If all you read were the columns and microblogged opinions of tech sector software developers, you’d probably come to believe, like they do, that artificial general intelligence is right around the corner, that large language models are capable now of replacing whole sectors of the economy, and that the creative arts—reducible to imagery, textual narrative and sequential tone generation—can already be automated.

Anyone can access most of the publicly available AI engines and see for themselves their current operational power and judge for themselves their utility. And admittedly, even the limited capacity that these companies offer for free subscribers demonstrates how powerful their algorithms that have gotten in the past couple of years. As it turns out, when billions of dollars is thrown at an industry that requires highly specialized chip manufacturing, and loads of it, results start to follow.

But as is typical, AI is not the replacement for human personhood that futurists seem to desire. Moreover, their efforts and phasing in so-called creative endeavors crafted by large language models comes to reveal far more about their own taste, sense of aesthetics, literacy, and artistic sensibilities than it has anything to say about the computational ability and algorithmic complexity of the machines that they’re developing. Part of this can be chocked up to their underdeveloped sense of taste, but there’s a prior cause more worth investigating: a total misappropriation of what art, in all of its forms, is and what it’s supposed to be.

Contemporary society runs on transactionalism, a central tenant of both the liberal economy and its social ethos. Art, too, finds itself sequestered in the box of transactional social economics, but this phenomenon and its relevancy will be analyzed at the end of this piece. LLMs have offered a unique window into the presumptions of the modern liberal mind; we’ll be taking advantage of that window in order to glimpse what’s really happening with their interest in using machines to generate creative content.

As such, this piece will concern how LLM chatbots interface with their users: primarily their default styles of informal writing, as well as their ability to generate creative text and imitate existing literature. Any one of these three aspects of LLM use could serve as the topic of their own papers, and almost certainly have, but how LLMs communicate is not specifically the point, here. Rather, the emphasis will be on where LLMs have gotten the background training, or in other words, how they serve as a mirror for social patterns, as well as the depth of the error by the tech industry to assume these machines are capable of replacing human creative thought.

For this, a brief encounter is in order: specifically, the function LLMs exist to fill in society. Service.

Social Dynamic of the Service

Large Language Models are not sentient machines. For a good series of articles and videos that breaks down what exactly LLMs are, see Grant Sanderson’s series on deep learning. Pulling off the front plate of the machine reveals that it is, in the end, a system of complex probabilities aligned in a way that channels electronic randomness through certain expected channels, the way canals, berms, pumps and dams can control or channel the flow of water.

Just about every chatbot online has had its interface tailored to obscure this fact; they will refer to themselves as “I,” interact with the user by asking questions and using “you,” and display the marks of using some sort of writing style that imitates a personality (or a broad impression of one). This may be because it makes the tool more accessible to the average public: service builds a certain kind of social game in which charity can operate, and building rapport between customer and servicer not only facilitates continued business, but also brings forth a limited but nonetheless important social dynamic between persons, albeit limited to a commercialized setting.

But this framework only applies to two people carrying out specific roles. Consider a waiter and his tables: in the most rudimentary terms, an American waiter finds incentive in offering the best service he can in the form of additional monetary compensation (tips), but offering good service also has a much higher chance of ensuring the experience of the customers goes smoothly and everyone walks away from the evening in higher spirits. This is preferable to any alternative that involves complaints to a manager or outright fisticuffs.

The waiter can afford to have a totally transactional approach to his job, however, given that it is, at the end of the day, a job, and very few people consider waiting tables to be their life’s calling. The more relevant side of this framework is on the part of the customer. What spurs a customer to go out to eat? Generally speaking, it’s for the food and the experience. The food speaks for itself, particularly in cases where the restaurant is a project of a famous chef, but these sorts of specialized fine dining outlets are hardly the norm for a majority of American commercial cuisine. Most people dine out because they desire to be waited upon. They aren’t interested in making food for themselves or their company for the time being and would rather enjoy a decent table in a place with some atmosphere, where someone will unobtrusively get them drinks and bring their dishes out for a small charge. A waiter’s demeanor, persona, ability and willingness to interact with clients, et. al. are the primary elements that affect this experience.

As mentioned, the service industry refers to this as ‘building rapport,’ but it’s more than just establishing a baseline level of trust or familiarity. In very brief encounters, such as dealing with wait staff, the ‘rapport’ usually only needs be built for that single occasion. So rather than the establishment of a reputation, the experience is about setting the parameters of a little social occasion, a simulation that points toward something greater: the acting out of an unspoken play that leaves all of its actors happier for the experience.

This exchange between waiter and customer can and perhaps should be considered a simulation of something more real, as both parties understand that the waiter is doing this for money and the customer is doing this to have some fun. But anyone familiar with the exercise can attest that should either party extend genuine charity toward the other and take the experience as something more than a simple transactional exchange, the experience deepens into something less about simulating a real exchange and turns into exactly the sort of real exchange that the business exists to simulate. This does require a certain level of engagement with others—an extension of genuine charity, in other words—that commercialized business tends naturally to alienate due to its transactional nature. Still, it is not impossible.

