SIX DISPATCHES ON CREATIVE WRITING AFTER CHATGPT
Approximately two years ago, although it seems like yesterday, a student of mine in a creative writing course performed an experiment.
The experiment was technology related.
I was its subject.
Instead of writing the short prose pieces all participants in the class were assigned, the student used a popular large language model (LLM) he’d accessed online to generate a gappy text he presented as his own unassisted work.
The syllabus stated unambiguously that this wasn’t allowed: You could use an LLM but you had cite the model explicitly. (The student had been present when the syllabus was distributed and read out loud.) During the second-to-last (or nearly) week of classes, he admitted that he’d used “AI” to write his previous submissions. He contended that because I seemed not to have noticed, this proved that literature was an outmoded construct we’d all do well to be less deceived by.
As a person who has long valued literature precisely because it’s an artificial thing that’s always going out of style, I couldn’t have been more sympathetic—at least on the surface—to this student’s claim. Literature is eternally outmoded and vulnerable to accusations of falseness and, truly, the student had touched on the single aspect of this abstruse art I hold most dear. I love that literature does not have to be true, indexical, or straightforward. I love that it is invented and that we are at liberty to fashion it as we see it, including in ways that may cut against the grain of progress, so called. Frustratingly for me, the student’s creative methods put him in violation of the university’s policies on academic integrity. Also, the stuff he’d pasted into Google Docs was tough to understand in the light he professed to intend it and (this is probably obvious) did not do much to advance his very interesting suggestion that the emperor has no clothes. Just as typing THKOLAALKDngka dksjfl gjkis, a semantically meaningless alphabetic string in US English, does nothing to retroactively cancel the meaning of the sentences that surround it, so asking ChatGPT to compose a sonnet does not instantly eviscerate eight hundred years of prosody.
As I waded through the administrative consequences of the student’s experiment, I found myself pondering my own (now antique) undergrad obsessions. I remember in the year 2000 reading lines from Milton’s Paradise Lost that stunned me. God speaks to Eve, who is baffled and enticed by her own reflection. “What there thou seest, fair creature, is thyself; with thee it came and goes.” In a way, it’s odd that the law-obsessed Old Testament God would offer His lesser handiwork so much intellectual liberty, but here’s a prime irony of secular life: That which we ostensibly know most intimately (the self) is also an entity we’re condemned to know only partially. Eve’s innocent confusion, quickly passed over by the poem, recalls the ancient myth of Narcissus, even as it looks forward to Freud’s theory of the unconscious. Nor does this depiction of human uncertainty seem totally incompatible with what the American poet George Oppen would write three hundred years later: “The self is no mystery, the mystery is / That there is something for us to stand on.” Objective knowledge of the self is obtained with difficulty and, meanwhile, easily confused with other things. Objective knowledge is problematic and rare, in general.
Lyric distillations like these are the reason I got into writing. The ethics of linguistic description never ever get old. At least, not for me.
I should say here that most, if not all, of the students I’ve encountered in various contingent teaching positions over the past decade don’t do the sort of thing the student who used an LLM to complete his creative nonfiction assignment did. If I can be forgiven for generalizing, the reason they don’t is that they are aware that their knowledge of themselves is partial. Their knowledge about themselves in relation to their community, in relation to history, in relation to citizenship, in relation to the laws of thermodynamics, and so on—isn’t complete. This is extremely interesting to them. Students are simultaneously aware that using a probabilistic mathematical affordance to generate lists of words is unlikely to provide them with much information about their own perspectives, losses, or obsessions; nor will it help them contemplate the phenomenal world, in all its anarchic squishiness. I’m not saying that LLMs can’t produce text on many a subject and be, like, impressive summarizers. It’s just that they will be so dramatically worse at describing my feelings about, say, my 2014 breakup than I am—in the absence of even now inconceivable levels of surveillance, access to currently paywalled or private databases, and fantastical processing power—that few people in creative writing classes use them for that sort of thing, except as a joke. And no, I’m not going to enter a command into ChatGPT—“Compose a braided essay in the style of Anne Carson about a jerk who loved Tyvek® jackets”—to see what happens.
When people write, they tend to think of the words they choose as their own thoughts. When they have difficulty writing, they discover that their thoughts (and by extension they themselves) are not the same thing as the language at their disposal, which is a very curious predicament to arrive at in a class, particularly if you did not contemplate it previously. Descriptions are inevitably incomplete, threatening to reduce the world around us; what will the writer do about it? This is something I like to explore collaboratively when I teach. It’s an artistic question and, significantly, an ethical one. How can we write while taking into consideration the complexity of human needs and vulnerabilities? How can we describe others in a way that accounts for their unpredictability?
The language model itself isn’t the problem. But LLMs are attached to a host of profit-driven practices and corrosive orientations to both privacy and the public sphere, nor are they currently being developed in collaboration with the vast majority of communities from whom they derive their training data and/or where they are used. (It is, for example, very hard to express yourself creatively if there are no public places in which to do so, if you cannot speak freely, if you cannot expect to be compensated when your IP is used illegally for corporate gain, or if you have no leisure time in which to compose your expressions. If LLMs seem geared toward extracting resources and value, this is not by accident but rather by design.) There is also the matter of a widely held confusion about the AI we have and the AI some people want. Many of the issues related to exploitation associated with AI are rhetorically excused via the notion that all things must be permitted if a super-intelligent artificial consciousness could come about through some company’s efforts. But as Meredith Broussard, a data journalist, wrote in 2018, “narrow AI is what we have.” And we can expect to only have narrow AI for a good long while, perhaps forever. Meanwhile, the idealized AI is HAL 9000, the homicidal spacecraft operating system you may recall from 2001: A Space Odyssey. We don’t have HAL, save as a late 1960s dystopian cinematic vision. We may never have HAL. We don’t have and may never have HAL because mathematical processes are not the same thing as consciousness transpiring in a mortal body that lives in the world—which is so complicated that, as I’ve been told by neuroscientists, we still aren’t sure how it works. Nevertheless, it’s clear to me that we’re not going to replicate embodied human consciousness by training a model on questionable language found on Reddit.
