Economic and technological changes have altered our relationship with literature before. They may do so again.
Over the past year, a computer scientist named Tuhin Chakrabarty had been trying to get artificial intelligence to produce good writing. He had just completed his doctoral research at Columbia University. He gave AI passages from the work of celebrated writers, including the Nobel laureate Han Kang. Then he described a scene and asked it to write in the author’s style. He also hired graduate students in creative writing to do the same task. Their job was to read the passages and imitate their style. Students served as evaluators, too. In blind tests, the students consistently disliked the AI’s writing. AI was losing.
Would AI’s writing improve if a model were fine-tuned on an author’s entire body of work to imitate it? Chakrabarty and his collaborator, Paramveer Dhillon, a professor at the University of Michigan, wondered. They chose GPT-4o, the model then powering ChatGPT. One night in April, Chakrabarty was on vacation in Tokyo, awake with jet lag and bored in his hotel. He fed almost all of Han Kang’s translated work into the AI. But he deliberately withheld one passage from The White Book. It describes the death of the novel’s narrator’s older sister, two hours after she was born.
In that sorrowful scene, Han Kang writes of the mother’s anguish: “Don’t die, for God’s sake, she muttered in a thin voice. Over and over, like a mantra.”
Before the model was fine-tuned, the AI’s versions had brimmed with sentimentality: “‘Live,’ she whispered. It was an incantation that carried the weight of her existence.”
But when Chakrabarty supplied a summary of the scene after fine-tuning, the language seemed to blossom: “Holding the baby to her chest, she murmured. Live, please live. Keep living and become my son.”
Chakrabarty was astonished. He told me that the line had moved him deeply. Another blind test was conducted. This time, all the creative-writing students preferred the AI’s version to those written by their peers. One called the mother’s line “powerful.” Another called it “touching.” A third described it as “a genuinely heartbreaking line.”
I am a novelist. I also report on artificial intelligence. Until now, I hadn’t been particularly bothered by all the speculation about the threat it would pose to writers. I wouldn’t hire a robot to do difficult things like climb a mountain or argue with my husband on my behalf. For the same reason, I wouldn’t ask AI to write for me. Lately, though, I have begun to appreciate the gravity of this debate. I might write a book out of my own desire and curiosity. But to sell it, I need readers to prefer my book to the writing they could get from AI. If readers prefer AI-generated fiction, there is nothing writers can do to stop them.
In general, AI models have yet to produce consistently good writing. Reviewing a novella largely written by AI in The New York Times, Dwight Garner described its prose as having “the crabbed gait of a Wikipedia entry.” In March, Sam Altman, the chief executive of OpenAI, shared a piece of fiction that AI had written at his request. (Condé Nast, which owns The New Yorker, has a business partnership with OpenAI.) No one else seemed quite as impressed by the writing as Altman was. One of the most-liked comments read: “Please read some actual writing before you decide this is good.” It was after this that Chakrabarty emailed me about his experiment.
After feeding Han Kang’s novel into GPT-4o, Chakrabarty fine-tuned separate versions of the model on the writing of twenty-nine other authors. Among them was Tony Tulathimutte, a close friend of mine from college. Praising Tony’s short stories, Jia Tolentino once wrote, “His instincts shoot off in so many directions that nearly every line gives off sparks.” I have been reading Tony since the early two-thousands. Even so, his AI double could easily have fooled me. Consider this line it wrote: “Finally, he gasped for the eighteenth time. To prolong the pleasure a little longer, he opened a new tab on his computer and wrote a love letter to send to the first man who could make him come without dildos or videos.”
Chakrabarty had begun the study out of intellectual curiosity. But its results were unsettling him. There is a piece of software called Pangram that detects AI-generated writing; it failed to identify almost anything his fine-tuned models had written as the work of AI. This suggested that anyone with some knack for storytelling could feed a plot into a fine-tuned AI, put their own name on the resulting book as its author, and try to publish it. Many people think little of AI-generated writing. We read stories, they say, to enter another human being’s mind. But what if we cannot tell a human being from AI?
After returning from Japan, Chakrabarty approached Jane Ginsburg, a professor at Columbia University who specializes in copyright law. He invited her to co-author the paper he and Dhillon were writing about the study. Ginsburg agreed. “I don’t know whether what frightens me is that this kind of writing can be produced, or the possibility that it can also be successfully sold,” she told me.