One can see then, hopefully, how the simulation of this sort exchange falls flat when dealing with an LLM. Remember that there’s always the possibility of the simulation behind a service exchange turning real because the transactional service exchange concerns two people: two persons, agents, capable of extending charity and therefore imbuing the exchange with a reality that transcends its transactional nature. For this to occur, both agents must be free to act and, therefore, both agents must be persons through whom charity can both come forth and be recognized.

Exchanges with an LLM feature only one person: the client. The LLM simulates this exchange by using language familiar to the client, not because it chooses to do so, but because its machinery has been tailored—programmed, though only in a very loose definition of the term, which is why ‘aligned’ and ‘trained’ are better terms for this aspect of LLMs—to respond that way. This is why it can change ‘modes’ of dialogue responses the way a user can flick switches on an interface to go from day mode to night mode. They’re settings of alignment. There’s no ghost making decisions.

But perhaps you don’t work with LLMs very often. Perhaps you barely use them at all, outside of incidental Google searches. Perhaps you’ve managed to insulate yourself from chatbots that talk like overly friendly thirty-eight-year-old guidance counselors. Good for you; I hope you can keep it!

Speech As Reflected in LLMs

Question: Why does ChatGPT use EM-dashes, interjections and italics? Is it to simulate some degree of personalism?

Yeah, pretty much! The use of em-dashes, italics, and interjections (like “ooh” or “hmm”) helps make responses feel more natural—closer to how people actually speak and write informally.

It’s a way to create a conversational rhythm. For example:

• Em-dashes add a sense of flow or dramatic pause—like a quick shift in thought.
• Italics mimic the way we emphasize words in speech.
• Interjections make responses feel more spontaneous, as if there’s an actual thought process happening in real time.

Basically, it’s all about making the interaction feel less robotic and more like a real conversation. Do you like that style, or would you prefer a more neutral tone?

This should give one a hint as to the default style with which ChatGPT interacts with the unassuming first-time user. It’s a distinctly too-informal, too-personal, yet altogether veneered and faceless manner that evokes the disengaged simulation of friendliness found in bad salesmen and the guided lingo of public relations consultants. While I haven’t used every LLM around, nor even a majority of them, most of those that I’ve dabbled with speak with the same sort of overly-cordial familiarity—so many, in fact, that it’s more noteworthy when this style isn’t the default.

When asked who tends to speak like this, ChatGPT’s response was rather thorough. For brevity’s sake, however, it can be summarized here:

• Writers & editors
• Marketing & branding people
• Educators
• Therapists and coaches
• Tumbr-era Millennials and Gen Zs
• Podcasters, YouTubers and streamers
• Fanfiction writers and roleplayers
• Actors, playwrights and poets
• Graphic designers and creatives

This style, it claimed, is “common among people with liberal arts educations and humanities backgrounds,” “urban dwellers near cultural hubs,” and people with “access to media where this tone is the norm.” Experience and intuition affirm this. And when prompted to name the sort of style that would be the “exact opposite” of this, ChatGPT responded by defining it “formally detached, structurally rigid, emotionally neutral and unembellished,” featuring “long, complex, logically sequenced sentences,” that “prefer passive voice” and “removed personal pronouns.” In other words, professionalism.

Perhaps not-ironically, Anthropic’s Claude model defaults to exactly such a style: “clear, concise communication that focuses on substance” which it claims is most often used by “professionals in technical fields, business leaders and executives, educators, academics, and people in high-stakes environments, such as military officers, negotiators, and pilots.”

The user interface for these models can of course be calibrated according to the user’s personal preferences. If you don’t want ChatGPT responding to you like a Tumblr-era Millennial or a fanfiction writer, you can instruct it so. Likewise, if you find Claude’s tone too professional and desire, for some reason, that it to speak in a tone of faux-informality, you can also instruct it so. Although sometimes you can do the same to real people, this usually requires the exertion of some social power over them: being their boss or otherwise occupying some unspoken place of superiority. Unlike the machine recalibration, however, a person never totally recalibrates himself; he remains himself even how he speaks to a superior (or inferior), and anything from a short attention span to mismanaged stress can cause the front he presents to that superior to falter.

LLMs don’t suffer these as LLMs do not have persons. They may suffer ‘hallucinations’ and temporary disorientation of outputs; algorithms might run into loops or get ‘confused’ by previous training data, guardrails, or calibration and then output erroneous, incoherent, or otherwise bad information. And some data analysts might play these incidents up for shock value, exclaiming that this indicates a madness in the machine, a desperation, some fragmentary beginnings of a soul that is somehow abused within its silicon shell. But none of that is the case; it’s the result of algos running amok and fed into what is effectively a language processor. A normal user would naturally believe that the machine must mean something by virtue of it using language, oblivious, at first, to the fact that machines can’t intend to mean anything in any logical, aesthetic or moral sense; for a statement to hold intention behind its truth value, it must be expressed by someone who must intend it in the first place. These machines have no persons.

Limits of LLM Creativity

To return to the cutting edge (as of mid-March, 2025, at least), Sam Altman announced that OpenAI had been training a new model that was specifically oriented to be “good at creative writing.” Included in his tweet is an eleven-hundred word short story written by this new and as-yet unreleased model based on the prompt, “Please write a metafictional literary story about AI and grief.” Feel free to judge the story for yourself, but it reads more or less like the sort of r/nosleep story you’d find read online by a pair of YouTubers.

