Consulting the Oracle
Is AI like other tools of art?
If there’s anything to compare to the existential crisis presented to art by the advent of generative AI, it was the introduction of photography.
The materials and manner of application for artistic production have changed over the millennia: frescoes, tempera, watercolor, oils, and acrylics, but the role of artist or craftsman has always involved some sort of human labor, often intensive and increasingly skilled.
But the big violation of photography was its ease of use.
Around 1826 Nicéphore Niépce captured the world’s first photograph on a pewter plate using a camera obscura and a photo-sensitive substance called the Bitumen of Judaea, an intensive process to be sure, but already by 1888 George Eastman introduced the Kodak camera, which came pre-loaded with film and allowed regular people to take photographs and develop them for relatively cheap (the negatives had to be shipped to Eastman’s lab in Rochester, New York).
Inherent in early criticism of photography is the fear of contamination, the fear that the relative mechanical ease of visual reproduction would enable the masses to dilute a sacred form that demanded difficulty, in which difficulty served as moat against the dilettante or amateur.
Undoubtedly this same fear of contamination-by-ease is present in contemporary reactions against generativeAI. Though it’s not the only argument, it is the one most gleefully trumpeted by some pro-AI practitioners, who assert that the same elitist bias against photography is repeating itself with the nascent art form of generative AI.
It is true that at a technical level photography does indeed make the ability to preserve the static image of a scene trivial in a way that is quite difficult to achieve otherwise. It does take skill in perspective and anatomy to convincingly draw or paint a scene from nature, and photography, especially with modern point-and-shoot smartphone cameras, makes this trivial to do.
If AI is no different from a camera, then what is the issue? After all, photography is now considered a legitimate art form (cf. figures as diverse as Ansel Adams to Annie Leibovitz to Weegee) despite its accessibility and convenience, so criticisms that generative AI falls outside the category of art surely are equally unsound.
The argument goes that the process of prompting, curation, and refinement through techniques like in-painting represent a form of artistic control. As anyone who has attempted to generate an image in Midjourney or Sora knows, the process is akin to trying to stab a bullseye with a pool noodle. A specific vision is often frustratingly asymptotic to execute in an AI tool. It’s quite difficult to steer the tool exactly where you want it to go, despite it being able to produce almost anything you want.
But still it is discontinuous.
At best prompts are suggestions, capitulations of total intentionality to an opaque system, however close to the mark these suggestions may come.
In this disconnect between intention and execution, it’s more like consulting an oracle than directing a tool, and it is this discontinuity that separates it from all modes of art that have come before.
Art is tough term to define, but surely artistic intention figures prominently in that definition. A caveman intends to stencil his hand in pigment. A painter makes a deliberate brushstroke on her canvas. And photography - despite initial skepticism - survived the test because the selection of what to frame in the viewfinder, what lens to use, what film, what angle, what shutter speed and aperture all function as intentional artistic choices. Even a smartphone's automatic exposure algorithm still requires a human to frame the subject.
All these tools, then, become extensions of the artist because there is a direct path that can be traced from human intention to what appears on the page or screen. The tools are tools because they become engulfed in human intentionality in an unbroken chain, extensions and not replacements of human will.
By contrast, generative AI violates this chain because its mechanism is ultimately inscrutable. Even AI researchers cannot yet fully explain why specific prompts produce specific outputs, and though the field of mechanistic interpretability is advancing, it’s still a long ways off. There is no deterministic 1:1 mapping of language input to output at the practical, consumer level outside of very controlled environments, and even then the complexity of combinatorial inputs precludes any familiarity approaching that of an artist with his own hand manipulating a brush.
Yes, if I prompt an image of a puppy, I’ll get a puppy. But I’ll never know why I got that particular puppy, no matter how detailed my prompt. With the exact same verbiage I could get a different image tomorrow, a different output on a different model. There may be rules that determine what the relationship between prompt and output are, but they are inexact.
In no other artistic medium is this an issue. A skilled artist holding a pencil has deep kinesthetic familiarity with the exact pressure needed to achieve a certain depth of shade. There is no surprise at how the action of our own body, no “check” to make sure the stroke corresponded to what we intended - indeed, such a thing is impossible. We cannot act without will.
A criticism may be made that this drastically overestimates the average technical understanding of most photographers. I’ll concede that many do not know the nitty-gritty details of how lenses bend light onto photo-sensitive surfaces.
But there is a difference. Even though the average person may not understand the underlying principles, photography and optics are well understood by at least some experts. No one is surprised that cameras function.
But even experts cannot fully account for what AI does: the object is conjured by the permutation of a billion unknowable parameters. There is no straight line of intent.
Sure, AI practitioners may develop some familiarity with the “rules” of prompting a particularly weighted model (which are ever shifting for consumer AIs) but they will always be a bit fuzzy and unreliable.
We must be honest and make our categorization precise. If a key necessary condition for art is intentionality, then it is difficult to place generatively created images in the same category as other art forms given this discontinuity of intention. It is more akin to a found object beckoned by incantation, an invocation of stochastic possibility that may then be curated, iterated, and refined, but never explicable.
So for now perhaps the best we can say is that generative AI remains unclassifiable, a sort of public daemon invoked and provoked with varying degrees of skill. From LOAB (the emergent and haunting image of a woman’s face from genAI’s yesteryear) to “Shape Store,” it’s clear that despite these questions of categorization, AI and the humans prompting them can elicit compelling - if lurid - experiences.
But the abdication of will, however brief, must be accounted for.
Exceptional piece!