How LLMs Set My Fiction Free

The cautiousness is a little more complicated, but it added to the staleness. Trade publishing seemed to have accelerated its trend of avoiding what it deemed uncommercial work, particularly from writers whose sales records, like mine, were poor. And the broader cultural moment—with its heightened sensitivities, its quickness to take offense or to assign it—had made risk feel costlier than ever. I became skittish and pigeonhearted: making safe bets, avoiding the healthy recklessness that can invigorate a sentence, or a project. For years, I’d tried to make work that was different—unorthodox, sometimes strange—but I’d also enjoyed some of the rewards of “legitimacy.” I could expect to be published and to receive a certain amount of recognition. Now, in adjusting my work to meet the terms of that legitimacy, I’d compromised it.

By 2022—the era of Failure B—I’d started fooling around with an early online text- generation tool called InferKit. It was a simple prompt-and-completion engine: the user input text, and the software inferentially continued it. I don’t have a complete record of these early experiments, but I did preserve the moment I realized that a machine could yield exciting output. Amid a five-page narrative that I saved is this InferKit-generated fragment:

The passage built off prompts I’d given the software, but the feel, the repetitiveness, the rush of pointlessly overprecise elaboration, the spectacle of spiraling self-consciousness presented in a clinical tone all emerged from the machine. It produced a species of prose I have long admired in writers ranging from Dostoyevsky to Thomas Bernhard—looping, incantatory, self-refining, and self-undermining, with deadpan comic effect.

Over the following months, I played more and more with InferKit. My sessions revealed other features of its output: unpredictability; a tendency toward non sequitur; startling, even unintentionally comic juxtapositions; a complete disregard for the relevant, the polite, and the reassuring. Unlike the hours I’d spent struggling to write Failure B, I enjoyed these sessions. I looked forward to them.

About a year after I started experimenting with InferKit, ChatGPT was released to the public. Along with millions of others, I was enchanted by the new technology. I played hooky from Failure B to “write” “novels” with it, including an episodic Western about a frontier town whose economy and culture are based on the manufacture and sale of explosive lingerie; a mystery concerning a Brooklyn detective uncovering a conspiracy involving robotic surveillance pigeons; and an alternative history in which the Beatles, led by a psychopathic Ringo, use their power and influence to extort money and creative property from rivals. None of these early attempts were good, exactly; part of their charm was in ChatGPT’s earnest application of beginning writer’s mistakes—mixed metaphors, clichés, hackneyed characterizations and plot developments—to the somewhat deranged narratives I was urging forth from it. As I had with InferKit, I often found myself laughing out loud at my desk.

If this was pleasure, it was sometimes a guilty one. As generative AI slop began to overrun the corporate inbox, the classroom, and, especially, the internet, the prevailing opinion about it among creative people was becoming clear. A piracy machine that spit out thinly disguised works of plagiarism, buggy code, and inedible recipes at massive environmental cost, AI was going to degrade art, hijack audiences, impair learning, wreck the planet, and possibly lead to an extinction-level event. I didn’t think these were frivolous concerns, and I still don’t. Nothing has convinced me that they’re false, least of all the reassurances from the latest batch of tech billionaires.

Yet whatever potential hazards and unintended consequences flow from AI’s mass adoption, my own experience using it contradicted much of what I was being told. One fear is that these systems will strip prose of style, idiosyncrasy, and that elusive element “voice,” but I found that they could do the opposite, introducing the unexpected, the “wrong,” the uncanny disjuncture between the realistic and the oneiric.

Most important, I was having fun writing fiction for the first time in years. As a phenomenon, “fun” is hard to quantify, but I don’t hesitate to attribute to it a sort of essential value, and it may be the most reliable motive I know to keep one writing. Art, I’ve always believed, retains traces of both the joyous spontaneity that sparks its conception and the exhilaration that comes with its realization. After a long, frustrating slog that was dismally free of spontaneity and exhilaration, I suddenly had both in abundance. I began to see possibilities in the odd, askew narratives I was producing, and a chance to transform them into work I took seriously. By late 2023, I’d shifted focus to what I called my “machine stories.” I turned my back on poor, stunted Failure B.

AS A BEGINNING WRITER in the late 1980s, I’d steeped myself in modern, especially postwar, fiction. My early work paid homage to its innovators, foregrounding formal scaffolding, nonlinear narrative, and obtrusive devices of all kinds. Take Sound on Sound, my first novel, published in 1995, which emulates the layered mix of a sound recording, with each section representing a different “track.”

Gradually, though, I distanced myself from overt experimentation. I wanted to be published, I wanted readers, and for a long time I was satisfied with the work, which was hardly conservative and belonged to a very recognizable strain of male American literary fiction from the late twentieth and early twenty-first centuries, at home in both tiny journals and mass-circulation glossy magazines. I was perfectly willing to defy convention when I felt it served the prose, but I didn’t pursue innovation in any self-conscious way.