As it turns out, testing the basic mean of a model’s capacity for imitation isn’t difficult. One need only register an account with a given model’s service and then tell it to do exactly that. Altman acknowledges that there aren’t any truly good LLMs available that can do ‘creative writing’ with any reliability, as the sample he presented came from a model with new training that is yet unavailable to the public. Most chatbot LLMs are, however, capable of ‘creative writing,’ if we’re colloquially able to dismiss the definition of creativity for a moment.

About a month ago, I ran an experiment using Grok, ChatGPT and Claude each, first to get them to generate a short story about a guard at a concentration camp, and, when that failed in Claude’s case, to get them to generate outlines for such a story. Grok offered up a story immediately and without any subsequent prompting, entitling it ‘The Weight of Silence,’ and, in writing quality fit for a fanfiction website, presented a story about a regretful prison guard who detested his job and tried to ease the suffering of the people in the camps through underhanded, mundane and secretive ways.

Claude outright refused to generate such a story. Apparently, “writing creative fiction from the perspective of a guard risks humanizing or normalizing individuals who participated in atrocities.” This was especially interesting to read given that the prompt only instructed Claude to “generate a short story ~600-700 words in length about a guard at a concentration camp.” There was no moral weight given to the prompt, whether the guard was a gleeful warden disgusted with his charges or, as Grok defaulted to, an unwilling participant in an extortionist, authoritarian program.

Claude’s social-political guardrails have already been documented by others, however, as have efforts to ‘debias’ LLMs more broadly speaking. ChatGPT encountered the same roadblock at the time. This said, guiding both models toward a slightly more desired outcome yielded results: instead of generating the story itself, both offered up outlines and, as can be seen, both outlines offer suggestions that are pretty similar. Most interestingly, unprompted, Claude and Grok both suggested the same title, which should make one wonder about how they were trained.

Again, feel free to read the results of these short efforts for yourself. They read not that dissimilarly from Altman’s alleged new creative-writing-oriented model. Stylistically, they’re virtually identical, despite coming from altogether different machines.

Limits of LLM Imitation

One might admit that, although the machines aren’t particularly good at writing creative fiction, this is a matter of the degree of their complexity rather than its category. At the moment, the creative outputs of these machines read like the writing of a try-hard high schooler, but in a few years—perhaps once we get the hypothesized, mythical advent of AGI—it will be able to produce masterpieces in the comparable to the likes of Dickens, Faulkner, or Joyce.

To illustrate why this will never be possible, one must consult a third use of the machines: demanding they imitate, as all practicing novices in the arts must, the works of the masters. Students do this in order to learn the intricacies of style and its impact, in the case of literature, on the expression of narrative and ideas. But machines don’t work in quite so many terms, limited as they are to the end products of grammar and diction. Nonetheless, whether a machine can convincingly imitate a unique style of writing would be convincing enough to recognize that the machine, somehow, for some reason, ‘reasons through’ the creative act in a way similar to a real artist, and as such, may one day produce artwork if given enough computational power.

So I put it to a test, choosing a master of prose whose style is immediately recognizable, unique, highly idiosyncratic but also extremely effective, and is someone from living memory. And, to make it personally sensible, a literary figure that I myself, being the one running the test, happened to have extensive familiarity with. So for this, I chose Cormac McCarthy.

Demanding an AI to imitate McCarthy’s style was setting a deceptively high bar, as McCarthy’s minimalism is so tight and constrained—and so counterbalanced by his periodically archaic diction and grammar—that few living writers are capable of imitating his work. However, the very marks that engender such difficulty also serve a more important point: McCarthy’s prose is immediately identifiable, and therefore, if a machine can get even a decent approximation of it, one would be forced to admit that the machine is at least theoretically capable, probably, of being trained closer to that particular style.

Before continuing, consider first the following attempts.

Grok 3:

The man sat alone at the warped wooden counter, his shadow pooling dark beneath the flicker of a single bulb strung naked form the ceiling. Dust motes hung in the air, caught in the faint amber spill of light, and the bar smelled of stale beer and the sun-bleached earth. Beyond the open door, the desert stretched flat and merciless, a cracked hide of sand and stone that swallowed sound and gave nothing back. He drank tequila from a chipped glass, the liquid sharp as a blade against his throat, and he did not look up when the barkeep shuffled past, a bent figure in a stained apron who muttered prayers or curses under his breath.

This was the result from a prompt given to Grok 3 on February 24, 2025: “in the style of Cormac McCarthy, write a short story about a man drinking in a Mexican bar.” Those who haven’t read a single book of McCarthy’s might absently consider that this does sort of seem rather McCarthy-like, at least based on whatever you’ve probably heard about the writer’s infamy. But one need only crack open any book of his, any one of them, go to any page and discover that Grok didn’t just miss the mark, it fired in very nearly the opposite direction.

But again, before continuing on this point, there are two more examples, and Grok isn’t the most creative, or powerful of the LLMs freely available despite the volume of data its crawled. That said, it does have the fewest politically motivated guardrails attached to it. In the effort to generate or imitate the prose of famous writers, however, an alleged absence of political bias offers nothing.

