The much-discussed arrival of AI-generated text into the literary realm has cast us into a new era of authorship: nothing you publish online is free from its radar, ever. In this excerpt from her investigative book Searches: Selfhood in the Digital Age (HarperCollins, 2025), Vauhini Vara writes about how the accessibility of these tools, trained on huge data sets but constantly trawling the internet for more, brought an avalanche of ‘polished’, writing. Now, with updated public models of the GPT Vauhini examined for her book as well as other companies’ offerings, writers face an onslaught of opportunities to sharpen their ideas, improve their creativity or standardise their grammar. Meanwhile, big tech firms like OpenAI partner with publishers to mine material for their models. But language is composed and sculpted on human effort; and style, Vauhini reminds us, is fundamentally inimitable.
This excerpt complements our interview with Vauhini, where we spoke about what technology such as AI language models mediate for us emotionally. You can read it here.
My favorite description of what it feels like to write comes from Marcel Proust’s Swann’s Way, when Proust’s narrator recounts a moment from childhood when, while riding in a coach, he noticed three steeples of distant churches. Because the coach was moving down a curved road, the steeples seemed to keep rearranging themselves. The sun was setting, too, and the light kept hitting the spires at different angles. The narrator recalls feeling a wash of inexplicable delight—¬yet also having a sense of “not penetrating to the full depth of my impression.” He borrowed a pencil and paper from a grown-up and wrote a page describing what he’d seen. Proust describes how the narrator felt after this exercise was complete: ‘I found such a sense of happiness, felt that it had so entirely relieved my mind of the obsession of the steeples, and of the mystery which they concealed, that, as though I myself were a hen and had just laid an egg, I began to sing at the top of my voice.’
For me, as for Proust, writing is an attempt to put into language what the world is like from where I stand in it. The language doesn’t exist before the attempt begins; it’s the attempt itself that conjures the language into existence.
One doesn’t need to have published anything to recognize that feeling. One needs only to be human. But to be a writer is to make a career of the pursuit of that feeling. And to be a good writer, Zadie Smith argues in an essay from 2007 called “Fail Better,” is to succeed in it. “A writer’s personality is his manner of being in the world: his writing style is the unavoidable trace of that manner,” Smith writes. “When you understand style in these terms, you don’t think of it as merely a matter of fanciful syntax, or as the flamboyant icing atop a plain literary cake, nor as the uncontrollable result of some mysterious velocity coiled within language itself. Rather, you see style as a personal necessity, as the only possible expression of a particular human consciousness. Style is a writer’s way of telling the truth. Literary success or failure, by this measure, depends not only on the refinement of words on a page, but in the refinement of a consciousness, what Aristotle called the education of the emotions.”
***
But AI language models, being mere parrots, do not have communicative intent.
They neither represent an individual perspective nor model the perspective of a potential reader. The language they emit is all signifier, stripped of significance; any significance we perceive is a mirage. In the line of “Ghosts” in which my sister holds my hand, it might seem, at first glance, that GPT-3 is conjuring my perspective. But there’s a problem with that interpretation—because what it described never happened. I don’t remember any moment when we were driving home from Clarke Beach and my sister took my hand. And it’s not just that. The truth is that I can’t even easily imagine something like it; my sister and I were never so sentimental. Maybe that’s why I found myself so attracted to the line. It was a kind of wish fulfillment. Yet it wasn’t true, which is the reason that, with each iteration, I kept deleting GPT-3’s words and replacing them with mine. The machine-generated falsehoods compelled me to assert my own consciousness by writing against the falsehoods.
In “Ghosts,” I diminished GPT-¬3’s role over the course of the nine attempts, writing a growing proportion of the text myself. In the version of the essay published in The Believer, I gave GPT-¬3 the last lines. In the final paragraph, I wrote, “Once upon a time, my sister taught me to read. She taught me to wait for a mosquito to swell on my arm and then slap it and see the blood spurt out. She taught me to insult racists back. To swim. To pronounce En-glish so I sounded less Indian. To shave my legs without cutting myself. To lie to our parents believably.” It continued, “To do math. To tell stories. Once upon a time, she taught me to exist.” But after its publication and subsequent reception, I decided to revise the piece, reclaiming the last lines for myself. The revised version is the one in these pages. I wanted to make sure it came across that the essay is as much about what technological capitalism promises us as it is about the perversion, and ultimate betrayal, of that promise.
