Synthetic Disenchantment

We are only seeking man. 

We have no need of other worlds. 

We need mirrors. 

–Stanisław Lem, Solaris 

Machines, artificial intelligences (AI), can now produce creative works such as art, language, and gameplay so impressive that they are occasionally indistinguishable from the creative works of skilled humans. Despite this indistinguishability, many of us still have the intuition that an AI’s creative work is less valuable than the creative work of a human. Are those of us who devalue AI works simply misguided anthropocentrists, unjustly biased against the machine and ignoring value in the world? I think we are anthropocentric but not misguided. I want to convince you that there are good reasons to value the creative work of humans over the creative work of current AI.  Let me take you back to Danto’s gallery, I think the question I want to investigate is best posed by setting up a contemporary wing in this old thought experiment: 

Imagine that you are standing before two perceptually indistinguishable canvases. As it stands, you have no reason to value one over the other, there are no visible or material differences between them. That is, until the curator informs you that they have distinct origins: the work on the left RedH, was made by a human, while the work on the right RedAI, was made by an image generator, an artificial intelligence. 

When I discover that RedAI was made by an artificial intelligence, I feel the meat fall off its bones—I come to devalue the work. Recent psychological evidence suggests that I am not alone. I think we can value and devalue the work of AI for many reasons. I want to focus on only one reason: that we should devalue the work of an AI because it is (in its current form) less creative than the work of a human. I think this could be the case for any creative works of AI, such as scientific theories or mathematical proofs, though I’ll continue to use artworks as my example.  

A nice lesson from Arthur Danto’s red-square thought experiment is that it reveals “really, no material differences need distinguish the artwork from real thing.” I think people get caught up in how good the products of these machines can look. But looks don’t always matter, the causal history of an artifact can sometimes change its value: the law would find the dollar bill minted by the country’s mint valuable, but the dollar bill made in the basement by a crook a valueless counterfeit; an artistic forgery is less inspired and profound than the work it was forged from. RedAI and RedH are perceptually indistinguishable, but so are good counterfeits of money and art: no matter how good the machine mimics our representations, its product can still be less valuable than our product. Is the AI then, a mere basement crook? Of course, we can’t just go throwing around criminal accusations. I must first show that the value of creative products depends at least a bit on their causal history. That something is created by an AI shouldn’t lead us to devalue it unquestionably, lest we be unjustly anthropocentric. 

The value of creative products does in fact depend on their causal history. This becomes clear once we see what it is we are evaluating when we call a work creative. “Creative” is used variously to describe a person, a product, an action, and a thought process. Although these four uses of “creative” seem distinct, we can unite them by recognizing that creation is a relation between a creator and their creation. When we appreciate a work as creative, we appreciate that it bears this relation: that it was the result of an action by a creator. Kant says of the product of creative genius that “originality must be its primary characteristic,” meaning that the creative genius was the originator of the rules by which he generated the work. Tomas in Creativity in Art says we congratulate the creative person because of their chosen actions: “he was the originator of the rules he implicitly followed while he was painting or writing.” And Henri Poincaré in Mathematical Creation writes: “The true work of the inventor consists in choosing among these combinations so as to eliminate the useless ones.” Determining whether something is creative requires reference to its causal history: the means by which it was generated. It is only by recognizing what was done by the creator that we can properly appreciate the creativity of a work. 

What about “what the creator did” are we evaluating? Creative actions are minimally exploratory: they are actions we take in order to seek knowledge of unknown or presumed to be non-ideal possible actions and their values. This captures the idea that creativity involves novelty: when we explore, we discover new information about whether actions are valuable or not. But exploratory actions are a broad category, I explore when I watch the movie The Fly for the first time on the advice of my brother, but watching The Fly is not a creative action. If we go back to Tomas, and Poincaré, and Kant, they all note that we are appreciating the choices made by the creator. What this suggests is that the creator can’t simply be exploring by blindly following instructions (like I do on the advice of my brother) or randomly acting (for there is no choice in this)—they must originate the action. The creator chooses to go in a certain direction because they suspect for some internal reason that there is something to be found there. Leaving complex questions of agency and action aside (in order to give the machine its best shot), I will take “originate” to simply mean that the action is not merely the following of an instruction or randomly acting. 

