Next month I will be delivering a presentation at Illuxcon, which I hope will, in part, separate some of the facts from a good deal of the fiction surrounding the impact of “AI Art Generators” on the human art experience. Over the past few weeks, I’ve been keeping an eye out for some useful examples of problematic articles that REALLY push fiction in this regard. Yesterday, I came across a real “doozy” posted to the online art-market publication, MutualArt. The essay, The Romantic Artist Is Dead. AI Has Killed Him, written by author and artist Michael Pearce had it all as far as problems go: alarmist hyperbole, nebulous language, fallacy, and a host of unsupported assertions.
I did have a limited exchange with the author regarding this subject on social media. However, none of the issues that I brought up during that exchange were addressed productively, with the interaction eventually being reduced to an exercise in “label-gazing” (a form of conversation stalling in which participants get mired in term-based fallacious arguments like appeals to definition.)
So what is the general position that I am labeling problematic? It is the position that AI Art Generators are capable of replacing or eliminating the role of humans in what we understand as “art.” So does this new essay from Mr. Pearce promote this position? I believe so.
Let’s begin our examination of this article by first establishing some basic terms essential for making this exercise productive. (I acknowledge that the author may be applying different usages compared to the ones I have shared below. This is a valid concern in any exercise like this. However, upon the first read, it does seem that the author’s term usage does not seem dissimilar to usages commonly assumed by the average reader familiar with these subjects.)
AI Art/Image Generator: During the past 50 years, innovative creatives have written a good number of computer programs to generate what some might describe colloquially as “art.” Generally speaking, these generators are impressive computer programs that analyze massive amounts of data in specific ways to produce a range of “original” output products that hold certain kinships with the source pool.
Current (text-to-image) generators do vary in their specific processes, but most can be generally understood to be using the following broad strokes:
First, a generator maps text from a user prompt onto a virtual representation space (like a virtual blank canvas.) Then, it compares that text against massive amounts of captions and alt text for collected images that have corresponding semantic (meaning/usage) elements. The source image pool is then “deconstructed” and encoded into a collection of unrecognizable (relative to source) component parts through various processes.
(For example, the Stable Diffusion generator gets its component materials by “training” on a “de-noising” process. In other words, the generator takes image data and adds what we might call ‘noise’ to it. It then repeats that noise-adding process over and over (up to ~1000 times) to obtain a result that you or I might describe as “pure noise.” The noisier products are then fed back to the generator with the instructions to predict a denoised version of the image.
Following this type of training, the generator is then able to generate image data by processing randomly sampled noise from the learned de-noising process.
Next, the generators pull from their newly encoded, unrecognizable component pool, applying data onto the aforementioned virtual representational mapping space (virtual canvas) built from the initial text prompt. The program then stochastically (somewhat randomly) assembles selected component parts into a configuration that we might describe as “recognizable.” Lastly, it uses additional textual information to assess whether what it produced is semantically consistent with the initial input prompt. The massive number of potential variations in this reassembling process means that not only is the image not an exact reconstruction of the source material but also that you never get the exact same image twice. As such, the product, in terms of content and form, can be aptly described in certain contexts as “original.”
Want this in even wildly simpler terms? “…imagine you want a one-of-a-kind, custom-made table. Now think of a forest. Cut down all those trees and put them in a woodchipper. Take all the tiny pieces of bark and sawdust and reassemble them within a rough outline of a tree. Now cut that tree down. Chop it into wood and then make the planks that somewhat resemble an Ikea table. Check the instructions and assemble. Congratulations, you’ve just constructed a custom-made table.” (Adapted from The Jasper Whisperer, Medium.com (MLearning.ai)
Now, this gross simplification does no justice to the many amazing specifics of the AI Image generation process, but I think it provides the general idea. It is important to note that Mr. Pearce does appear to clearly understand the broad strokes of the generation process, as seen in paragraph four of his article. It is actually one of the few parts that a reader does not need to worry about drowning in a torrent of flowery language.