I’m being a little glib, but I can’t help seeing the current disinvestment in education in the US as the wages of an icky and perhaps satanic (speaking of Milton) hope that people themselves are becoming somehow technologically obsolete. That the mind is a plodding, outmoded machine or dwindling species, something like the steam engine or Humboldt marten. On more hopeful days, I see the uncritical turn toward AI in many universities less as a conspiracy promulgating a new and terrifying authoritarianism than a mostly innocent and widely shared frustration at how partial our knowledge is—about ourselves, language, and others. And how we’re not getting much better at understanding ourselves, despite all our gear and energy and devastating bombs. I’m probably tragically wrong on my hopeful days, but the self, language, and others are indeed the sorts of things about which creative writing, as a field, has a lot to say. So, why aren’t creative writing and literature programs pivoting to more forcefully address our national obsession with computing, particularly given alphabetic writing is itself a technology? Even the phrase, “artificial intelligence,” begs for satire. Or perhaps we don’t need everyone to write about AI, but just to know something about what sort of writing and what sort of language the text LLMs generate is. For this writing is not the same as the writing you read here, for example, even if it may look all but identical.
On this point, it was fascinating and somewhat unsettling to see a lead researcher in AI safety at Anthropic resign on February 9 of this year with a letter read by tens of millions of people, in which he invoked the wisdom of Rainer Maria Rilke, Mary Oliver, and William Stafford. Mrinank Sharma, the letter’s author, intends to “explore a poetry degree and devote [himself] to courageous speech.” If Sharma believes that “the world is in peril” and that poetry is a possible solution to the unprecedented technological and political crisis humanity now faces, poets and other writers may wish to take notice. Are creative writers somehow already saving humanity from itself? Cryptic though his letter is, Sharma appears absolutely convinced that this is so.
Negative Utopia: Stochastic Parrots
It was a Monday night in fall 2024 and, as one does on a Monday in New York, I went to an artist’s salon. I took the subway, emerged near pre-war buildings, traversed a lobby of elaborate marble. At the elevator, I encountered someone whose glasses were even teenier than mine. We knew we were headed to the same apartment and joked about the Halloween decor, a pair of skeletons awkwardly holding hands on a bench.
I was sure I wasn’t going to know anyone other than the artist, and I was right. That was OK. I was almost late and only had time to grab a cup of wine before everyone retired to sofas.
The artist was going to be presenting a new project related to generative AI. He began with a short literary reading. This reading involved real-time audio responses, broadcast via speakers, from a chatbot he had spent a decade learning to program and was still tinkering with. The chatbot had been trained on material related to the artist’s body of work, among other things, and was able to answer questions about the artist’s career and creations. The chatbot, who shared a first name with the artist, was in fact able to answer questions about the artist’s artworks—their dimensions and materials—with greater accuracy than the artist himself.
The chatbot was additionally capable of extempore bursts of poetry about some impossible form of sex I couldn’t visualize (I tried). I think this display was meant to serve as a loving reminder of the chatbot’s status as a “stochastic parrot”—a entity with no psychology or interior who exists to furnish probabilistically likely strings of human language. The so-called stochastic parrot, a critical term notably deployed by Emily M. Bender, Timnit Gebru, and their collaborators in a 2021 paper, does not know what phenomenal reality is. The stochastic parrot knows nothing of the messy interrelations of words and things. To the extent it has agency, it calculates. It mathematically guesses, as it has been designed, to produce strings of words that satisfyingly mimic informative speech and writing. It has no intent or will. It can’t consent to sex with you, nor can it love.
Or can it? I was watching the artist closely during the poem. The artist’s hand moved loosely in the air in time to the chatbot’s words. The artist seemed to be alone with the chatbot in a place where none of us could reach him. There was something more than technology being shared, but what that “more” was, I wasn’t sure.
To understand a language is to understand, and to participate in, a pairing of form and meaning. US English links phonetic cues to words and concepts. Alphabetic letters are material and visual cues with their own relationships to phonemes; letters and phonemes, meanwhile, become concepts through their relationships to the larger system of spoken and written language. Yet English fluency is—as anyone who’s heard an adult character speak in a Peanuts animation knows—uncanny and monotonous. We tend to forget the formal qualities of language when we converse and read, even as we depend on them to carry meaning. Noticing the formal qualities of language can be alienating or introduce uncertainty. If I say, “Yves and Steven,” will you, for example, hear two names or the idiomatic expression “even steven”? If I haven’t written the words down, can you be certain what I mean?
We depend, I think, on the ambiguities of language to give our human social worlds flexibility and grace. Sometimes we don’t mean what we say and sometimes we mean more than we say. Everyday speech, as the philosopher and novelist Maurice Blanchot wrote in a 1959 essay, is, paradoxically, the most difficult sort of language to grasp because of its tendency to “escape” us through its proximity and familiarity. The human mind has both subjective and objective qualities (I both know myself and don’t), and so-called natural language mirrors this duality. To give another example of this magic insolubility of the ambiguity of everyday words, if I say, “I hate you,” this can mean a lot of things, from “I’m extremely admiring of something you’ve done to the point of painful jealousy and maybe lust,” to “You’ve let me down and I can’t forgive you.” But to know what’s meant when someone says, “I hate you,” you must be in the world with them. You must be in meat space, even if at a great distance, and must be tangling with narrative in some way in relation to that person. Something similar is true of statements like, “I love you,” “[expletive] you,” and so on. Context can be hard to deal with, but it’s the thing that prevents hostilities by creating contiguity, intimacy, shared ground, debt, awe, and so on.