GPT-¬3 couldn’t satisfy me as a writer. This was, for me, the point.
***
ChatGPT’s unveiling, in November 2022, was most people’s first introduction to an AI language model. Two months later, it had become, by one metric, the fastest-growing consumer application to have ever existed. But my own experiments trying to get ChatGPT to write were confusingly disappointing. No matter how many times I ran my queries, the output would be full of familiar language and plot developments. When I pointed out the clichés and asked it to try again, it would just spit out a different set of clichés. At one point, I opened a website from a startup called Sudowrite, which claimed to be able to use AI models including the one underlying ChatGPT to generate fiction. I dropped in a prompt describing the premise of a story I’d already published, called “I, Buffalo.” The story begins with an alcoholic woman who has vomited somewhere in her house but can’t remember where. In my published version, I continued from there in a vein that was meant to be darkly comic. Sudowrite’s version could not have been more different. It featured a corny redemption arc, ending with the protagonist resolving to clean up her act: “She wanted to find the answer to the chaos she had created, and maybe, just maybe, find a way to make it right again.”
I felt somewhat better when I learned that ChatGPT was also disfiguring other people’s writing.
In a Harper’s piece about information on the internet, the writer Ben Lerner described his premise to ChatGPT—involving a young male poet trying to rewrite Wikipedia’s version of history—and asked it to produce an ending for him. ChatGPT responded with seven perfectly anodyne paragraphs, finishing with “And with a heart full of humility and purpose, he continued his journey, guided not by the desire to conquer, but by the genuine pursuit of truth, both as an artist and as a seeker of wisdom in a world teeming with knowledge.” In a New York Times review of maybe the highest-profile AI-generated book to date—a novella called Death of an Author, written using ChatGPT, Sudowrite, and a platform from a startup called Cohere—Dwight Garner dismissed the prose as having “the crabwise gait of a Wikipedia entry.”
I didn’t understand what was happening until I talked to Sil Hamilton, an AI researcher at McGill University who studies the language of language models. Hamilton explained that ChatGPT’s bad writing was probably a result of OpenAI fine-tuning it for the purpose of following instructions. “They want the model to sound very corporate, very safe, very AP English,” he explained.
. If, in building this product, OpenAI’s engineers sought to give it something resembling a single human perspective, it would be that of an extremely well-trained customer-¬service representative.
The outrage by published literary authors struck them as classist and exclusionary, maybe even ableist.
Elizabeth Ann West, an author on Sudowrite’s payroll at the time, who also made a living writing Pride and Prejudice spin-offs, wrote, “Well I am PROUD to be a criminal against the arts if it means now everyone, of all abilities, can write the book they’ve always dreamed of writing.” I also found a comment from a mother who didn’t like the bookstore options for stories to read to her son. She was using the product to compose her own adventure tale for him.
Maybe, I realized, these products supposedly built for writers are actually of more interest to those who identify primarily as readers. I could imagine many of the people employed as authors, people like me, limiting their use of AI or declining to use it altogether while a new generation of reader-¬writers began using AI to produce the stories they want. For them, it might not matter that these tools are parrots, as long as they parrot convincingly enough. And it didn’t seem like a stretch to imagine that, over time, the products’ parroting would grow more sophisticated.
***
One day I opened WhatsApp and saw a message from my dad, who grows mangoes in his yard in the coastal Florida town of Merritt Island. It was a picture he’d taken of his computer screen, with these words:
Sweet golden mango,
Merritt Island’s delight,
Juice drips, pure delight.
The poem belonged to my dad in two senses: he had brought it into existence and was in possession of it.
I stared at it for a while, trying to assess whether it was a good haiku—whether the doubling of the word “delight” was ungainly or subversive. I couldn’t decide. But then I realized my opinion didn’t matter. It was my dad’s poem, not mine.
My dad kept sending me ChatGPT-authored writings. Once, he texted a long list of reasons for the oppression of Dalit people, a community to which we belong, branded as “untouchable” under India’s caste system. He wrote, “AI answer,” then added, “100% correct.” Later, when I sent him passages of this manuscript in which he appeared, he had ChatGPT edit an email to me objecting to my superficial treatment of his experience: “One needs to spend time with people and work with them to write about their struggles. Only then can you produce great work.” It seemed he felt ChatGPT helped him express himself better. But it also subtly changed the tone of his writing. In his original email, which he also sent me, he had written: “Then only you can produce great work.” That version sounded much more like him. The construction “then only” might sound strange to a speaker of American English and, therefore, to ChatGPT, largely trained on American English, but it is perfectly standard Indian English.