The Art Experience: What we tend to describe as “art” colloquially is, in actuality, a description of an experience–not a physical object or property of the environment (although such usage can be practically useful in many contexts.) Art is no more an inherent property of an object than “red” is of a Liberty apple. I find that great difficulties (contradictions, irrational positions, etc.) arise when navigating conversations about the nature of art when we try and hold to the latter concept. This is not to say that our regular assignments of great value to aspects of our environment (e.g., objects) that we describe as art are, therefore, absurd. It is quite the opposite, as these aspects or objects facilitate or elicit the art experience—thus, the objects are a form of “pleasure technology.” In his essay, “We Make Art Because We Can," experimental psychologist and author Steven Pinker explains the nature of a pleasure technology using the example of cheesecake. He writes, “In explaining the hypothesis that conventional art is a pleasure technology, I once used the analogy of cheesecake: something we enjoy not because natural selection specifically adapted us to do so, but because we invented it for our own enjoyment. What is the adaptive function of cheesecake?—The answer is, obviously, none. Cheesecake is a by-product of our evolved tastes and our evolved inventiveness. This doesn’t mean that we can ignore adaptation in understanding why people like cheesecake: each kind of ingredient was adaptive in the world in which we evolved. It is a package of supernormal doses of things that in more moderate and natural amounts clearly were adaptive, like sugar, fat, and protein. But if you ask, ‘Why is cheesecake adaptive?’, you are posing the wrong question. Humans are intelligent enough to collect things that, in their natural settings, were wholesome or useful and use them to create bombs of pure pleasure.”
Examining the nature of the art experience through this lens, we can look at art objects as metaphorical pieces of software that run a potent neural program–one that activates a number of adaptation subroutines that we experience as a “bomb of pleasure.” For example, we value and enjoy drawings and paintings, not because of an affinity for marks on paper or blobs of pigment and oils on a piece of cloth was selected for. This selection wouldn’t seem to have provided any significant advantage for our ancestors working to pass along their genes. Instead, current evidence points to the fact that our response to drawings and paintings occurs as a consequence of a collection of adaptations that WERE selected for. In terms of representational content–the perception of certain faces, bodies, patterns and habitats gives humans aesthetic pleasure because they serve as perceptual cues that signal understandable, safe, productive, nutritious, or fertile things in the world.
But why then would we value an original work of art over a reproduction–even if they were otherwise identical? Wouldn’t both be offering the same neural program?
To answer this, we have to understand our essentialist nature. A simple definition of essentialism is the view that objects have a set of attributes that are necessary to their identity. Paul Bloom of Yale University explains how this is relevant to our purpose here, "one of the most exciting ideas in cognitive science is the theory that people have a default assumption that things, people, and events have invisible essences that make them what they are…
[Essentialism is] a particularly clever and important adaptation that drives us to focus on the deeper aspect of things. For a lot of things, the capacity to go beyond the surface makes a big difference. It matters, when you look at people, not to be entirely moved by what they look like, but to also be influenced by what you believe to be their histories and their hidden properties. For food, it matters where it came from and what it touched. For animals, you want to know what they can do to you and how they behave, not just their surface appearance. For these reasons, I think we’ve evolved to have an essentialist bias…
…a painting might depict something that is beautiful, it might be associated with a positive memory, it might raise one’s social status, and so on. And then there is essentialism. Consistent with the broader theory defended in How Pleasure Works, I would argue, building from the work of the philosopher Denis Dutton, that our experience of an artwork is profoundly affected by our belief about what that artwork really is. It matters to us who created it and how it was created."
So yes, if our perceptions and knowledge were hopelessly limited to the most superficial inherent properties of an object, then we would likely NOT place a higher value on an original work compared to an identical reproduction. However, evolution has left us with a drive to access the deeper essence of a thing. We work to learn more about what we encounter, so as to better inform our potential interactions with it–thus increasing the likelihood of behavioral success. Hence, the original and the reproduction would only be valued as equals if we were deceived–or our essentialist nature disabled.
Romanticism: This was first defined as an aesthetic in literary criticism around 1800, gained momentum as an artistic movement in France and Britain in the early decades of the nineteenth century, and flourished until the mid-century. With its emphasis on imagination and emotion, Romanticism emerged as a response to the disillusionment with the Enlightenment values of reason and order in the aftermath of the French Revolution of 1789. It was the German poet Friedrich Schlegel who first defined the word romantic as: “literature depicting emotional matter in an imaginative form.”