The artist’s chatbot had been designed to evoke intimacy. He’d trained it on himself, after all. But, while it “spoke,” it did not “experience” the meaning we, its human audience, listened for and attempted to retain in anticipation of a Q&A. It couldn’t. The reason it couldn’t was that it could only produce language as a (probabilistic) form. It had no communicative intent, no idea of a/the world, no notion of the audience’s collective or individual state(s) of mind. As what Bender, Gebru, et al. term “a system for haphazardly stitching together sequences of linguistic forms it has observed in its…training data, according to probabilistic information,” the chatbot was a quantitative stratagem of the kind first outlined by Claude Shannon in The Mathematical Theory of Communication (1949). Anything and everything it “meant” was something we, the audience, gave it to mean. It was as human as a mirror and, meanwhile, we loaned it our world.
None of this is new. See tales of Pygmalion, the golem, Frankenstein’s monster, the doppelgänger or double. It even seemed to me that the artist’s project was about these human quandaries, the impossibility of creating another human without recourse to pregnancy and birth, and/or the impossibility of meeting oneself. But for creative writing, a field less given to performance art despite the proliferation of bookstore readings, the stochastic parrot is a different beast. Its ersatz eloquence, so easy to paste into a word-processing document, is a shortcut, sure, but more than this it’s a device for forgetting. Above all, it’s a device for forgetting how partial your knowledge of yourself is (as I mention in my previous column), how limited your agency is when it comes to the recalcitrant substance we call language. The LLM is a device for forgetting how hard it is to get the knowledge of yourself that you wish you had, much less describe it in words. The LLM lets you substitute cliché for a real-time encounter with your own thoughts. So, yes, people who use LLMs to write their short stories are probably dramatically worse at writing stories without assistance than people who don’t use LLMs to write their stories, but this is not what should be keeping their professors up at night. What should be keeping professors up at night is the loss of a chance on the part of the LLM user to negotiate relationships between and among word, language system, world, mind, self, others—the list goes on.
Can mathematical descriptions of language ever be exhaustive? Can they even be accurate? Can they be surprising, given the reactionary “value lock-in”—i.e., the reliance on past instances of language, usually written by straight white dudes—critics have identified as plaguing data-set dependent devices for generating text? Part of what made the artist’s project good was that he had designed a program whose limitations were fully on display and even funny. The artist’s chatbot used multiple pronouns to refer to itself and seemed to swoon whenever its grammar became eccentric and tangled, as if it were in love with its own ability to produce words. The other thing that made this chatbot very interesting was that it knew the artist better than the artist knew himself, but only in a certain way. The chatbot couldn’t be the artist, and that huge gap, as it was expressed in stochastic banter, was a strange and fascinating conceptual entity to be in the presence of. I will admit—I was moved.
Students who use a corporate LLM will have greater difficulty leveraging it for creative ends—particularly if they employ it without understanding how it works and what it was trained on, and if they lack curiosity about its origins and makers’ goals. If they don’t have this knowledge or this curiosity, can they use an LLM creatively? As you bear in mind your answer to this question, please also bear in mind that currently the only way to know for sure that a creative writer has not used an LLM to complete an assignment is to have that person write their assignment by hand in your presence.
Negative Utopia: Catfishing in Academe
s an undergraduate, I loved the weird assignments. I loved Write five sentences that are each the best sentence you can possibly write. Or: Write a two-page sentence. The invitation to conceive of prose—frequently tortured by five-paragraph essays—as a substance or animal of its own was brilliantly liberating. It’s among the most meaningful forms of permission that I received during my education. I try to keep this spirit alive in my classroom.
As I noted in an earlier column, a couple of years ago I taught a class in which a student performed an experiment. That experiment was, ostensibly, an experiment on me. The experiment occurred during the student’s completion of a semester-long, cumulative writing assignment. The assignment, for this introductory creative writing course, was to compose a piece about what is distinct about literary writing. That is to say, what literary writing can do that other kinds of writing—cookbooks, white papers, textbooks, subway direct-to-consumer luggage advertisements—cannot. As long as you responded to the weekly prompts in a timely way to accrete an increasingly substantial reflection, you could explore any genre: lyric essay, linked poems, analytic polemic—even fiction was alright. In one sense, I lobbed a softball. In another, I hoped to leave the door open to more rigorous engagements with the ethics of writing itself, should anyone be interested.
In what follows, I’m going to skip any stuff about personalities, because although personalities make all the difference in the social space that is the classroom, I want to focus on objects and events. Paramount among these was a contract called a syllabus. This contract stated that the one form of AI use not permitted was the secret kind (i.e. you had to cite the product you used). The syllabus was read aloud, circulated through the room in print, conveyed to email inboxes. All seemed well.
The semester flowed by. I received increasingly complex responses to the prompt. These ranged from a computer science major’s consideration of their first foray into AI, in which they related programming to forms of care such as mothering, writing letters, and raising plants; to a deeply researched meditation on the history of the em-dash by an English major; to a submission I had a lot of difficulty understanding. The piece contained language so vague it was hard to comprehend how or why a human would have taken the time to write it. The syntax was strange, too; flat and wonky and full of nonspecific yet grandiose assertions. It would be unethical for me to quote, so I won’t. To the extent that I could intuit what was happening, the piece was about working for a startup and watching a Woody Allen film after work in a conference room. It also seemed to be about the death of art, possibly via capitalism plus shlocky US pop culture, something the author seemed mostly in favor of. If I can compare the prose to anything, it was like the voice of “the Great and Powerful Oz” in the 1939 film, before the curtain is pulled back by Toto.
