The Yale Review | Christopher Sorrentino: “How LLMs Set My Writing…

The Yale Review

Subscribe<br>Donate

Folio

How LLMs Set My Fiction Free

Writing no longer felt like experimenting. AI made it fun again.

Christopher Sorrentino

MOST WRITERS TRAIL behind them an obscure history of failure: work that couldn’t be started, work that couldn’t be completed, and work that was dragged across the finish line and then banished to a drawer. Over the course of my career—thirty-three years, seven published books—I’ve experienced each of these kinds of failure. Sometimes these defeats felt total, but then I’d manage, through trial and error, to stumble onto something new.

Until the novel I’ll call Failure A. In June 2017, I finished a long section of the first draft. I had a nagging sense that there was something wrong with it, but I figured that after a short break I’d continue writing and fix the problems later.

“Later” turned out to be never. During my break, I wrote and sold a memoir. When I eventually returned to the draft, Failure A seemed even more flawed than I’d remembered, the problems I’d already identified only the most obvious instances of a deeper inadequacy I was no longer interested in remedying. Maybe the moment had passed: It happens. I decided to move on.

By then it was late 2019. Within a few months, my family and I were stuck in our apartment waiting out the pandemic, each of us withdrawn to a neutral corner. It was not a productive period. I went through my notebooks and flagged ideas. I started short stories and revisited unfinished ones. I wrote the opening chapters of at least three new novels. I outlined another from beginning to end. I attempted to revive previously abandoned novels.

Finally, in 2022, I began a book that seemed to stick. Call this one Failure B. I worked on it steadily over the next year or so. I experienced relief, but no joy. No eager anticipation before writing, no blissful immersion while writing. It was a project I could live with. That was it. It beat the way I felt before.

Two words immediately come to mind to describe Failures A and B: stale and cautious. Staleness has none of the unselfconscious virtue of schtick, which simply puts itself out there, confident in its effect. I could see that I was borrowing from my own better work and—even worse—I could see how I was trying to disguise it, the way you might cover a worn sofa with a throw and some pillows.

I became skittish and pigeonhearted: making safe bets, avoiding the healthy recklessness that can invigorate a sentence, or a project.

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:

I can tell you that my father never saw me as scared as I was on that Christmas day in London Town. I wanted to know what I wanted to know, I didn’t want to know what I didn’t want to know, and I didn’t want my father to know what I knew, what I didn’t know, what I wanted to know, and what I no longer was interested in knowing. In short, I wanted him to be completely ignorant of what knowledge I possessed, wanted to possess, no longer possessed, or no longer wanted to possess. But perhaps it was so obvious he could simply discern it on his own.

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...