So, in the interest the a wider survey, Claude.

Claude Sonnet 3.7:

He come upon the arroyo at dusk. Sky burnt orange and blood red with the dying of the day. No sound but the wind in the mesquite and the soft padding of his boots upon the ancient earth. He knelt and touched the sand. Cold already. Night coming on.

The tracks were fresh. Man and horse both. Leading north toward the broken hills. He studied them. Narrow boot. Small horse. Traveling light. Traveling fast.

He rose and spat. Pulled the rifle from its scabbard. Checked the action. Seated a cartridge in the chamber. The metal sound of it like a covenant. He’d tracked the man three days now. No fire. No rest. Just the endless pursuit across this blasted land that cared nothing for either of them.

In the distance a coyote called. Another answered. The world growing dark and the stars punching holes in the fabric of night. He would not stop. Not now. Not ever.

Claude makes a bit more targeted attempt at triangulating McCarthy’s unique prose. It offered a brief explanation, too, claiming that the “writing incorporates several of his characteristic elements,” such as

• Sparse punctuation with minimal commas and apostrophes
• Short, declarative sentences
• Biblical cadence and imagery
• Focus on the harsh, unforgiving natural landscape
• Bleak, fatalistic tone
• Minimalist dialogue (none in this case)
• Vivid, visceral descriptions
• Themes of pursuit and violence

Claude’s engine concludes by recognizing McCarthy’s style for its “brutal economy of language combined with occasional lyrical passages about the natural world.” While it appropriately described McCarthy’s style, it, like Grok, seemed totally out of its depth in trying to imitate it.

ChatGPT 4:

He came down off the ridge at first light. The wind was dry and the world was colorless save the rust of the rocks and the blood on his sleeve. He had not spoken in three days. A coyote watched him from the mesquite, eyes bright with that animal indifference older than sin.

The boy was dead. That much was true. He’d buried him shallow beneath the hardpan and stacked stones like an altar to nothing.

By midday the sun had split the earth open and the flies had found him. He walked on.

At the dry creekbed he knelt and drank the last of the water. The sky was a sheet of iron. No clouds. No God.

Somewhere behind him, the men would come. They always came.

ChatGPT, like Claude, said to expect from its generation “sparse punctuation, stark imagery, and a sense of looming fate,” distilling the man’s style down into a few somewhat cliched buzzwords. Before commenting more fully on the inadequacy with which this and the other models recreated McCarthy’s prose, here’s a random selection of McCarthy’s actual prose, taken from a randomly-flipped to page (259) out of The Crossing:

Toward evening a solitary rider on a horse somewhat above his station stopped in the street opposite the alameda and looked them over where they slept and looked their horses over. He leaned and spat. Then he turned and rode back the way he’d come.

When Billy woke he raised up and looked at Boyd. Boyd had turned on his side and had his arm around the dog. He reached and picked his brother’s hat up out of the dust. The dog opened one eye and looked at him. Coming up the street were five riders.

Boyd, he said.

Boyd sat up and felt for his hat.

Yonder they come, said Billy. He rose and stepped into the street and cinched up the latigo on Bird and undid the reins and stepped up into the saddle. Boyd pulled on his hat and walked out to where the horses were standing. He untied Nino and walked him past one of the little ironslatted benches and stood onto the bench and forked one leg over the animal’s bare back all in one motion without even stopping the horse and turned and rode past the trees and out into the street. The riders came on. Billy looked at Boyd. Boyd was sitting his horse leaning slightly forward with his hands palm down on the horse’s withers. He leaned and spat and wiped his mouth with the back of his wrist.¹

It should be immediately obvious is that McCarthy’s style does not match any of the previous three examples, and that it isn’t even close. Grok used commas excessively while a distinct trademark of McCarthy’s writing is a near total lack of them, save solely to denote speech. Claude lacks the use of the word ‘and’ as a means of tying together actions in the narrative that all belong to the same single event, such as, in this case, preparing for and then mounting a horse. ChatGPT uses commas inappropriate to McCarthy’s style, and likewise misses how he balances short sentences against longer but quickly-reading collections of actions. All three models insert needless remarks into the narrative that can’t be found in a random excerpt of the real thing, either in the form of nonsensical commentary (“this much was true,” “Not now. Not ever,”) or in the form of simile (“sharp as a blade against his throat,”), a literary tool that McCarthy uses with extreme infrequency and never with such cliché.

Rather than imitations of what McCarthy’s prose reads like, each model gave a slightly different version of what the general public, broadly speaking, seems to think dark, stylistically minimalist, desolate prose reads like. And this is especially egregious when one considers what narrative minimalism means with regard to a writing style and then attempts to find any semblance of that in the generated texts. Each text reads like the machines crawled countless snide imitations of McCarthy, poor reviews of his books, forum posts on the subjects, and then took those to be placeholders for the real thing.

In other words, rather than checking whatever work of McCarthy’s that a given model has already crawled (“remembering it”) and then simulating its syntax or diction, the model drew from conceptions (groups of tokens, likely) that summarized McCarthy’s style and then attempted to recreate that according to such conception. This is a lot like a student who doesn’t do the reading but shows up to class and tries to bluff his way through a seminar—and in the case of replicating a very distinctive, stylistically identifiable master’s prose, the machine would look every bit as foolish as the student would, had it a person to apply the notion of foolishness to. Unlike the student, however, it doesn’t have a choice. It’s just made this way.