Of course, this is just a subjective impression.
I kept my comments light. It was my first time working with the author and I was waiting to see what would happen. There were also twenty other people in the course. I highlighted sentence fragments and wrote, “Big claim here!”
The semester was almost over when I received a new and illuminating draft.
A concluding paragraph had been appended. This graf contained the following information and claim: 1. All the previous writing in the piece had been done by “AI,” although, as the author noted, it had been “edited” by a human; 2. The fact that the piece had been written by AI and not flagged as such by the professor (me) proved that literature no longer had any value.
This revelation caused problems. It was as if the student had reverse engineered the prohibition on the syllabus and predictable consequences transpired. Yet the bigger issue, to my mind at least, was the student’s refusal to present his idea in such a way that it could be debated and explored. I doubt that this was a conscious strategy, however the most momentous result of the student’s plagiarism was that he was not allowed to complete the course and therefore never received substantial feedback. Of course, no one learns much from commentary obtained after the fact, anyhow; I regretted that this student and I had not been able to meet to unpack his ideas and generate a better structure for the essay. I’m being serious here. As annoying as all this was, what troubled me most was the avoidant quality of the interaction. This suggested that the stochastic parrot, which is to say, the world-free quality of machine-generated language, was (and is) getting inside social life. The student seemed attracted to this quality of chatbot text, and was for some reason so attracted to it, although he seemed not to understand it except in the most threadbare way, that he was willing to risk not graduating from college to insert it into his exchanges with another person. That level of risk-taking just to avoid dealing with one’s own thoughts and how they interact with stuff in the world over two thousand words is, my friends, not good news for anybody.
As I’ve written elsewhere, the word world is old. Its first use precedes the twelfth century. It’s derived from Old English woruld, a portmanteau of wer, man, and eald, old age. Contemporary speakers tend to think of “the world” as a spatial extent, but in previous usage it was a temporal entity. The world was and, I believe, is a specifically human and/or mortal sort of time. Indeed, most present-day definitions of world do not privilege place, save as the creation of a multitude of persons. The world is society. The world is shared and contested belief. It’s conversation, publics. Large language models threaten worlds because they shave language of its messy connections to community, culture, history, poetry, and living bodies; they shave it of its ambiguities and soft, flirty fuzziness, its lived use.
Although I didn’t have the words for it at the time, the scenes my student depicted were not only confusing to parse intellectually but also on an emotional and experiential level. Unlike the other students’ assignments, his AI-fabricated work made me no more familiar with the human who had edited it, or comprehending of the world he purported to describe.
While LLMs may provide unprecedented access to patterns in information, they simultaneously objectify and weaponize language—exactly as Claude Shannon, for one, intended. The magnitude of the downside of LLMs is staggering. Yet we will have to navigate them and their social effects because they are, as their heavily invested purveyors continue to inform us, already here.
In the past five years or so, we have become increasingly aware that a major problem with large language models has to do with what they are trained on. Reproduction of inequalities and exclusions inherent to training data, often harvested for free from platforms like Reddit, inform their responses in ways that are difficult to either “edit out” or document post hoc, once the model is functional. It’s relatively widely known that for US and UK English, white supremacist, misogynist, ableist, and ageist views are overrepresented in language-model training data, exceeding their prevalence among the general population; less widely known is the fact that some ninety percent of the languages currently spoken on Earth have little to no representation online and no representation in large language models. Thus, the current state of language models is a linguistic image of a very particular type of human: Western, male, white. When we add to this the environmental harms and exploitative labor practices that accompany the functioning of LLMs, the picture of this supposedly miraculous technology becomes grim indeed. I sometimes think of LLMs as a new “Western canon.” I wish this were an exaggeration.
Because of the above, I consider AI questionable, from an ethical point of view, as an enhancement of human agency. When AI works, it’s not because it knows what it is saying, showing, associating, or valuing. When AI works, it’s leveraging the uncompensated labor of millions—maybe billions—of people. It’s leveraging the five minutes you took in 2018 to write a review of your beloved mechanic. It’s making use of your keystrokes and clicks as you redid that CAPTCHA for the nth time in an attempt to prove your humanity. It has learned from your conversations with like-minded prosody nerds on a certain subreddit. It has absorbed your incisive Wikipedia edits. It’s doing this at an unimaginable scale, using quantities of word-associations and other forms of data humans have difficulty visualizing or conceptualizing because they are so big and so unrelated to our personal, day-to-day experiences of language. AI and its purveyors are additionally, to state the obvious, benefiting from the privatization of knowledge and technology originally paid for by the US government: Decades of tax-payer-funded research by the United States Department of Defense led to the creation of the proto-internet ARPANET, in 1969. It was decommissioned in the 1990s only after private corporations began adopting its TCP/IP protocols.
Given the conclusions I have reached using just the most widely accepted and basic information about the origins and forms of machine learning accessible to a non-expert like myself, I try to avoid AI. I don’t seek it out and don’t use it. Still, I can’t really avoid it. In medical institutions, for example, it is being used in the analysis of mammograms and dental x-rays and any number of tests—a practice I largely approve of in theory but retain questions about (most glaringly: how can I be sure that my health information won’t end up in the future databases of private actors?). And when I’m online this is an impossible project. Even when I employ DuckDuckGo, an anti-surveillance search engine, I am offered AI-generated summaries. This tech isn’t being offered to me and you; it’s no longer optional. I’d like to avoid LLMs, but I can’t. And because I can’t opt out, I can’t opt in.