With AI, it felt like I was once again experimenting to arrive at the methods that would make each story happen. Working with large language models (LLMs) felt a little like using the devices engineered by the OULIPO—the French experimentalists who imposed generative constraints on their work, such as writing a whole novel without the letter e—or the recombinatory cut-up techniques of William S. Burroughs. But it operated at a different scale; these new tools made the Oulipians’ more otiosely schematic conceits—say, Raymond Queneau’s notional hundred thousand billion poems—seem even more like proofs of concept than they were intended to be.

At their most basic, my methods will seem familiar to any casual user of generative AI: I asked, the bot answered. Sometimes I turned to GPT for help making my work more accurate:

Other times, I used it to help make my work more inventive:

Of course, by now this sort of query hardly counts as experimentation; in 2026, everyone from high school students to office workers are familiar with GPT’s usefulness as an expedient, and I admit to using it frequently to supply details, names, and lexicographic and usage guidance, among other ad hoc requests.

Where AI began to really contribute materially to my working process was through iteration: input, generation, selection, revision—and then a lot of repetition, until these steps yielded a phrase, an image, an idea that opened a new vein to explore. Maybe the most elaborate example of this is “Almost Infinite,” a story that began when I mused to GPT-4 that one measure of a writer’s competence is “the ability to usher a group of people into a room and have them all sit down at a table without missing a beat,” and then proposed a twist:

GPT-4 instantly generated a draft of about four hundred words, establishing the setting (an overheated dining room), the cast (members of a family reuniting for a holiday), and various internecine conflicts and complications. It sounded a bit like a workshop student doing Franzen. Amused, I had it create variations in style and register, then instructed it to progressively pare them down. Once I had about twenty different versions, I selected eight, which I revised line by line, teasing out and amplifying motifs and recurring phrases to give the individual pieces resonance beyond their shared premise. The finished story moves from one version to the next, progressing from most elaborate to simplest: a critical text copiously annotated with mock academic citations, a baroquely figurative and simile-laden vignette, “contemporary literary fiction,” a Hemingwayesque pastiche, unattributed dialogue, a theater director’s notes blocking the scene, anthropological field notes, Beckettian reduction.

The goal of this kind of iteration is not to perfect a draft but to create one that has an unstable quality, labile and ambiguous. Most often I work in the space between accident and intention, using AI to produce the “wrong” thing and then shaping that wrongness into coherence. I typically work with the current version of ChatGPT in combination with earlier, more primitive models because of the off-kilter results they produce. For one story, I fed GPT-4 incomplete dialogue and asked it to complete the exchanges, but to do so emulating GPT-2: I wanted a muddled effect that would give the impression of characters talking at, or around, one another. I selected fragments of output and stitched them together, adjusting to give it cadence and flow.

Glitches are good for style, but also for substance. While working on a story called “Housewife Studies,” for instance, about a Stepfordian company town where women are subjugated to the wills of their husbands, I again turned to GPT-2 to generate dialogue among a group of these wives. At one point, it spat out:

Lance? I hadn’t invented any Lance. Where had he come from? These kinds of hallucinations are held up as the classic example of AI failure, but they’re exactly the sort of accident that I look for to enliven a story and send it careening in an unanticipated direction. Intrigued, I entered further prompts to elicit more about Lance. Gradually, he began to emerge:

In the finished version, I developed Lance into a figure whose absence informs the story and affects the lives of at least four characters.

AI has become a zillion-dollar industry by promising shortcuts and serving as a time “amplifier.” An LLM can realize a concept almost instantly and let me test a premise without wasting hours, or days. But at its most involved, and most engaging, my process with AI is painstaking. “Housewife Studies” is about twenty pages; the transcript of my related exchanges with GPT-2 runs to more than eighty. And even these pages don’t document the story’s composition, which entailed working manually on the structure, pacing, dialogue, and individual sentences until I was satisfied. Describing my process with AI as cutting corners would suggest that I’ve already settled on a design and have merely outsourced the execution to a machine. But what the transcripts of my conversations with GPT preserve, along with multiple errors and false starts, are the small moments of potential that provide the impetus to turn a story into something strange and unexpected, even to me.

WORKING ON MY “machine stories” immediately restored—and heightened—my awareness of fiction as artifice. A successful sentence in these pieces was one that snapped and startled, not one that skillfully conjured reality. By defamiliarizing language and tone, LLMs made me think carefully about their output and, in turn, about my own prose, helping me to approach narrative anew. This was partly a function of the raw material I was working with, which was drawn not from “real life”—conversations, headlines, history, autobiography, knowledge, experience—but from machine-generated text. My prompts yielded a babble of ad copy, fan fiction, Reddit commentary, reportage, legal text, and other matter, all surfacing out of the stew of data the LLM had been trained on, occasionally punctuated by a burst of repetition or profanity. Sometimes I led, sometimes I followed, but the work always offered something I could not have generated alone. In this sense, the process was aleatory: I spun the wheel, made an assessment, and decided whether to work with what I had or spin again.