Romantic Art is characterized by its prioritization of emotion, a celebration of the individual, promotion of the “common man” (and childhood), and emphasis on the imaginative idealization of nature. On the surface, Romantic paintings can often be described as having a “painterly” appearance. Physical characteristics of these work typically include unrefined outlines, unrestrained brushstrokes, and a notable emphasis on color over form, all of which tends to communicate an energetic process and sense of immediacy. While the subject matter of Romanticists varied greatly (landscapes, religion, revolution), artists belonging to this group shared the prioritization of exploring various emotional and psychological states as well as moods in their work.
So, with those basic terms defined, let’s start to look at the article in question:
“The Romantic Artist Is Dead. AI Has Killed Him -With computational power humans could never reach, Artificial Intelligence has penetrated the artistic domain, democratizing creativity.”
The title and blast line immediately come across as run-of-the-mill clickbait. If you are unfamiliar with this term, clickbait is a sensationalized headline that prioritizes attention over presenting objective facts and supported claims. Clickbait headlines often appeal to base emotions (often fear and/or outrage) and curiosity.
While it is true that the title (and many aspects here) could be interpreted in a number of ways, to the best of my knowledge, it has not been demonstrated that there is currently a shortage of artists creating work that adheres to the qualities of “romantic” work (let alone a shortage that has been causally linked to the advent of AI text-to-image generators.) I could not locate any news stories in recent months of museums removing the works of Friedrich, Turner, Constable, Delacroix, Goya, Blake, or any other artists from the Romantic period from their walls due to the advent of “AI.” In fact, I cannot find any objective, reputable source that might even vaguely support (independently) the alarmist claim that works that comport with romantic characteristics are “dead” in any sense of the word. Perhaps this is a shortcoming of my research technique, but no such news stories are cited within the essay. (Keep in mind that the carries the burden of proof for this headline claim.)
What about the claim that AI has computational power that “humans could never reach?”
Well, you might be surprised to learn that a computer called “The Frontier” at the Oak Ridge National Laboratory in Tennessee was recently reported to have entered the realm of “exascale” computing on May 30th of this year. Frontier clocked in at about 1.1 exaFLOP (it can perform at up to 1.69 exaflops at peak with a yet unrealized theoretical peak of 2 exaflops, or 1.1 quintillion operations per second.) The human brain, in contrast, is currently estimated to “compute” somewhere around 1 exaFLOP (10^18 operations) per second.) Frontier has a total of 8,730,112 cores capable of parallel computing tasks (a typical laptop has between five and nine.) At peak power, the computer generates so much heat that it requires four high-powered pumps to push more than 25,000 liters of water around the machine each minute.
Frontier now ranks as the fastest AI system on the planet, dishing out 6.88 ExaFlops of mixed-precision performance in the HPL-AI benchmark. That equates to 68 million instructions per second for each of the 86 billion neurons in the brain, highlighting the sheer computational horsepower.
But does the existence of a machine with more raw computing power than a human brain necessarily qualify as some threat to human art practices? In reviewing analogous scenarios in other arenas of human activity, we quickly find that the answer is a likely “no.”
“In October 2017, the DeepMind team published details of a new Go-playing system, AlphaGo Zero, that studied no human games at all. Instead, it started with the game’s rules and played against itself. The first moves it made were completely random. After each game, it folded in new knowledge of what led to a win and what didn’t. At the end of these scrimmages, AlphaGo Zero went head-to-head with the already superhuman version of AlphaGo that had beaten Lee Sedol. It won 100 games to zero.” - Joshua Sokol, Quanta Magazine (If you are interested, there is an excellent article on Alpha Go Zero here: ( AlphaZero: Shedding new light on chess, shogi, and Go )
So if computers exist that can consistently outperform human opponents in Go, Chess, and other games—why are there still people playing these games? According to the author of the article I am addressing—we should assume that players of chess, shogi, and go are currently “dead.” Yet, like the art that we might describe as romantic, these activities seem to be not only alive and well–but thriving. According to un.org, “About 70% of the adult population (US, UK, Germany, Russia, India) has played chess at some point in their lives, while approx. 605 million adults play chess regularly.” Chess.com reports that “From October 2020 to April 2022, Chess.com saw their number of monthly active users double from roughly 8 million to nearly 17 million.” By these sources, it doesn’t seem like chess players are “dead,” even though computers that can be absolutely be demonstrated to have superior “computational power” clearly exist.