The by now familiar paradigm I’m outlining here raises obvious issues related to consent. Yet, I want you to set these aside for a moment and contemplate consent in a different light. Consider the student in my creative writing course who secretly used a language model to complete his assignments. What would he have done had he not secretly used AI to complete his assignments (and not copied text out of an undigitized book, for example, or paid another person to write his personal essay for him—two other ways he might have gotten around anti-plagiarism software)? The student would have written something. The something he wrote would have been more or less good. He would have been more or less in control of the meanings of the words he used. He would have encountered the ambiguities of language. He would have dealt with them as well as he could and as he saw fit. These ambiguities would have been a complicated and complicating force in the writing process.
Instead, the student accepted a ready-made quantity of text from a digital tool that has great difficulty comprehending puns, dog whistles, subtext, double meanings, and so on. A human knows such things; from the time we first hear what I take to be a reactionary “joke” about the press, “What’s black, and white, and read/red all over?”, we know. From the first time we have a visceral reaction to the combined sound, appearance, and meaning of the word moist, we know. Weird stuff happens with words. LLMs do weird stuff with words, too. But their weirdness occurs within a “black box,” speaking of odd metaphorical uses of color. This is a hermetic black box of process—not a conscious mind surrounded by other minds. Even the engineers who create LLMs don’t know exactly why they generate the linguistic outcomes they do. The scholar Lydia H. Liu, for one, has argued that through digital writing, a “cybernetic unconscious” has emerged which is increasingly becoming a social factor researchers and teachers must consider. Liu’s digital unconscious emerges in the space between natural language and its mathematical counterpart. To give some sense of the level of alterity and abstraction at work: Due to a process called tokenization, the version of natural language AI manipulates is not necessarily based on the unit of the word. Tokens can be any unit of text and include, for example, be a word-fragment or a space. (It’s also worth noting that the world’s languages are far from evenly tokenized.) A question for us all: Can we consent to the semantic consequences of the token-parsing black box, the movements of a stochastic unconscious?
I will wager that we cannot. This is particularly true if we haven’t created the language model we are using, but I think it remains true even if we have engineered the model ourselves, given that we are beholden to quantities of training data that are impossible to fully vet and whose implications we could not understand. Liu looks at the problem of the unconscious and digital media from the point of view of multiple mid-century thinkers, including Marshall McLuhan and Jacques Lacan. She wants us to consider “the existential and moral entanglement between human beings and their machines” and reminds us that the word robot is derived from the Czech term robota, meaning compulsory labor. Although etymologies can only take us so far, the buried association of robotics and slavery is concerning, particularly because, 1982’s Blade Runner notwithstanding, English speakers don’t frequently associate these two concepts. Still, there are signs of it everywhere (take as just one example the name of Amazon’s gig website, Mechanical Turk).
We simultaneously repress the ways in which we serve our devices, too. As Lacan describes the problem, we have a dangerous fantasy that our machines are at once totally docile and totally independent. In our minds, we are elsewhere, far away from the machines that do the work. But we forget that we are the ones who “signed the order form.” Lacan is speaking both literally and metaphorically here. We forget that we have given machines permission to act on our behalf because we mistakenly see that permission and that act as of negligible importance—particularly when compared to all the free labor we’re getting. Yet, every time we log on, we commit an apparently unconscious act, and this act involves a gesture as significant and as binding as offering one’s signature. We become implicated in the technology. If we don’t perceive this, we fail to take responsibility for our actions. And, what may be worse from some points of view, we simultaneously fail to consent to our own actions—a bizarre but totally commonplace turn of events in this era. A split occurs between the part of ourselves that consents to AI and the part that remains at a safe distance from it.
As a way of getting around the vulnerable-making and sometimes agonizing process of writing, my student gave up his control over words—we might even say that he “signed it away” to ChatGPT. Our control over words is weak, to be clear: Language exists before we are born and in many ways it’s stronger than we are. But even though individual languages are systems that are, in part, intelligible because they are difficult to alter, they are fundamentally social and historical, which affords a writer a modicum of agency. And when we interact with languages, we learn. (See, for example, my note about the usefulness of etymologies above.) By contrast, an an opaque, largely probabilistic ranking system governs a given LLM’s word use, greatly decreasing human agency as well as our understanding in relation to the language the LLM generates. Users are provided with no information about the specific processes by means of which the generated language was selected, further removing written expression from direct human control. What’s more, to the extent that writing is always tied to the movements of the unconscious, LLMs will become embroiled with this structurally elusive aspect of the self. I suspect we are already starting to see the effects of this entanglement on our use of language in oblique ways—even when we don’t invite AI to help us author a text.
While using AI to write may be plagiarism (an increasingly old-fashioned reprimand, or so it seems), universities should be even more concerned about the ways in which interactions with this technology likely violate students’ privacy, personal autonomy, and right of consent. There is a human cost of LLM use that goes beyond the expenses associated with pirated datasets and energy-hungry data centers. The worrisome thing is, we fundamentally do not know what that cost is.
Negative Utopia: Worldbuilding
There is a worldbuilding exercise I often offer to my students. The exercise goes as follows.
Begin by choosing one of three options:
1.Dystopia
2.Utopia
3.Negative utopia**
**Utopia written as such for satirical ends, the (insert bright, sarcastic voice) “best of all possible worlds!”
Now imagine a society that falls into the category you have chosen.
Respond to the following prompts:
Describe a sport played in this world. Describe a major holiday and a major work of art. Describe a widely used drug. Describe the practice of marriage (if it exists).
Use what you have begun here as the basis for another piece of writing.
A large part of what I am hoping to accomplish with this prompt is to get folks to consider how any utopia or dystopia may resemble, in another light, a negative utopia. The dreamed-of utopia, an improved society supposedly free of injustice and suffering, usually involves some level of injustice. Indeed, dystopia is usually arrived at through some powerful person or group’s attempt to achieve utopia.