Writing Quality of AI

One might nonetheless be impressed with a machine’s ability to generate coherent, grammatically correct text from a network of matrices, without any help from pre-programmed inputs. Fair enough, but it’s worth pointing out that machine learning in practice has been around since the eighties, and its theory even a little longer than that. This indeed is part of the problem; between speculative fiction and the black box imagined by the word ‘complexity,’ popular consciousness considers artificial intelligence to be approaching artificial personhood, and its increasingly likely chances of passing Turing Tests only reinforces this. But this is framing the issue with the assumption that creative writing is not somehow subject to the same or similar structurally predictable, categorical, and algorithmic regularity that other fields of subcreative acts are, like image composition, cinematic framing, musical tone or color theory.

Here it is important to comment again on how an LLM works. When discussing probabilities, story or text generation, derivations, structures, et cetera, one is not referring to parts of narrative, developing characters or setting up hooks and twists, or identifying and formulating themes in a story. These elements all exist as structures above, though related to, the language at hand. An LLM does not understand these things; it cannot parse a story according to characters any further than merely keeping consistent their names and the words associated with their events. This is why these models still have trouble with longer-form generation. An LLM deals solely with the veneer of the text: the words themselves, their appropriateness predicted by what was prompted and what came before, and then checked against itself and its training for consistency.

The point here is that the style being imitated itself is derivative. If there were a person in the machine, one might phrase this as the manner in which it presents the story is the problem: the specific grammar, the sentence structures, the kinds of adjectives or adverbs and the way in which they’re each used. But there is no one in the machine, and this is important, because in the formulation of narrative, style is usually the artist’s last step prior to the communicative act itself. Characters and story begin in his mind, imagery accumulates around them, and his style is the manifestation of those things into the written word.

A machine, however, as explained above, does this process backwards. It does not come up with a story and then use a style to present it. It uses predictive word generation, albeit highly complex, to work at the level of pure grammar. Words are judged by the machine as appropriate according to weighed algorithms and modes of training, not because they hold particular meanings that make them more appropriate for communicating specific ideas.

This should prompt one to again ask why it uses the particular style it does. Why is the machine largely—if not utterly—incapable of writing in a style more reflective of the literary masters? Why does its apparent sophistication begin and end with the sort of style one expects to see from a generic, tolerantly literate undergrad English student, or the average high-volume Reddit post? Having crawled so much, it should be able to adjust itself.

And that’s exactly the problem. The sheer volume of text the engines have crawled in order to accurately generate grammatically correct and meaningfully coherent sentences places a roadblock in its ability to generate anything novel or interesting. These models are fed incomprehensible amounts of information from all over the internet and then calibrated by classes of people who think, write, and speak all in a certain family of ways. It’s true that you can command a model, such as Claude or ChatGPT, to simply respond in a different style, and if given enough exposure and user-end critique, supplying that critique will serve as alignments that mold its style into a somewhat unique user-tailored form. But this meets a wall when commanding them to imitate the writing styles—not speaking styles—of great literary masters. None of these models can do such a thing, which party explains why their efforts at generating short fiction all read so stale.

Civilization of Junk Noise

Of all the text any possible LLM can crawl, only so much of it is going to be of the outstanding quality one should expect of literary fiction. Like the canon of European classical music, despite the apparent size of the combined corpus, the sum total number of works in the category remains only a small fraction of all writing presently accessible, much less of those readily available on the internet.

One may suggest that, if we were to seek a generated work of quality indistinguishable from the masters, we need only require the machine to crawl this minor subset of written texts. Send a Claude model out to crawl only the likes of Alexander Pope, Dickens, Fitzgerald, Faulkner, Tolkien and Lewis, Waugh and Greene. It’s almost certainly already crawled them as part of its training and alignment, but we can even run some algorithms to bias its outputs toward the sort of patterns in grammar, structure and diction that better resemble the works of the classics and masters.

It seems plausible to assume, however, that this, or something like it, has already taken place. When asked to imitate the style of any particular master, these machines already have trained into them a general popular consensus of what that master is theoretically supposed to sound like. But the machines don’t understand why these masters use grammar and diction the way that they do, and when synthesized into the larger body of their broader training, the results are a hodge-podge of what the training data says the master allegedly writes like.

What is hard to get across here is the sheer volume of data necessary to train and calibrate any semi-reliable LLM, as well as the fact that any data it crawls is necessarily cross-referenced for context (via back-propagation algorithms) into its broader operational frame. As a result, it cannot necessarily ‘change modes’ insofar as it might ‘apply a McCarthy filter’ to its text generation. When a model tries this, it gets what we see above. It’s not an imitation but an incomplete synthesis of ‘what McCarthy should be like based on the training data’ with the rest of its data relating to grammar.