Before proceeding, I highly recommend you carefully read the essay I am addressing here (the link is at the top of my post here.) I want to make sure that my simplifications are in no way misrepresenting the author’s position. I’m NOT going to go line by line in my address as there is a dense forest of flowery language that does not seem conducive to that manner of approach. Instead, I am going to simplify statements as best I can. However, If at any point you think that anything that I am stating even remotely misrepresents what is actually written or intended, please let me know. If an error on my part can be verified, I will adjust my post accordingly.
Paragraph 1:
Simplification: Artists were once awestruck by nature and were moved to create grand expressions of the wonders around them. (a characteristic not limited to Romantic artists, but indeed a common characteristic of the genre.) (This is fine. Agreed!)
Paragraph 2:
While this paragraph is really a jaunt through a jam-packed flower shop, I can extract that the author is stating that:
Simplification: the imagination can be seen as an essential feature of the artist. The creations of the artists (born of imagination and inspiration) and the ritual-like creative processes that were assumed to be in play elicited great admiration. (Again, all of this is fine, and I agree!.)
Paragraph 3:
This paragraph begins with, “But the electric storm of artificial intelligence has seized imagination with a breathtaking efficiency. In the new age of picture-bots we have no need for romantics – the machines have replaced them.”
Ok, now we are off the rails. In case you lost track, this is what we have been presented with so far:
- Artists were inspired by nature.
- Artists used imagination and inspiration to create works that garnered much adoration.
- AI has “seized” imagination, and now there is “no need for romantics.”
Wait…what? How? Why? The first two paragraphs are fine in that we have mountains of evidence that artists were indeed inspired by nature and that artists engaged in “imaginative” work, which did garner much attention from the public. However, there is no such mountain of evidence for the problematic claims that follow in paragraph three, and there is nothing in paragraphs one and two that might justify one to conclude that the claims in paragraph three are true or even likely true.
Clearly, (at least to me), I can imagine whatever I want right now (arguments of free will aside), and I can experience joy in doing so. Try it for yourself right now. Imagine something. Did some disembodied AI show up to stop you?
The existence of AI, even if it is infinitely more powerful than I in terms of computing, clearly does not stop me from being inspired, imagining anything, or expressing what I have imagined for myriad reasons.
There is a concept I’d like to introduce here as I feel it is very relevant. (It is something that I plan to introduce in my upcoming lecture as well.) It is a concept known as an evolutionary trap. Within evolutionary biology, this term has been used to refer to cases in which an evolved and presumably adaptive trait has suddenly become maladaptive, potentially leading to the extinction of the species. One such case of this type of trap occurred in Western Australia, where Australian jewel beetles were mistaking a certain type of discarded beer bottle (a 370-milliliter golden brown “stubbie”) for a female of the species. The color and dimpled surface of the bottles were observed to trigger the males into lengthy mating attempts. Males would spend useless hours attempting to fornicate, even dying under the hot Australian sun while leaving no heirs. Thankfully, in an effort to save the at-risk beetles, beer companies decided to change their bottles. The little bumps were eliminated to be replaced by smooth glass, and as such–the beetles lost all interest in bottles.
Now some may read this and be quick to think that AI art, in the author’s view, is the bottle to our human-made art (females.) But upon further consideration, you should quickly realize how wrong this is. The objects in the environment that we dub “art objects” are, in fact, the beetle’s bottles. Paintings, drawings, sculptures, etc., are not metaphorical females here. Instead–they ARE the bottles. They are the supernormal stimuli or “pleasure technology” capable of eliciting experiences shaped by adaptations and consequential tag-alongs.