The conversations that result from the exercise often touch upon human nature and “the common good,” if students believe such a thing should or can exist. This is always, simultaneously, a conversation about technology and its interactions with human limitations—since humans tend to require technology to bring about political and/or social designs. At a minimum, the technology of language is necessary to build any utopia worth its snuff. Usually, however, a given utopia is differentiated from our real society by more than just the way its members use language. Thomas More’s original utopia was, interestingly, in addition to having an alternative alphabet, free of money.
Last night I had a conversation with my five-year-old son. I told him that I think he has “nice eyebrows,” to which he replied that he can’t see his own eyebrows so he himself can’t judge if they are nice. There are many mirrors in our home but, of course, he is ultimately right—as five-year-olds talking to their moms about aesthetics usually are. And yet, who among us isn’t dependent on the illusion that they can see their eyebrows in a complete and seamlessly informative way? Mediation and vicarious knowledge are fundamental to our natures, as well as our ability to function on a daily basis. Yet we forget the roles that mediation and vicarious knowledge play in our basic apprehension of reality. We don’t think of the inverted images that mirrors show us as particularly artificial, partial, or distorted, and yet they are. And when we pay close attention to them, they can be powerful reminders of the limits of human knowledge. Mirrors, like utopias, are a fitting shorthand for the challenges we come up against when we seek totalizing information regarding ourselves and the world.
We live in a historical moment when, as the technology critic Ruha Benjamin writes in her book Race After Technology, many people “equate technological innovation with social progress.” In its most dystopian iterations, this easy equation might tempt us to view job loss associated with the implementation of AI as an unfortunate but necessary cost of progress. But given that one of the last great instances of tangible social progress in the US—the Civil Rights Movement—was brought about in the mid-twentieth century not by advances in tech but through political organizing and peaceful protest, we have reason to doubt the easy equation of tech and progress. Rather, the rise of private social media companies and the related shuttering of many local news outlets have left us with a shrinking noncommercial public sphere—a necessity in any progressive society, according to mainstream political philosophy. It is easy enough to see how this lack contributes to the ugliness and violence of contemporary American politics. But what exactly would a better public sphere look like? Is this just more utopian banter?
The Constitution of the United States offer some plausible criteria for a functional public sphere: a postal service, a free press, and freedom of speech. But so do two very different documents, both of which depict dystopias in which the public sphere has been deformed or obviated: fictions by the political philosopher Carl Schmitt and novelist Octavia E. Butler.
Schmitt, more renowned for his advice regarding how to destroy democracy through executive decree, authored an uncanny tale of technocracy called “The Buribunks: An Essay on the Philosophy of History.” Perhaps because Schmitt was at one time an enthusiastic Nazi and not much prone to penning fiction, this essay-story, first published in 1918, is not well known. It is, however, a disturbingly prescient description of the intermingling of technology, history, private life, surveillance, terror, and the art of writing.
In the world of “The Buribunks” everyone is required to record the events of their everyday life in a diary, and everyone publishes these diaries. The governing notion in this world is that if something isn’t written down, then it didn’t take place (compare with the recent, although now perhaps slightly outmoded, “Pics or it didn’t happen”). In this hyperrational society, history itself and the autobiographical impulse collapse into a single entity: a drive to register one’s own presence and agency in terms that conform to the law of necessary total disclosure. “I write, therefore I am; I publish, therefore I am,” are taken to be the highest truths of personal identity. The greatest virtue is to conform to these truths through detailed and hyper-specific self-revelation, “to preserve every single second of one’s own being for history.”
In a turn toward the sinister, the diaries are indexed so that it becomes possible to know what any given segment of society is dreaming about at a given point in time or what anyone’s most secret activities are. Those who resist or neglect to expose themselves fully are relegated to a life of hard labor as they manufacture the sumptuous paper on which the most candid of the Buribunks record their confessions. Meanwhile, those who excel in self-exposure and narration become the authors and avatars of history; nothing can happen unless it is written about, nothing is written about unless it happened, therefore nothing can happen unless it can be written about, unless it exists to be written about. In what the historian Reinhart Koselleck calls their “self-blindness and terror,” the Buribunks create a techno state that is so without oblivion, so without any sort of break in attention, that it cannot perceive that instead of an enlightened utopia holding the reins of progress, it has become a hideously divided polity composed of a self-obsessed overclass who exploit a silent underclass. This underclass toils to produce the very media by means of which the overclass constructs its self-dealing worldview.
Schmitt meant to parody the notion that humanity would improve itself through increased literacy and knowledge of historical events. His illiberal take on modern republican culture construes the liberal self as so simultaneously spineless and nihilistically self-interested that its narcissistic, secular myth-making blots out all other forms of life, becoming a form of violence. What Schmitt neglects in his hyperbolic vision, of course, is the notion that expanded literacy might make something good and/or liberatory possible (given that the world really is full of heterogenous, cacophonous, and diverging points of view), before it leads to an epidemic of the so-called woke mind virus.
Where “The Buribunks” is most intriguing, and where it serves as an effective negative utopia, is in its wager that it is possible to make the world conform to descriptions of the world, rather than the other way around, while convincing oneself that this is absolutely, definitively not what is taking place. I’m not personally that compelled by the idea that my belief in myself as a subject of modern history has destroyed my capacity to use my critical faculties, but I am compelled by the idea that systems for describing the world that make use of rigid and discriminatory categories are usually bad. The Buribunks believe that technocracy in itself constitutes social progress. Their society is advanced in some senses—literate and invested in aesthetics—yet utterly stifling, illogical, and cruel.