Attributions of creativity are also often associated with intelligence, ingenuity, and insight. Lindsay Brainard has recently argued that creativity has epistemic value. The creator doesn’t just make exploratory choices, they do so intelligently; they suspect that an unknown path is worth taking when they see something at the end of it, it is at least partially this insightfulness that we appreciate. By this I do not mean to endorse the idea that creativity must be accompanied by a particular phenomenology of insight (you don’t have to experience the ah-ha! feeling in order to have created something), but I mean that the creator must have engaged in some form of ampliative reasoning; the creator must have made some kind of leap, a choice to go in a direction because they intelligently saw, predicted or otherwise anticipated that there would be something beyond the woods. This might be accomplished by induction, abduction, reasoning by metaphor or via exemplars, etc. I think any of these are likely sufficient. This insight is part of what we appreciate when we appreciate creativity. 

Of course, we don’t usually witness the creative process, we just see the artifact itself (or the final performance, idea, etc.), not what was done. David Davies argues that we rationally reconstruct what was done when evaluating artworks (for Davies, artworks just are the performances that result in works). The product is clearly the effect of a particular causal chain, and we can usually work backwards to reach it. When we pull the decorated vessel from the earth, we see the mark of man in it: the winds and the water could only weather such a marvel, not manifest it, for its complexity reveals the need for fine and intelligent manipulation. Assuming that the product was made by a human like us, we can often access this information by thinking about what we would have done to generate it. Of course, part of what is so jarring about AI is that it disrupts this ability to read creative products, just like the good forgery does; but once we find out the forgery as such, we reassess our reconstruction of the process that led to it and realize that this process has less creativity than the original artist’s. 

That creativity requires a kind of reconstruction explains why people can better appreciate the creative products in a domain as their skills in it increase: they can better identify what was done by the creator. We need to evaluate what was done when we’re assessing the creativity of something. This is, to me, the best explanation of this relationship between skill and the evaluation of creative products. “Brilliant” chess moves, for example, are associated with those that are unlikely to be noticed by the weaker chess engine. At the most extreme, I can’t even tell whether a move in chess is creative or not if I don’t know how to play chess. If I am cut off from being able to tell what was done, if I don’t know the decision space, then I can’t seem to make an evaluation of creativity at all. 

If this is right, then evaluating the creativity of a work requires, evaluating the ampliative, exploratory choices that were made by the creator (perhaps it also requires evaluating other things, I am leaving this possibility open). This suggests that we devalue AI art because we cannot appreciate or do not value the ampliative, exploratory choices made by the AI. For the record, I am not going to argue here that AI aren’t creative at all (just less creative). I think they at least sometimes can be creative, and I get into the nitty gritty of this elsewhere. So, assuming that at least some machines are creative, why might we devalue their creative actions? I will provide two explanations: the first is that AI-created works provide us with less actionable knowledge than the creative products of our fellow humans; the second is that creativity is about engaging in a collective project with fellow humans about finding value and meaning in the world together. 

Figure 1: Polaroid-mimicry generated with DALL-E 3 via ChatGPT

The first explanation is that the possible actions revealed by the TIG are, at worst, impossible for us to replicate, or at best highly cumbersome given our current physical and cognitive architecture: they reveal paths of action with their creations that we will never use ourselves, and do not bear much influence on our lives. Machines almost always create in very different ways than we do. This has two implications: first, we might have a harder time understanding what they did (and thus evaluating their work); and second, they are exploring different possible actions than we are—usually ones we can’t access. To create the image above, I would have to go to a house with a Polaroid camera (likely an Sx-70 with low-grain film), under the right environmental conditions (the cold to get the blue-green tint), then bend the Polaroid to get the imperfection in the bottom left corner. Image generators such as DALL-E and Imagen produce images via generative diffusion conditioned on text prompts—by iteratively removing noise from a random patch of noise. If appreciating creative works requires understanding what the creator did to make them, then we might worry that we can’t appreciate the machine’s creativity because we can’t get close enough to understanding the machine’s actions. The human artist is like me, and so it is easier for me to rationally reconstruct what she did given our shared form of experience. Furthermore, I can appreciate her process internally in a way that I cannot do with the machine’s internal process. I know what it is like to shoot a Polaroid and this informs me of her process. I do not, however, have the computational power required to iteratively remove noise from a random patch of it. So, I am barred from the same intimacy of appreciation. Turing famously writes in “Computing Machinery and Intelligence”:

“The inability to enjoy strawberries and cream may have struck the reader as frivolous…What is important about this disability is that it contributes to some other disabilities, e.g., to the difficulty of the same kind of friendliness occurring between man and machine… [as between man and man]”

Turing’s lesson, runs both ways. Of course, this doesn’t mean that the work of the image generator is not valuable, it just means that we might not have access to the value: this is ultimately an epistemic problem of our own. 

Since AI image generators create in different ways than we do, they are exploring a different part of possibility space than we are. RedAI is “a work of art made by generative diffusion,” rather than “a work of art constructed by hand.” By recognizing this we can recognize just how narrow a space the image generator explores: it only explores the space of “generative diffusions” and can’t leave it: it can’t decide to cut up the Polaroid and transfer it to paper because it has no hands. It can only ever produce images via generative diffusions. Marta Halina argues that AI creativity does not marshal domain general intelligence—intelligence that is transferrable between domains. If this is correct, then the creative work of current AI will likely result in less integrated knowledge than the work of human creatives, and less epistemic value as a result. There’s also a question about how useful the machine’s explorations are to us, we don’t create generative diffusions, we create works of digital art; we may not even be able to take the actions it’s revealing. The machine creates with distinct limitations, what has practical value to us is what we can create with our limitations, actions that we can possibly take. 

I think this is right, but I also think my devaluation of AI creativity goes deeper. This leads me to a second explanation: humans are engaged in a collective project of worldbuilding and meaning-making that we take pleasure out of pursuing together, and ceding creation to the machines takes this away from us. As knowledge-gatherers and feeling-perceivers, part of being human is figuring out how to live well and do well by others, we do this by seeking and sharing value in the world. Creation is part of this human project. That we separate the natural from the artificial (those things defined historically by human manipulation) and separate ourselves from the rest of the world is evidence of this. What is creation for but the collective building of our world: about trying things out and saying to others: “Hey! Over here!, What about this?! Isn’t that cool?” and then celebrating these finds together. It makes sense to think that the work of someone like us would have more value to us, it is part of our shared project, we enjoy the pleasure and excitement of another, and it comes to us already infused with meaning for us. Samuel Coleridge said: “Art is…the power of humanizing nature, of infusing the thoughts and passions of man into every thing which is the object of his contemplations.” When we let the machines build with us, we cede power to them, that pleasureless, soulless being, an amalgamated version of us, begins shaping us—we lose agency. Furthermore, we cede this power, as it stands at least, to an imperfect mimicry of us: the way an AI would build the world is not the way we would, and as it stands it does not even take real pleasure in doing so. Arnold Berleant writes: “Art is, I am convinced, fundamentally social, even when engaged in most privately. Appreciation, at its heart, joins us in a collaborative act…in a shared intimacy of experience.” Appreciating the creative work of another is not just appreciating a thing detached from ourselves, but it is a little bit about appreciating what we collectively can do and value doing.

Maybe, if we eventually treat AI as community members, we will come to value their creativity. As of now however, the machine has deficiencies that interfere with our projects: it cannot feel the value of the things it pursues. Machines, by definition, cannot infuse the passion of man in their projects, but the creative work of any human can. This isn’t to say that the products of AI are not valuable as products (a medical breakthrough is valuable as a medical breakthrough), nor that AI will always be less valuable as beings, it is only to say that their creative products are less valuable for not being a part of this collective project.

Of course, here lies a fear of intense ignorance: “All that is lovely in himself he loves, and in his witless way he wants himself” writes Ovid of Narcissus. If AI is the mirror of man, then are we the witless narcissist? Maybe we are the narcissist, but we are not witless, in fact, we are finally shown by this thing, trained on our representations and reflecting them hollowed out back to us, what we actually value—not these mere appearances, but the knowledge, meaning, passion, and collective success we find within them.