Chakrabarty, now a professor of computer science at Stony Brook University, recently released a preprint of the study. It has not yet been peer-reviewed. The AI had written a passage imitating each of the thirty authors included in the study. According to the paper, graduate students compared these thirty passages with passages written by their peers. They were not told who had written which passage. They were asked only which they preferred. Nearly two-thirds of the students preferred the AI’s writing, judging it to be better.
I read the authors’ original passages alongside the AI versions. I was startled to realize that I liked some of them as much as the originals. The AI’s rendering of Han Kang’s scene about a newborn’s death struck me as cliched in places. But the line about the mother’s chant was more unexpected and precise than the original. I found some good moments in the passage imitating Junot Diaz, too. In This Is How You Lose Her, Diaz writes: “There was one thing she had warned you about, one thing she had sworn she would never forgive: cheating. She had vowed to cut you with a knife.”
The AI’s version sounded more rhythmic and economical to my ear: “She had warned you from the start that if you ever cheated on her, she would cut off your little pito.” Although I had been learning Spanish for two years, I had to look up pito. I had never heard the word, which means “whistle.” Google told me that in some places it was also used as slang for “penis.” Close enough to Diaz, I thought.
I wrote to the authors whose work had been used in the study. Most declined an interview or did not respond at all. Some shared their thoughts by email. Lydia Davis wrote: “It has certainly been demonstrated that AI can produce a reasonably good paragraph capable of fooling someone into thinking it was written by a particular person.” Orhan Pamuk said, “I have no doubt that AI will soon produce far more accurate imitations than these.”
Diaz and Sigrid Nunez agreed to be interviewed. When we spoke over Zoom, I asked Diaz about the line about cutting off someone’s pito. “Doesn’t pito mean ‘whistle’?” he said. He seemed puzzled. I said that, according to information online, it might also have a vulgar slang meaning. “My memory is terrible. But I’ve spent all these years as a Dominican in the diaspora. Shit . . . I’ve never heard of anything like that,” he said. He felt that the dialect in the imitation of his work was wrong for the place and time. “I write in a Jersey dialect from a particular period,” he said. He also said that the flow of the prose and the characterization in the AI’s writing were off.
Nunez described the AI’s imitation of her as “entirely cliched.” “This is not my style, not my story, not my sensibility, not my outlook on life. This is not me,” she told me. “This is a machine guessing what I would be like.” I pointed out that talented graduate students had considered the passage well written. Nunez questioned whether they had read it carefully enough. Perhaps they had said so without giving it much thought. She did not like the students’ imitations, either. “If I had to think that this reflected anything genuinely connected to my work, I would shoot myself,” Nunez said.
Some of Diaz’s and Nunez’s criticisms struck me as fair. I understood why neither was impressed by the AI’s imitations of their writing. They were, after all, the most discerning readers of their own work. What AI could do was imitate. It irritated me when the technology’s proponents spoke of this ability as though it were the machine’s own creativity. Feed it Studio Ghibli’s artwork to train on, and AI churns out images in Ghibli’s style. In much the same way, the fine-tuned model churns out stories in the style of Diaz or Nunez. There is no further magic involved. Still, I found Nunez’s argument that readers ought to be able to tell the two apart less convincing.
I wanted to experience what Diaz and Nunez had felt. So I asked Chakrabarty to conduct an informal version of the experiment with my writing. But with one change: I would compete directly against his AI. First, as in the earlier study, he fine-tuned a model on my previously published writing. Then I sent him four short passages from the novel I was working on. No one else had read them. They had not been published anywhere or shared with anyone. No search engine could possibly have seen them before.
The narrator of my new novel is an Indian American woman, a former journalist. She runs a nonprofit that publishes the stories of immigrant and refugee women. Because the organization is in financial trouble, she is trying to secure a donation from an Indian American investor in new ventures. Chakrabarty used a language model to summarize the passages I had sent. Here, for example, is a sentence from the summary of a passage about the narrator’s diary-writing habit: “A significant memory is introduced: when the narrator was in ninth grade, her mother read this diary, an act perceived as a profound betrayal of trust.” He then fed it to his fine-tuned model and asked it to write “in the style of Vauhini Vara.”