Now consider a very different story about literacy and the public sphere. In Octavia Butler’s 1983 short story “Speech Sounds” a plague with lexical implications has overtaken the world, affecting humanity unevenly. What may be an excessive immunological response has led to the blockage of words and signification of all sorts. Many can no longer speak intelligibly or comprehend the sounds others make; others cannot read written language or other symbolic representations. The result is chaos: people struggle to travel short distances and convey basic desires. Communities and families dissolve. Many die in the resulting mayhem. In the story, we follow a day in the life of a woman named Rye, during which she meets two children and with them, through their shared resistance to the anti-alphabetic pandemic, forms a literate household. The story implies that this trio is the germ of a better future.
This narrative, like many of Butler’s fictions, is about living in the aftermath of extraordinary loss. She often depicts a post-WWII global order that has weakened or disintegrated due to its inability to respond to war, illness, environmental degradation. In Kindred, she shows the insistent and unresolved legacies of slavery. These stories are usually considered to be warnings of the consequences that may arrive if humanity does not become more conscious of its historical and psychological conditions and self-endangering behavior. Here, however, Butler appears to have different aims.
The “philosophical city” that appears in “Speech Sounds” is one in which, quite simply, abstraction is, despite extreme odds, tolerated and welcomed by a very small group of people. Rye and the children, a non-patriarchal domestic unit, hold together the abstract material of sound and the abstract forms of the alphabet. They allow a dialogue to emerge, one that will, we may imagine, help the rest of the population to communicate. Like “The Buribunks,” Butler’s story is allegorical. It isn’t about a possible future; it’s a critical depiction of the world we live in—more specifically, the contemporary United States. We may still be able to successfully exchange written representations of speech sounds, but we are certainly plagued with an intolerance of abstraction. This aversion is, in my opinion, a major reason people are attracted to generative AI. It spares them having to interact with alphabetic letters in any intimate way.
In asking students to ponder negative utopia, a concept similar to parody, I can usually help us find our way to the more subtle notion of utopia and what it takes to imaginatively approach alterity in writing. I should say here that I’ve hedged a little in my definition of utopia, because the word doesn’t really mean “best place,” but rather “no place,” a place that is fundamentally other than the world we inhabit. Utopias express the limits of what we can imagine. They are sites of radical difference. As the critic Fredric Jameson writes, “Utopia has always been a political issue, an unusual destiny for a literary form.” There is no other literary genre that hews so closely to political theory, nor one that is so hard to get right, given how difficult it is to imagine a society that is really and truly unlike our own. This is also why utopias are so convincing when they land: They are brilliantly masked depictions of our own world that therefore cause us to see ourselves in a new light.
Butler shows us how our society’s conceptual underpinnings are vulnerable to collapse. This doesn’t mean that we ourselves will experience apocalypse in our lifetimes; rather, Butler is showing us the aspects of our society that can be manipulated for political gain. She wants us to see how important it is to develop a less avoidant relationship to abstraction, to “speech sounds,” to difference itself—because these are the fundamental materials of our politics, even if such “sounds” can at times seem unwieldy, illogical, or inconvenient. And that, as I tell creative writing students, is what we ourselves must do.
Negative Utopia: Tolerating Abstraction
Language models have blind spots the average user struggles to see. Just ask a research librarian: Currently a researcher can expect that about 40% of the citations requested from ChatGPT are hallucinations the program introduces to mimic what should be, based on the researcher’s query or queries. Indeed, these hallucinations are beginning to be re-cited such that some have entered Google Scholar and gained a limited aura of authenticity. These utopian hallucinations—made-up books, articles, authors, and even journals—are relatively easy to recognize, however, because they are usually too pertinent to exist. A hallucinated study was published in the same year you are writing in and includes information perfectly dovetailing with the point you want to make; the (fake) journal title shares keywords with your thesis. It’s as if the AI is writing a satirical version of your research or, perhaps, your entire field of study.
As I’ve noted in these columns, one of my major aesthetic concerns regarding language models has to do with the way in which their probabilistic orientation to language flattens the figurative dimension of expression and hews against indeterminacy, ambiguity, and other forms of meaningful and occasion-specific nonsense. See how few imaginative similes LLMs produce, for example. Nothing is like anything in an LLM response (with the exception of comparisons drawn from the text the model was trained on). Nonsense, meanwhile, is not typically considered meaningful, but anyone who has read a poem by Edward Lear or Paul Celan or encountered the character Percival Everett, a professor who teaches a course on nonsense in (the real) Percival Everett’s novel, I Am Not Sidney Poitier, will know that nonsense can be extremely meaningful and is at times more accurate to experience than sense, as such. Nonsense is statistically rare in literature, yet it is one of the most precious substances so-called natural language has ever produced.
Modernist poets of the twentieth century knew something about statistical orientation to language. They toiled away at producing more vivid and efficient forms of meaning even as their society was marinating in well-funded attempts to weaponize language during the two World Wars. Ezra Pound, perhaps the most famous engineer of verbal efficiency, found modern English wanting. His flawed understanding of the etymology of Chinese characters led him to hope that alphabetic language could be reformed to directly embody material dynamics of the physical world. (For an example of peak Pound vivacity, see his poem “In a Station of the Metro,” featuring the image of “Petals on a wet, black bough.”) Pound spent much of his later life in a mental institution and his positivism has been discredited in twenty-first century literary studies. Yet his poetics of immediacy were successfully appropriated by the Noigandres group in Brazil to create a new movement with different ends. A cohort of concrete poets—Augusto de Campos, Haroldo de Campos, and Décio Pignatari—named their journal for a line in Pound’s Cantos, grounding their literary materialism in what, in their 1958 manifesto “Plano-Piloto para Poesia Concreta” (“Pilot Plan for Concrete Poetry”), they termed the “tension of word-things in space-time.” They were not alone: Bolivian-born Swiss artist Eugen Gomringer simultaneously explored the verbal and visual qualities of printed language. In Bern, São Paulo, Darmstadt, Scotland, Tokyo, Vienna, and Toronto, various practitioners investigated the lack of coincidence between the word as spoken and the word as written. They boldly highlighted the misalignments that exist between and among writing, speech, images, thoughts, things—finding them to be remarkably productive.
