait artist, and writer, this feels both intimidating and exciting to me. I love to make art.
I suspect we’re all experiencing a kind of “Goldilocks zone” right now, a calm before the storm of our eventual artistic obsolescence. AI is currently good enough to act as a fantastic tool for making art, but not yet good enough to replace the artist entirely.
When said replacement does eventually occur2, what will happen to human artists?
Well, some will continue on like nothing happened; these are people for whom art is not an income source and who derive most of their fulfillment from making art, not sharing it. Others will become destitute and may abandon art; these are people who make a living by sharing their art.
Either way, the demand for human-generated art will shrink drastically. I predict the job of “professional artist” will become extremely rare, and the industry of human-made art will be worth only a tiny percentage of what it is today.
AI is already rapidly learning to make several of the most popular kinds of art we make — namely music, images, videos, 3D sculpture, and writing — and is close to being capable of making those things better, faster and cheaper than us.
For those who scoff at this because the current state-of-the-art methods are “still not that good,” I invite you to look at how much better they are now compared to just a single decade ago. The trend, to me, appears very clear.3 AI will continue to get better at making art.
A common argument I hear is that “well, even if AI will create some art, it will never create masterpieces like the best human artists, so art is safe.” Even if we assume this is true, it still wouldn’t prevent AI from dominating the art industry!
I propose that the quality of art follows some Gaussian distribution. Some art is “bad”, some is “good”, but most lies in the realm of “eh, it’s a thing that exists”. Importantly, people pay for this “middle-tier” art — in fact most of the art people pay for is middle-tier.4 Thus AI doesn’t need to create masterpieces. It only needs to create middle-of-the-bell-curve art to displace the majority of the industry, because the majority of the industry is (by definition5) middle-of-the-bell-curve!
I think AI has already handily proven that it will be able to generate middle-of-the-bell-curve art. DALL•E 2 is already extremely close to meeting that criterion for clip-art.6
For what it’s worth, I also anticipate AI will create masterpieces, it’ll just take a little longer to get to that level. In the end, we stratify art based on how it evokes an emotional response in our minds, right? So human emotional response7 will simply become another term in the AI’s loss function, and eventually it will learn to play on our heartstrings like a fiddle.
We might like to think that we would somehow be immune to this kind of direct optimization against our own feelings, but the fact that we as humans can read and manipulate one another’s emotions so effectively makes me strongly suspect that AI will be able to as well.8 You might even argue that AI is already manipulating human emotions due to its role in social networks and recommendation algorithms!
This whole discussion — of AI and the death of the human-made art industry — is one where people often disagree based on emotion instead of reason. It is admittedly worrying to imagine a world where human art has lost almost all of its value — but just because something is scary doesn’t mean it is unlikely or impossible.
To the reader: do you believe the current trends in AI art will continue until we arrive in a future similar to what I’ve described here? Do you believe I’ve overlooked key ideas that prove me wrong? I would love to hear your thoughts.
Thanks for reading,
Kai
Quoting from Sam Altman’s recent blog post:
“A decade ago, the conventional wisdom was that AI would first impact physical labor, and then cognitive labor, and then maybe someday it could do creative work. It now looks like it’s going to go in the opposite order.”
that is, when the large majority of all art ingested by consumers is AI-generated
It’s especially clear in image and video generation — keep in mind that the GAN was only invented 8 years ago at the time I’m writing this. We went from being nearly incapable of generating novel high quality images with AI to StyleGAN 3 and DALL•E 2 in just 8 years.
This is not intended to sound snooty, it’s just how things are. I personally consume middle-tier art myself very frequently.
I recognize that this is only “by definition” if you agree with my assumption of a semi-Gaussian distribution.
I would even argue that a significant portion of DALL•E 2’s outputs (based on the limited examples I’ve seen released by those who have access to the API) are already good enough to use as clip art in articles without most of the audience ever noticing they weren’t human-made.
probably measured with yet another neural network trained to do nuanced emotional state encoding from video and/or audio of the art consumer