***
I bristled, at first, at my dad’s reaction to the writing I’d sent him; he’d also called the parts about him “dry,” with “no spice and no gravy.” But then it occurred to me that I’d probably be frustrated, too, if I read someone else’s depiction of me and found it lacking. My dad had previously sent me a poignant letter he had written to friends years ago, before we moved to Oklahoma for his studies in occupational medicine. I asked if I could quote from it, and he gave me permission. “I am quite excited about these new adventures—some people think that I am going through mid-life crisis,” he wrote. “Others think that I am pure crazy to leave regular practice and stable lifestyle for an uncertain future.” He added that my sister and I were reluctant to move and that we’d all miss Saskatchewan—“friendly, quiet, and quite beautiful.” He closed the letter on an anxious note: “I hope that things will work out OK in the end.”
I also sent my mom the sections of this manuscript involving her, to which she responded with a single-¬spaced six-page document elaborating on her experiences. This, too, annoyed me, until I read the document. My mom tells stories all the time—I learned to tell stories from her—but there were several I hadn’t remembered hearing. In one place, she described joining a PhD program in political science at a Canadian university when she was a young mother. “One of the professors said I was getting too high of a grade on my papers and that he was having doubts about my integrity re: whether they were my papers because I was unable to eloquently explain my reasoning to them verbally,” she wrote. “I was scared to speak in En¬glish because of my accent as everyone was asking me to repeat what I said. That was the final straw for me to decide to quit that program.”
***
It can reasonably be expected that with time AI companies will address some of their products’ early issues. OpenAI found that GPT-4, the large language model that came after GPT-3, improved on some of its earlier models’ shortcomings, though not all, and promised that future models will be better. (Sam Altman called GPT-4 “mildly embarrassing at best” and “the dumbest model that any of you will ever have to use again by a lot.”) When it comes to language models, improving will depend in part on finding more language with which to train the models.
Researchers have found that language models become more accurate when they train on more material, but the text freely available online is running out.
A lot of my own internet-borne language, and yours, is already eligible to be used by corporations—only sometimes with our consent—to build their AI products. Google can use what we publish online to create AI-generated summaries of search results; Amazon can use our products reviews to create AI-generated review synopses; Meta can use our public Instagram pictures to create AI-generated images; OpenAI can use our chats (unless we opt out) to improve ChatGPT’s AI-generated discourse. But even all this isn’t enough. The New York Times reported that OpenAI researchers addressed the need for even more material by transcribing more than one million hours of YouTube videos and that Meta considered trying to acquire the publisher Simon & Schuster for access to its books.
OpenAI is pursuing another tactic, too: paying publishers for their content. It is happening as a lot of the people running media and publishing companies seem to be warming up to AI’s potential. In the Financial Times, the CEO of Bertelsmann, Thomas Rabe said, “It can be very positive provided we stay on top of it and understand its potential and threats.” He suggested, for example, that authors might train AI on their own past writing: “If it’s your content, for which you own the copyright, and then you use it to train the software, you can in theory generate content like never before.” The implication was that if authors like me were to train AI models on our own text, we could produce future writing in our own style, presumably avoiding the ethical problems associated with using language produced by models trained on other people’s language.
But while some researchers theorize that over time language models might need progressively less text, the absolute minimum needed would still likely be far more than the mere thousands of words an author typically composes in a lifetime.
That means that even if I were to fine-tune a model to write language like my own, that model would have to have already been trained on other people’s language.
On top of that, all my past writing represents my past self, my past insights. It, unlike me, is fixed in time. It’s functionally impossible, then, for ChatGPT or any product like it to generate language close to what I, in any given moment, would come up with.