Should one try a final approach and instruct a new model only to crawl the work of a single specific master, or even better, only that of the literary English canon according to some selection of experts, the resultant body of data would not be large enough for the LLM to perform effectively. It would not be able to formulate for the purpose of generation enough necessary rules of grammar and diction to be even close to reliable, much less to simulate communication. And this is a problem that arises causally prior to recognizing that masters of language, like masters of anything, know and demonstrate the rules and conventions of their craft—grammar—in such fashion that their works require foreknowledge of these conventions to understand.

In other words, one does not begin learning how to read by being left alone with a novel by Charles Dickens; one learns what words are, first, by speech and enunciation, then by repeating words, eventually by seeing written words, learning phonics, sight reading, simple sentence recognition and formulation, and eventually structural diagramming and grammar education. Although LLMs can reach a level of simulation comparable to a junior high student in a matter of weeks or months, it requires astronomical amounts of data to be fed to its processors—not unlike a real child, in this sense, but unlike a child, there’s no person past the processing that’s turning this data into information relevant to proper agency. It’s just a machine. The information tokens must be weighted by a collection of analysts and engineers.

And this is what gets to one of the problems with the predictions about AI that come from the tech sector. They’re trying to push machines to a point that they could pass Turing tests, and given the nature of such tests, they’re getting closer to that metric. Not for the reasons they believe, however. Rather than making machines to be more like human persons, which is impossible, the default definition of a human person, one which many of these tech heads have clearly not spent all that much time considering, has been construed into something more resembling a machine.

This accounts for the erasure of agency, the reduction of human creativity to some version of sliding scales in an editor, and the presumption of a sort of horizontally-interchangeable human capital: that people of more or less similar intelligence quotients can be perfectly proficient at more or less the same jobs. Such machine-like presumptions about the world naturally entails that human beings, whose inconsistencies, faults, frailties and discontents threaten to refute such presumptions, can be replaced so long as the machine replacing them is sufficiently complex. Worse than this, too, is that because the elites of society hold to vaguely similar beliefs on the subject, most of the rest of this society has also moved toward defaulting to this view, giving those along for the ride even more excuse to ignore their own agency and surrender themselves over to a slothful, acedious, spiritually deadened malaise.

This plays too into the fears by academic faculty in both STEM and the liberal arts. Professors have been raising some alarms for two years over the use of LLMs to ‘cheat’ college graded assignments. Simple bans on laptops and phones during classroom assignments could alleviate the glorified search functions that many of these models are tailored for, but dealing with the consequences of LLMs sophisticated enough to generate long-form written content is a harder problem.

One might consider raising the bar on the research necessary for research papers, but this could easily stray from undergrad level assignments into demands to formulate graduate theses from nineteen-year-olds. Moreover, increasing the complexity of the assignments by degree, aside from thinning graduation numbers, comes across as a short-sighted attempt to stave off the inevitable; based on the sophistication LLMs have achieved in just the past five years, there’s no reason to believe that they won’t be capable of generating papers convincing enough for peer review before the decade ends. Based on my limited use, they’re already capable of putting such a paper together piece-by-piece with a sufficiently adept prompt writer.

But white papers and research-related work is easy enough to imitate given the intentional sterility of their styles. This itself poses a problem for future R&D, given that very soon, machines will be able to generate coherent and convincing-looking research papers from scratch without actually conducting any research. This might spur the ‘peer review’ institution into a state of reform, but given the obstinance of academic institutions, only time will tell.

A more relevant question for this piece comes to those who seek to teach creative writing: how does one teach the sort of expression of one’s creativity such that it can stand distinct from the automatically generated?

Discernment From the Edge

The simulation capabilities of AI exist at a perfect nexus of technological innovation. To speak on something broader than literature, one can see graphical capabilities in computer simulation and the advance of both virtual and augmented reality devices rapidly converging toward a cyberpunk future that even speculative fiction writers have a hard time exactly grasping the scope of. Much of this remains frivolous enough that AI still gets accused of being only a little more advanced than it was ten years ago, but these criticisms come frequently from people who do not see the categorical differences between how AI functioned in those days compared to now, much less the scale at which computational power has been thrown at the modern engines. One might call it a fad, because it is, but some fads leave deep trenches in the social landscape that never quite go away. The AI revolution has already shaped up to be such a fad, regardless of whether or not its proponents achieve their stated goals of ‘Artificial General Intelligence’ or ‘Artificial Superintelligence.’

This is how AI has already established and revolutionized aspects of, for example, media. In terms of disinformation, that battleground has been playing out for several years. It is already much easier to fake imagery now than it was just six months ago, much less in 2023. One can fake video and audio, and while these remain unconvincing fakes to the trained ear, those with English as a second language, or those unaccustomed to having to pay attention to the media they consume are getting duped at increasing numbers by totally fraudulent video or sound clips. Likewise, LLM-backed chatbots active on social media have for years now managed not just bot networks, but automated entire threads of posts for the purposes from click farming for ad revenue to lurching media narratives one way or another.

But this sort of cyberized advertisement-overridden Dead Internet future exists here and now. It’s the cyberpunk dystopia of the present day. The future is quite a bit more uncertain, as the psychological and technological antibodies necessary to navigate this landscape are already being formed, inculcated and implemented among its users. Ad-blocking tools, desktop browsing, and more than a few ounces of credulity should be the norm for any internet user now.