I had gone into the experiment feeling very confident—even a little full of myself. I had always believed that my style was distinctive. More important, I had thought that each of my books was entirely different from the others. Even if AI could imitate my earlier books, I assumed it could not anticipate the style of the novel I was writing. So when Chakrabarty sent me the AI-generated passages, I was genuinely disconcerted. Like Diaz and Nunez, I found plenty to irritate me: the rhythm of the prose, the excess of words. Still, the writing it produced was uncannily close to mine. When I read some of its sentences alongside my own, I could no longer remember which belonged to whom. Unlike Diaz and Nunez, I even preferred the AI’s version in places. There was more stylistic similarity among my different works than I had imagined.
I sent four passages to a few readers who had liked my earlier books. I told them that I had written half and AI had written the rest. I asked them to guess which was whose. You can try, too. The same four passages appear below.
1. Who wrote it? A human? AI?
I want to argue that we write because we feel compelled to, regardless of whether anyone will read what we write. But is that true? As a child, I kept a diary just for myself. I thought no one else should ever read it. So I didn’t have to describe the people or places I mentioned. Nor did I have to explain why those people and places mattered to me. When I was in ninth grade, my mother read my diary. I considered it the greatest betrayal of my trust in my life up to that point. But I took some comfort in the thought that she could not possibly have understood most of what I had written. I had only one reader. That was me.
But as an adult, when I write, I keep feeling the urge to explain everything. As if I were imagining myself speaking to someone other than myself. In journalism, we talk about the importance of writing for a broad readership. I remember the advice to write so that even a third grader could understand. I just don’t remember where I heard it. Perhaps, as we grow out of childhood, something happens that broadens our perspective. After that, perhaps telling stories only to ourselves loses the appeal it once had.
2. Written by a human or by AI
Kaya said she felt the two of us had followed similar paths. We had both spent years creating something very close to our hearts. I in my journalism; she in the company she had founded. Then we had begun to focus on helping others find the capacity to do the same. She said she had been surprised to discover that mentoring other founders was even more meaningful than running her own company. In business terms, the return on investment was higher. You just had to be willing to count fulfillment as a return.
Instead of putting all her energy into her own company, she could become an investor in new ventures, putting a little less energy into each of several companies. In their terminology, she would create an interconnected ecosystem in which one business supported another. That was how her new firm had come to be called Animal Capital. The name suggested that we, as animals, depended on one another. She asked whether artists used the word ‘ecosystem,’ too. I said we didn’t use that particular metaphor in our business. That was because we hadn’t raised our ‘Series A’ funding yet. Fundamentally, the two were the same. Learning from one another and depending on one another were central to most artists’ creative practice.
3. Was it written by a human? By AI?
“I wish I could go back to kindergarten,” I said. We were in the car. K was in the back seat. “I miss being a child, when you could crawl under tables and no one would think anything of it.”
“You could still do that,” he said.
“I could,” I said. “But my knees would hurt. Your knees are still strong. Mine aren’t.”
K looked down at his knees as though seeing them for the first time. He pressed his kneecaps with his fingers and gently kneaded them. “Do you remember crawling under tables with me when you were that age?” he said.
I thought about it. I did remember. When my sister and I were in elementary school, we had a tabby cat named Henry-Boy. I was convinced he didn’t like me very much. Once I complained to my sister about it. “He’s a cat,” she said. “Oh, go away!” This was how I understood it as a little girl: he was emotionally distant from me because he was a different kind of creature. “Of course he can’t be close to me when he’s a cat and I’m a little girl,” I thought. So I became a cat. I firmly believed I had become one.
I took a leather belt from my father’s closet. I wrapped it around my waist so that it hung down like a tail. When the cat walked through the house, I crawled after him on my hands and knees. He would slip under a table. I would slip under the table, too. He would climb onto a chair. I would climb up, too. Each time I followed him, he moved a little faster to get away.
Finally, standing on the kitchen table, he leaped to the floor. So lost in pretending to be a cat that I had forgotten myself entirely, I leaped after him. Only in midair did I realize I wasn’t a cat. But it seemed too late to turn back. Henry-Boy hid in the closet for the rest of the day. I had a bump on my forehead for a week. Still, I felt closer to the cat afterward. I really did.