In May 2024, The Wall Street Journal, the place where I wrote the first articles of my career, published an article about OpenAI’s latest deal. “Wall Street Journal owner News Corp struck a major content-¬licensing pact with the generative artificial-intelligence company OpenAI, aiming to cash in on a technology that promises to have a profound impact on the news-publishing industry,” the article began, explaining that “OpenAI would use content from News Corp’s consumer-¬facing news publications, including archives, to answer users’ queries and train its technology.” The reporters added that OpenAI could pay News Corp the equivalent of $250 million over five years—specifying that this would come in the form of both cash and “credits for use of OpenAI technology.” In an email to employees, News Corp’s CEO, Robert Thomson, called it a “providential opportunity.”
Other news executives quoted in the article echoed Thomson’s rhetoric. William Lewis, the CEO of The Washington Post—which is owned by Jeff Bezos—said that his paper was “in the market for significant partnerships.” Louis Dreyfus, the CEO of Le Monde, the French newspaper, said, “It is in my interest to find agreements with everyone.” Dreyfus was more direct than the others about his reasoning: “Without an agreement, they will use our content in a more or less rigorous and more or less clandestine manner without any benefit for us.” OpenAI’s Sam Altman declared, meanwhile, “Together, we are setting the foundation for a future where AI deeply respects, enhances, and upholds the standards of world-class journalism.”
When I read The Wall Street Journal’s news, I wrote to the union representing Journal reporters, asking what my rights were over my articles for the paper—more than five hundred of them published over a decade.
Tim Martell, the executive director of the union, wrote back about a week later, apologizing for the late response; I would later realize that in the interim the Journal had laid off eight reporters, following a round of major layoffs earlier in the year. “We still do not know all (or actually, any) of the details of News Corp’s deal with OpenAI,” he wrote. “On the licensing of material from WSJ, if OpenAI has access to all Journal content, that means they’ll be able to use any material that appeared under your byline while you were here.” He attached an agreement that all new employees must sign at hiring, assigning the Journal all “right, title, and interest” to journalism published for the paper. “You likely signed one,” he wrote. I likely did.
***
Nothing, it seemed, could be done about it. That deal with my former employer was one of several that OpenAI made around that time; the Associated Press and The Atlantic also signed up, among lots of other high-profile publications. OpenAI started testing search features using material from its publishing partners; someone might ask when the upcoming Olympics would take place, for example, and get a text answer with a link to the information source in parentheses. I could understand the pragmatic defeatism of the Le Monde CEO. Google’s AI-¬generated search summaries also relied on information from publishers, but in its case without partnering with and compensating them. OpenAI’s approach at least seemed fairer than that.
OpenAI isn’t just courting publishers; it’s also courting writers themselves.
One afternoon in August 2024, I got an email from someone named Jay Dixit, a former journalist who described himself as a member of OpenAI’s community team. He explained in the email that his role was to engage with writers; he’d contacted me after reading “Ghosts” and feeling moved. “I think your story could be an inspiring example for other writers on how to use ChatGPT as a creative collaborator that works in service of the writer’s vision, using it not to generate copy but as a catalyst for their own creativity,” he wrote.
I agreed to a Google Meet call with Dixit. He came across as congenial and open; we bonded over our shared Canadian roots, and when I told him I had a bunch of questions, he obligingly answered them. I was curious, most of all, about his gig—about the goal, for OpenAI, of all this courtship of authors. He explained that he planned to promote published AI-¬assisted writing as inspiration for others interested in trying it; he also wanted to recruit authors using AI to help inform product development. It was clear that Dixit, a technophile, felt he’d scored a dream gig; he’d heard more than one thousand people had applied. I imagined the position probably paid better than what a journalist could reasonably expect to make. I didn’t ask him about his compensation, but when I pulled up the job listing for a similar role at OpenAI dealing with visual artists, it listed a pay range of $190,000 to $240,000 a year.
The community team would include, along with Dixit, a liaison each for visual artists, educators, and internet creators. I asked why the outreach team was focused on these arts-and culture-oriented people and not, say, physicists or retail workers. “Physicists already use it,” he said. They didn’t need outreach; artists did. “A lot of writers and other artists have been skeptical and critical, rightly so, and that’s led to them avoiding using it,” he added, “and I think there’s a danger there of them being left behind.” My first reaction was to scoff. This was exactly what a hired propagandist, paid to disarm culturally influential skeptics, would claim. But then, he wasn’t lying about the risk for people in cultural fields of being left behind, especially now that those responsible for compensating us were already folding. The question, it seemed to me, was what to do about it.
