The tragedy of the artistic intellect and artificial intellect

Is AI art just stolen from human art? Let me put it this way. Imagine you had a fellow who grew up in a black room, the only thing that they ever saw was images on a screen. They say saw millions of images, sometimes with captions explaining what they were. Then, one day they were taught to draw, given captions without images, and asked to create images that matched them. Call this person Adam. Would it be fair to say that the images created by Adam were plagiarised from the images they saw during their training? Not really.

Okay, but is there a relevant sense in which mediation by a human is different from meditation by a neural network? That’s very hard to know. I don’t think anyone in the world really understands either humans or neural networks well enough to say.

[We know staggeringly little about how neural nets work. This is tragic. So many of us thought that the journey towards AI would be a journey towards learning about ourselves. In some ways it has been, in other ways we’ve just been building machines that we understand almost as poorly as ourselves].

Certainly, we can say that the images created by AI are most often not particularly like any single image in their training set, just like the person in the dark room’s images probably won’t be like any specific image he saw. Some say emotion makes the difference, but this isn’t plausible to me. Yes, humans experience emotion as part of the process of creation in a way that transformer models don’t, but it’s very unclear to me that this makes the difference between plagiarism and non-plagiarism. Would an apathetic artist necessarily be a thief?

Even when I set Midjourney up to plagiarize, it doesn’t. I asked for Sunflowers in a vase in the style of Van Gogh and it gave me this:

Pretty enough, doesn’t look anything like the original. It doesn’t even look particularly like other work by Van Gogh, although there is a subtle trace, almost like the sunflowers are seeking to burst out into the air around them.

Here’s the original for comparison.

It’s unsurprising that Midjourney can’t copy very well, even when I try to use it to plagiarize a famous work. Transformer models generally can’t reproduce the images in their training sets because they don’t store artworks. They were trained on millions of images and are only a few gigabytes large. The images wouldn’t fit. The models are way too small to contain all of that data, just like our hypothetical room dweller’s brain is far to small to remember all the millions of images that have flashed up on screen for them.

It would be more accurate, but still wrong, to say that they understand or grasp millions of works they are trained on, rather than remembering them. But it’s a kind of understanding or grasping which isn’t much like human understanding- or if it is like human understanding, is like only a part of it. The machines see the highly abstract structures that repeat across the dataset.

There’s a lot of untruths going around about how AI works. Consider this image by @Zededge:

These machines do not, in any sense, copy and paste from their dataset. They don’t even store the information to copy and paste. The only sense in which they can’t visualize things “without references” is the exact same sense in which you and I also can’t visualize things without references- I couldn’t visualize a dog if I’d never seen one or heard it described, nor could Stable-diffusion or Midjourney.

In the words of Upton Sinclair: “it is difficult to get a man to understand something when his salary depends on his not understanding it.” We Marxists mostly apply this to the bougies and middle management, but it’s common among workers as well.

Sadly, the first to lose out will be new artists who were just about to start earning their palettes. As Kendric Tonn noted on Twitter, a lot of the cheap commissions that went the new artists- the D&D portraits etc.- will be the first go.

Sad as this is, no amount of complaining about AI art will make it stop . No amount of distortion about how AI art works will make it go away. A huge number of artists are going to lose their livelihood. This is going to happen. It won’t stop with visual artists. Voice artists (singers included), musicians, and yes, authors like me will be among the next targeted.

There are few options. Trying to establish a loyal clientele, looking for other work, trying to kick off a discussion about how all this reflects poorly on capitalism… Different people will take different. As a guy who’d been planning on making a career in writing, I’m thinking about all these options myself, but none of them will change the reality that these jobs, already undervalued relative to their social contribution, are going to go away.

Still, I’d prefer people not misrepresent what AI is doing in a doomed attempt to save their incomes. AI is about to fundamentally transform the world, for both good and evil. We need the public to understand AI as far as is possible. Misrepresentations of AI that make it out to be a copyist might well blind people to the profound power these models have.

Our demand needs to be that the value these machines create should go to the great mass of people, and not to the 1%. A few artists thumping the table to demand we don’t employ the engines of creation won’t save us.

With all that said, I really do think there is a sense in which artists are getting a raw deal in relationship to AI art. They are not being compensated for what they provided AI with. As we will see, the sense in which they are being screwed over is, sadly, the exact same way that capitalism has always screwed artists over.

The problems created by AI art for artists reflect the structural problems of art under capitalism- viz, a large portion, perhaps the majority of the value art creates is not recognized, let alone compensated. Art is a public good. Art informs our imagination even when we don’t directly copy it, even when we only see it for a moment, hence capitalism’s attempt to incorporate the value of art- copyright- the commodification of art as if its benefits were just about specific buyers and sellers, was always a poor stopgap for any kind of real support of the arts. Copyright only captures the value the artist gives the the direct purchaser, not the value they give to the community as a whole.

Our current model of compensating artists- compensating them when someone buys their work- does not recognize the broad benefits that art provide

There is a theorem in welfare economics that says that goods with positive externalities will be underproduced. Intuitively this is obvious. Suppose every time someone in a town buys bicycle, air pollution falls 0.1% percent. The whole village benefits every time someone eats one, but the purchaser does not really benefit enough to notice. Maybe people will start buying them out of a sense of public spiritedness… maybe… but maybe not. If the production and distribution of bicycles weren’t done purely on the basis of whether it suited a specific individual to buy them, that might be better for everyone. Maybe it would be in the interests of the town to partly cover the cost of bicycles, for example. The innate tendency to disregard positive and negative externalities is one of the three major failings of capitalism (the other two are economic inequality, and the absence of democratic management of the economy).

Art has always been like those bicycles in that village. It’s a public good, it inspires everyone who sees it, far beyond just a direct purchaser. Thus art is underproduced relative to its social benefits. What we are seeing now is a great historical irony. We are seeing the broad benefits of human-created art- inspiration and crystallized wisdom that goes further than the purchaser- help destroy the institution of semi-professional human-created art. All this often-uncompensated value that is embodied in the art world- the sweat and vision of artists- the positive externalities created by artists- is being used to train machines that put artists out of a job.

The world will benefit from machines trained on the work of artists. That will be just another case of the benefits that artists provide going far beyond the initial buyer (if there was any). However, artists will not be compensated for this additional benefit- like so many other additional benefits they’ve provided.

In the Grundrisse, Marx talks about the idea of the general intellect- a kind of ambient buildup of knowledge, know-how, and information that becomes increasingly important as a factor of production. To my mind, that reservoir of wisdom is not just related to science, it includes art, culture, and the humanities. AI- requiring as it does a massive flow of data in order to come into being is built on the congealment of the general intellect. It is the expression of our collective intelligence. In order to make a machine that makes art, we concentrate the artistic component of the general intellect first into a database, and then into parameters, into a few billion numbers. Always we face this paradox- increased productivity should make us all wealthier- but instead, it will immiserate a group of workers. That irony is palpable in itself, but there is a double irony here- this congealed artistic intellect was made, in large part, by the very artists who will be immiserated by it [and the ML engineers- AI will come for their jobs eventually too].

The tragedy of art is like a garden filled with aromatic flowers that fill the land with beautiful scents. The garden, however, is much smaller than it should be, it’s scent much less widespread, and its gardener much poorer, because the gardener only makes money when someone comes to buy a specific flower. That’s the ancient tragedy of art. AI art is just the extra cruel, ironic twist at the end of that tragedy when we figure out how to manufacture and bottle that beloved scent, and the gardener goes completely broke. A better society might grow the garden, feed the gardener and manufacture and bottle the scent, all at once.