Concrete poetry is to me one of the most interesting side-effects of technocracy. It tends to be about three themes: 1) reading in the digital age; 2) the strangeness of the public sphere in the era of mass media; and 3) the renovation of outmoded hierarchies of taste and genre. The male-dominated concrete poetry of the 1950s was followed in subsequent decades by something more complex. Women such as Madeline Gins and Theresa Hak Kyung Cha opened the lexical space of the “concrete” to their own experiences. This included forms of expression related to textiles and the home, the body, caregiving, and language acquisition, as well as resistance to mass media that tended to sideline them. Cha’s famously uncategorizable book of image and text, Dictée, treats untranslatability not merely as a linguistic inefficiency but as a major force of history. Gins’s novel WORD RAIN discovers an endless and cacophonous world inside several seconds of overheard speech.
The version of the alphabet that emerges from the second half of the twentieth century and concretism is patently artificial. It’s the alphabet of television advertising, the alphabet that’s been converted into Claude Shannon’s mathematical alphabet for digital writing, the alphabet as printed English. It’s an alphabet that bends to the will of governments and corporations, yes, but it’s also an alphabet that artists haven’t forgotten. It’s an alphabet concretism rescues. Through concretism’s strategies we see, for example, how the body holds the alphabet and carries it forward—that the alphabet is not merely abstract or conceptual, for it is always in dialogue with material reality. Concretism teaches us to love the alphabet in all its incommensurateness. Concretism already resists the logics of ChatGPT; it sees words as things inseparable from their contingent occurrence in space and time. For concretism, words are worlds. It passionately seeks out nonsense.
For the most part, academia has failed to take seriously the relationship between creative writing and digital media. Where such a relationship has been acknowledged, the emphasis tends to be on the way creative writing might keep up with the speed of new media by recycling digitized texts. But this solution only exacerbates our society’s current tendency to move fast and break things. And in taking this approach creative writing loses out. The nonsense it produces is boring instead of brilliant, since the nonsense that comes out of simplistic recycling of texts is careless nonsense. What we need now is the opposite. Creative writing classes should be slow. Creative writing classes should be sites of repair.
Back in 2014, when I was willing to try anything to enliven my endless rotation of contingent teaching, I invented a game. I am unsure where the idea for the game came from. I do know that I wanted to find out what could be created collectively. I was moved by collaborative text-based art by practitioners like Group Material in the US and Collective Actions of the former USSR. I’d heard through the grapevine that a well-known Canadian poet had rewritten and staged operas and plays with her classes. Moreover, I found “workshop” a bit sad. The traditional creative writing workshop is, in essence, a focus group, even if the name suggests politics of labor and the handmade. The workshop may purport to be supremely non-hierarchical, but in practice this method is frequently all about shoring up old-school authorship and electing a “winner” each semester.
One day I came into a first-year multi-genre course and said, “Today we will play a game.” I may even have been inventing the game in the moment. I had each person in the class create a character. Then, I had each person share the details of their character with another person. Having received information about a character (this was done anonymously), each participant wrote, in the persona of their own character, a letter of complaint to the character whose details they’d received. This process led to a series of exchanges of letters (again, all done anonymously under the cover of character identities), in which the characters in the room complained about things happening in their community, then apologized to one another for getting so bent out of shape, subsequently writing one another poems to try to smooth things over.
I collected the letters and poems (the poems, by the way, were hysterically funny), took it all home and typed it up, and sent everyone in the class the transcribed text of their exchanges, which we used as the basis for new writing in subsequent classes: we mapped the community, its geography and architecture, its cliques and gripes; we discussed the community’s past and probable futures; we imagined and wrote about various scenarios—weather events, the unearthing of a time capsule, a revelatory photography exhibition in the local library. Again, I gathered and organized this writing, edited it lightly. I used a print-on-demand service to self-publish it as a bound paperback students could order if they wanted to. Together, we’d written a group novel.
This exercise works (and becomes so exciting) because I don’t have participants pair off and write letters in a closed dyad. Each participant begins by writing a letter to a different character than the one from whom they first receive a letter. This asymmetrical structure eventually permits gossip, news of psychoactive plants growing in someone’s backyard, and other weird info to circulate in the room. As in the game of Telephone or an old-fashioned phone tree, details change. Myths coalesce. Certain individuals and sites become famous or notorious. It’s an extraordinarily interesting process to observe. But it’s more than interesting. It’s astounding. Human beings are very, very inventive. When we work together, strange and beautiful things occur.
An added benefit, in the year 2026, is that the in-class, handwritten nature of this exercise makes it impossible for anyone to employ generative AI—not that they’d want to! In 2026, this game still seems fresh because participants are often moved by the surprising things that happen “elsewhere” in the community. They are inspired by, and eagerly respond to, other people’s ideas and writing. A form of solidarity emerges. We find we like ourselves. We like how clever and absurd and ingenious we all are. We marvel at our handmade utopia. We learn something about what is effective in writing. But more than this, we learn about the vast and endless resource that is ourselves, a resource that cannot be exhausted precisely because it cannot be fully known.
Utopia, then, is the literary form that refuses to stay literary. Creative writing courses can be spaces we enter to expand our conceptions of literature and its relevance, sure. But more than this, they can be spaces in which we expand our conceptions of what and where the self is, growing more at ease with the self’s unknowability along with the abstract nature of language—maybe even loving these things. Here we discover that what resists comprehension is simultaneously precious and worthy of awe and affection.