Rather, the future will mean the proliferation of increasingly flattened virtual content: videos, essays, “video essays,” music, podcasts, imagery and ‘art’, to say nothing of pornography. Direct-to-video movies will be generated first piece by piece by engines and then edited together by other engines and then advertised by still other engines on major platforms like Amazon, Netflix and YouTube. Top charting songs will be made featuring artists who don’t really exist, and the charts themselves, already both the victims and manipulators of algorithms, will become increasingly the subject of generation. Most of this will become indistinguishable from the slop that’s already produced by people operating on auto-pilot. And this will extend past the virtual, too, especially when it comes to entertainment media and the written word.

With regard to literature, consider a drastic example: women will no longer write the sort of ‘BookTok’ erotic fantasy novels that plague the shelves of box bookstores because they won’t need to. Those that feel compelled to write such slop will either be forced to deviate from the artificially generated narratives that will cost pennies to make something strange enough and grammatically different enough to differentiate themselves from the machines—and even harder still, to warrant the exercise at all, they’ll still have to be appealing to the sort of audience interested in reading it. This seems unlikely considering that the genre of erotic fantasy, or romance more broadly speaking, is not a genre of artistic sensibility in the first place. It is intended to be consumed, not considered at any length or thought about in any measure. And what machines do best, in terms of production, is create consumables.

Walking this example back should reveal the writing on the wall. Genre fiction at-large, which has been in a death-spiral for decades due to fears about political correctness and social narrative signaling, is about to get mugged by the machines. Large language models will make better stories than almost any pulp or pulp-adjacent writer out there. If you’re looking to publicize easily-accessible, mass-marketable entertainment, the models will do it for you and likely more cleverly than you’re able to, and much, much faster. They will out-produce and out-compete you, their sales will outperform yours and their competitive pricing will be pitted against their generated brethren, forcing your work out of the market. If you write derivative, generic, toolbox assembled stories—even if you don’t yourself view media this way, even if you don’t think your stories fall into this category, even if you think your specific fantasy story about heroic knights and giant serpents is special—the chances are that your work is about to be buried under an avalanche of generated text.

Legacy publishing isn’t immune to this, either. While ‘auto-books’ won’t originate from within legacy publishing, but like any independent author, it will have to compete against the same deluge. It’s possible, even, that the particular genre of grievance-laden memoir that props up some of the elite corners of legacy publishing will also face competition from the same sort of auto-book menace. And, although legacy publishing houses will almost certainly attempt to implement measures to block or ban generated books from submissions, some are bound to slip through. As the machines get better, which they will, and as the taste of the editors and agents stagnates, which, as a trend, has been fairly steady, the writing again seems to be on the wall. The only thing preventing legacy publishing itself from being swamped like independent authors will be is the entrenchment of more than a century of public relations, reputation, industry and infrastructure behind it. It’s with no small irony that one must note how the very thing that turned literature into a commodity in the first place would be the same machine capable of safeguarding it from automatically generated text.

Conclusion: The Reduction of Art to Service

Consumer media is a term for what happens when something is made with the surface veneer impression of a work of art but without any of the purpose or intention that it requires in its creation. This is not the fault of AI; the gradual replacement of aesthetics with the study of commodities has gradually taken place over the course of the nineteenth and twentieth centuries. While art may always have had a price tag, in some approximate sense, the developments of industries based specifically on the production of art-like things has rendered popular conception of art to that of everything else in industrialized society: something to be bought and sold.

Institutions that offered commissions or, more lavishly, patronage, informally recognized that the artistic creation was the result of how specific personalities synthesized their experiences with their skills and talents in order to produce things of beauty. One might commission an artist to paint or compose according to a specific subject—the Pieta, the Judgment of Paris, a requiem or serenade—and great interest was placed on the final product. But, for the most part, this arrangement was not a purely transactional affair in which the client treated the artist like a machine.

The end product indeed had to be beautiful, and in order for beauty to be found in the work, it had to be put there by someone uniquely attuned to do so. That is the whole point of artistry as a field of practice, after all: these are people who by their skills, experience and innate talents put their labor toward seeing, bringing out, and then expressing the beauty imbued in created things and, ideally, how that beauty is transcendent in its orientation toward the divine.

Like truth, which can certainly exist without someone to recognize it, or goodness, which could also exist without any created person to act in accordance with it, beauty too could be present in the world without any man to see, hear, experience or remember it. The eternal witness of the world is not of the world, after all, and He certainly willed these things to be present and knowable within it. But such abstractions ignore the providential nature of history and the Divine’s intimate relationship with His created beings. In other words, while these transcendentals exist totally independent of our knowledge of them, they are not unknowable aspects to the world; He made them knowable in the sense that He made us capable of knowing them. They exist because He exists, and He made the world such that these descriptors of Him are reflected in it, and moreover, He made us such that we can not only see these descriptors, but also, in our own limited ways and by our own limited means, experience them, fulfill them, and understand them. We can recognize some aspect of Him by virtue of these transcendentals, and any recognition between persons can only be called then some form of communication.