So then, perhaps the author and others promoting similar alarmist claims are fearful of AI producing “better” bottles, thus leading to us losing our desire to create our own bottles or marvel at the bottles of our neighbors? The problem with that fear is that an aspect of the appeal of our current bottles is that they were made by other members of the species (see essentialism in the art experience definition section above.) Part of our appreciation and aesthetic response to a hand-made drawing or painting is rooted in our assumptions about the process of production. Even without access to the artist, knowing that the creator IS a human, we generate default assumptions (often from our own experiences, biases, availability and representativeness heuristics, etc.) to assess difficulty. Often, the greater the assumed difficulty–the grander the perceived degree of demonstrated virtuosity. With this particular fact in play relative to the art experience as laid out here, it should be clear that AI bottles are incapable of replacing human bottles. AI may, in the future, be able to create “better bottles” according to some criteria, but those products won’t be created by a human. As such, they can never “tick” that particular box for the aforementioned “pleasure bomb.” (Note: consider that if some AI in the future does achieve “consciousness,” why would we expect it would even “want” to make bottles for us? Assuming they would seems to be a fairly narcissistic assumption.)
Paragraph 3 (continued):
Simplified: AI Art generators (text-to-image) are available to the public, and they are easy to use. **They have been used to create millions of images already. **
Quote: “This is death to the romantic artist. Their heroic intercession is no longer needed. Now, access to the imagination is cheap and easy. Picture bots have democratized creativity, and the twenties are experiencing a proliferation of imagery. Art is everywhere, now beauty is a bargain.”
Again, How? Why? Let’s try and transform this latter part of paragraph 3 into the form of an argument:
Premise 1: AI ART Generators are available to the public and are fast/easy to use.
Premise 2: Thousands have already used AI to create millions of images.
Conclusion: This is death to the romantic artist.
Placing the content into this form reveals a clear non-sequitur. The article content that precedes the statement “this is death to the romantic artist” does not support it in any way. In other words, a statement is presented that doesn’t follow logically nor finds support from the previous statements.
In addition, up to this point, the essay has maintained an inherent implication that qualifies as a false equivalence fallacy. This should also be addressed. False equivalency describes an argument or claim in which one shared characteristic between two or more items of comparison is used to attempt a demonstration of equivalence. This fallacy usually involves oversimplification, errors in orders of magnitude, or additional factors that have been simply ignored.
A fallacious argument looks like this:
"Thing 1 and thing 2 both share characteristic A.
Therefore, things 1 and 2 are equal."
So if we “mad-libs” the framework of the fallacy with Mr. Pearce’s content, we get:
"The products of AI generators and romantic artists share characteristics.
“Therefore, the products of AI generators and romantic artists are equivalent.”
At this point, I hope I have made clear that these products are not equivalent. They are not equivalent in the same way that a photograph and a photorealist painting are not equivalent, even if visually indistinguishable at a certain level of granularity.
Now I’m not even past paragraph three, and we are already drowning in problems that are factually and logically detrimental to any potential support system for the central claims of the essay we are looking at.
Unfortunately, there are 10(ish) more paragraphs. Newsflash: the essay doesn’t get any better from here on out in terms of compelling reasoned arguments, facts, supported claims, etc. I am happy to share some of the greatest hits below, though! (Can you guess how many are well-supported? I’ll give you three guesses, and the first two don’t count):
“But the electric storm of artificial intelligence has seized imagination with a breathtaking efficiency. In the new age of picture-bots we have no need for romantics – the machines have replaced them.”
“Anyone with a decent vocabulary can prompt hundreds of images a day, select the best, and take the place of a digital concept artist who has years of training.”
“Artificial intelligence is the god of imagination.”
“A word is worth a thousand pictures, and the prompt is the mother of lies.”
“Artists are no longer the intermediaries of imagination….”
Again, This was a good exercise for my upcoming lecture, but I can’t say I’ll get anything out of continuing on this particular examination. There is more ground I have to cover elsewhere before the event, so I happily hand this particular baton over to whoever is keen on taking it.
I do hope, though, that this exercise was insightful for those of you that take the time to read it.
Happy artmaking to you all!!!