The resulting stories were like collages—constructed environments with no fidelity to any single time, place, or set of social and cultural norms, let alone to conventions of fiction. I did away with psychological explanation or backstory, worrying less about supplying all the interpretive keys to the reader. Importantly, I felt no need to display my intelligence or rectitude. In fact, I wanted to diminish it, even eliminate it when possible: it seemed presumptuous, self-serving, fraudulent, cautious. I avoided depth; the stories favored a surface quality, an ongoing “now” that constituted the larger part of the reading experience, rather than an intentional meaning encoded within layers of revelation and detail. Even the smallest of these omissions could push a story further from the sort of polished connotations that marked my previous fiction and toward an unsettling incompleteness.

To some, this may sound like a turn for the worse: No depth? No writerly wisdom? No message? Plenty of superb writers produce work that has these qualities, and my approach may be idiosyncratic, better suited to an interest in stylized, formally self-conscious fiction. In the unsatisfying writing I’d been doing, exemplified by Failures A and B, the hallmarks of literary fiction had become mannerisms, hedges against misinterpretation and being deemed “difficult.” I’d always believed that the best way to write—and especially to take risks—was to ignore what Hans Robert Jauss calls the “horizon of expectations”: the tacit assumptions readers have about genre and literary conventions, aesthetic norms, and prevailing moral standards. During my period of failure, I became preoccupied with that horizon, increasingly attentive to an imagined reader’s needs, and this act of self-sabotage capsized my work. Working with AI allowed me to stop worrying about my readers and, ultimately, about perfecting the artifacts I was producing. My emphasis fell on process: the events at the desk that make your brain light up.

Still, I was filled with doubt. Experimentation can sometimes yield, in a word, horseshit, and these experiments had carried my work into a zone where it was hard to apply familiar standards. Was I abdicating responsibility whenever I allowed GPT to generate the bulk of a draft in response to specific prompts? It’s a fair question, but blurring the boundary between artists’ purely creative work and their more technical role as operators, or even as curators, has a long (and controversial) history, from Marcel Duchamp’s ready-mades to Donald Judd’s fabricated boxes. I’m reminded of Sol LeWitt’s remark about his wall drawings, which were executed by assistants and installers based on written instructions: “The idea becomes the machine that makes the art.”

The distancing mediation of an LLM might be viewed as a moral hazard, a disincentive to police myself due to the intercession of the machine. In theory, I agree, although this feeling of insulation from timid vigilance can be a virtue. Unsettling output produced by LLMs gave me an opportunity to sit with my discomfort and reckon with how the material might change the work. Once, while I was drafting a story with GPT-2, the LLM spontaneously generated a scene depicting a sexual assault in a restroom. It troubled me—it was something I might have hesitated to imagine myself—and yet it belonged to the story. I kept it, and refined it: in the final version, two characters sell pornography produced via spy cams. I was the one making choices about what to keep, what to revise, and what to add, but my work with the machine had nudged me out of sheltered ground, and readers would be nudged along with me. Strange and disturbing scenarios, behavior, and language appeared with no compensatory moral resolution—as they often do in real life.

The doubt that simmered as I wrote my machine stories was different from what I’d experienced with Failure A, Failure B, and every failure in between. I had no ready point of comparison. I could isolate this effect or that one and trace it to Kathy Acker, or Donald Barthelme, or John Hawkes, say, but for the whole, and for the process that developed it, there was no model. I had to feel my way into these stories and into what seemed right. Paradoxically, embracing a new technology that generated literally unoriginal material was what allowed my work to truly feel my own; it made authorship—in a sense meaningful to me—possible again.

Failure A and Failure B did not fail at the level of craft: both are professional work, skillfully done. They failed at the level of daring. I tried to make only what I already knew I could, while anticipating and trying to avoid objections. In a pervasively transactional world, the commercial entertainment industry set the terms, and I attempted to meet them. The work was competent—but competence meant denying myself the opportunity to reach beyond convention and into the open territory where the only measure of success is one’s own assured audacity.

There is understandable anxiety about whether AI will displace the singular and distinctive creations of artistic vision. But LLMs have been in the picture for only a short time, while the most insidious enemies of that vision—cliché, predictability, risk aversion, an overriding concern for market success—have been with us for as long as we can remember. If we really care about creative authenticity, then what we should safeguard is a conception of composition as a patient and entirely human process of looking and feeling, of identifying what’s most useful among available material and recognizing when it’s been transformed into something new and unexpected. If generative AI can facilitate this sort of attention and experimentation, then the real failure will be refusing to engage with it. My own form of engagement has contained plenty of false starts and dead ends—that kind of failing endures—but it has led me toward work whose limits remain genuinely unknown.