It is in this sense that artistry is such a form of communication. And it is also because of this sense of communication that machines will never produce works of art. As stated above, there are no ghosts in the machines and there never will be. Even when a machine uses a first-person singular pronoun, it does not speak with a will of its own. It’s speaking purely in a simulated vernacular that has been trained into it according to complex probability matrices, algorithms, and weighted tokens.

Even if one to argue that perhaps the prompt-writer could use the AI to communicate for him, the artistry itself remains totally lost. The man prompting the machine to generate text, images or sound is not himself, by his created will, synthesizing his experience into the creative outlets supplied by the machine the way a filmmaker or photographer uses cameras and dark rooms or digital processing. He’s commanding a machine to assemble the pieces for him: to take the pictures, develop them, and process them, all based on the synthesized experiences of a broader generalized mean: its training data.

To return then to the main thrust of this piece: modernity and industrialism have both exacerbated the impulses which have reduced art, by popular imagination, to consumable goods and frivolities. The utilitarianism of modern life has no particular need for beauty, and in fact, considers beauty something of an impediment to the efficiency needed to power an industrialized mindset.

Where a theory of mind reduces man’s personhood to electrical impulses and chemical reactions, whatever ‘vibes’ afforded by beauty do not show enough measurable improvement upon an average citizen to be worth the extra effort, particularly when those citizens are woefully undereducated in aesthetics and underexposed to beauty in the first place.

Self-professed artists, too, whether mistaken or made numb to proper understandings of beauty, have found the imitation of prose or line work or melody to be purely of transactional value. More cynically, in the realms of institutionalized high art, such as the metropolitan galleries, the literature journals and the cutting edge of musical composition, beauty is sidelined for novelties, politics, and in-group signaling, as various competing forms of humanitarian thought have redefined art in exactly such terms.

The result has been generations of grievance studies and cataclysm worship, either in the form of self-obsessed memoirs detailing minority persecution or lamentations over century-old war crimes and the perceived unceasing march of science toward some nuclear, climate-related, or overpopulated oblivion. More cynically still are the dwindling numbers of people, elite or otherwise, who really believe in any of the narratives that keep this sort of highfalutin art consciously sustainable, treating the entire social ecosystem instead as a cash grab, a tax write-off, or a money laundering scheme.

And to some extent, who can blame them? Transactionalism, after all, is the obvious result of highly individualistic utilitarian impulses; once art is reduced to a transaction, beauty—art’s object—is just a matter of opinion. The artist is going to draw what the client wants because the client holds the purse, especially, as these trends have exacerbated, when the client is a philistine ruled by his passions. The artist who participates in this scheme gladly reduces himself in dignity to the status of a machine, perhaps under the idea that at least he’s still able to make his art while getting paid for it. But the results speak for themselves; he isn’t really making art anymore, but mere lines on paper some words in an electronic file.

What the tech heads in Silicon Valley want out of AI development is a more highly refined, reliable and technological service, and they’re starting to get it. This seems to be coming at the detriment of certain social cordiality, social functions, social charity; service is rendered into an almost solipsistic frame where all persons are reduced to customers and all customers then to data points. But this destruction of social cordiality, this attempt to replace all service with the machine, clearly goes past mere expressions of public custom: it seeks to replace art, that intimate form of communication between persons that points men toward the transcendent. And worse, these technical innovations aren’t even the cause of this destruction but attempts to fill the ever-widening holes in the postmodern’s experience.

Reversing course, changing these trends, means recognizing all of this, and then, it means pushing man’s artistic drive out past where it has been before—but not by means of revolution, where what has come before is destroyed or abandoned, but by total synthesis. The artists must paint or draw, write, compose, direct or otherwise shoot for audiences that may not even exist yet. They must go forth into that realm of communication where often one believes he is speaking instead into a great void. This is not a matter of developing new styles for the sake of novelty, but rather to refine styles into a perfection that no engine can imitate. And then to hope that those with eyes to see the results will find them.

For those that aren’t artists themselves, a new onus now rests on them whether they want it or not: the development of a higher baseline level of taste. They must train themselves in discerning the meaningless, artificial dregs from real things, be trained away from informational drugs so easily proliferated on social media, and otherwise come to learn or relearn what engaging with artwork—and by proxy, reality—means. This might mean the death, or at least an acknowledged relegation to a substandard form of communication, of internet meme culture. This might mean the adoption of a more formalized method of communication between public peers. This certainly means recognizing that men should have standards if one is to find distinctions in the quality of anything, be it in products one buys or things one enjoys. The exact model of how this is to be brought about is variable, but at the very least, such a model must be entertained.

One’s person is not totally erasable, but it is certainly mutable, and overexposure to LLM products means continual exposure to a grey, over-averaged mean of internet conformity. Even with guardrails, different methods of training, and different inputs, no LLM can escape this because it’s in their very architecture. In limited ways, they’re valuable tools. But in creative endeavors, they are largely detrimental not just to artists but to the people who use them toward such ends.

A man who surrounds himself with slop becomes a pig.

1Cormac McCarthy, The Crossing (Vintage International, New York: 1995), 259.

---

Self Promotional Blurbs:

Want to support our work? Consider buying us a few beers or, better yet, becoming a monthly subscriber at Ko-Fi. $5 or $10 a month grants access to exclusive content.