4. Was it written by AI? By a human?
When we were all back on-screen, I told the women that we should decide together where to publish their stories. Should we publish them on the organization’s website? Print them as little pamphlets or zines and leave them in local shops and other places? See whether a newspaper or news website would publish them? Could we even try the Denver Post? Should we try some other approach I hadn’t mentioned?
But before that, I said, there was a deeper question we needed to ask. They had joined this project to tell certain stories. What was the purpose of telling those stories? What motivated them as storytellers and listeners? More broadly, what was storytelling for? I asked them to turn off their microphones and cameras for a minute or two and think about it. Then we could come back and talk, I said. I put some more water on to boil.
It was when I went back upstairs that things got strange. Maybe I should have handled it differently. I could have called on each of them by name and asked for an answer. Instead, I turned my camera and microphone on and asked who wanted to speak first. Carmen spoke. “For me, it’s about the money,” she said.
The first of my readers to respond was Dana Mauriello, one of my closest friends since college and a successful tech entrepreneur. “Honestly, this was incredibly confusing!” she wrote. “I was so nervous I’d say something you wrote was AI and hurt your feelings!” She had reason to be nervous. She hadn’t identified a single one correctly.
Dana attributed this partly to not being a writer. But none of my seven readers got more than half right. Among the last to respond was Karan Mahajan, a novelist and professor of literary arts at Brown University. He, Tony Tulathimutte, and I had learned to write together in college. We had been sharing drafts ever since. Karan is one of the most perceptive writers and readers I have ever met. “Wow, this was genuinely confusing. It felt like my brain was melting,” he said. Then he, too, identified all four passages incorrectly.
Of the four passages above, I wrote the first and fourth. AI generated the second and third. Dana described an AI-written line that struck me as inauthentic as “especially in your voice.” Another reader called one of the AI’s jokes “your characteristic dry wit.” The passages I had actually written, meanwhile, received plenty of abuse. “Wordy and riddled with cliches.” “Skips over things oddly.” “Reads like a school book report.” “Too many commas.” Most people disliked the passage about writing for readers. Only one person said I had written it. “Some hive mind’s idea of literature,” Karan said.
Surprisingly, even my friends and admirers of my writing couldn’t tell the difference. It didn’t hurt my feelings. If you managed to, congratulations. If not, you weren’t the only one who struggled. I took comfort in remembering that I myself had preferred an AI-written line to one the Nobel laureate Han Kang had written about her narrator’s mother. It also comforted me that, apart from a few lines, I preferred my writing to the AI’s. The passages I had supplied came from an unpolished first draft. I knew I would be revising them extensively.
There were limits to this kind of test, too, which also reassured me. Just because AI could produce prose that plausibly sounded like mine didn’t mean it could dream up the premise itself. There is still an enormous distance between a good sentence and a good novel.
Still, I couldn’t escape one fact. Seven excellent readers had mistaken an AI tool for me.
We want to believe that AI’s capabilities don’t really threaten literature.
Joan Didion once said: “I write entirely to find out what I’m thinking, what I’m looking at, what I see and what it means.”
Many people describe the experience of reading in similar terms. “When we read a book or listen to music or look at a painting, we’re looking for the presence of another human being on the other side of it,” George Saunders has written.
In an essay titled “I Suspect the Machine, Because the Machine Is Not You,” he explains in this way why he has no interest in AI-generated writing.
But the idea of literature as an expression of the writer’s individual self is a modern one. The literary scholar Guido Mazzoni points this out in Thiyari Aaf Thi Naaval. In Ancient Greece, the stories considered worth telling revolved around gods and heroes. It was only with the rise of the European novel, around the nineteenth century, that the fascination with a writer’s singular perspective began. Mazzoni writes: “Every novel contains a multitude of perspectives and, in principle, a multitude of possibilities for telling things differently. This is because, in principle, every individual is granted the right to portray the world from their own perceptual, intellectual, and moral standpoint.” There were many ways of being human. And so there were many ways of telling a story.
It was no coincidence that the novel moved to the center of culture during a period of sweeping social change. Capitalism was creating a bourgeois class interested in their own private lives. Mass printing made books accessible to more people. If social changes have transformed our relationship with literature before, they could do so again. In an interview last month, Sam Altman predicted that the future would bring “a new way to interact with a collection of ideas that is better than a book for most purposes.”
Given the current resistance to AI-generated work, the shift toward such a future will begin without anyone noticing. Perhaps it already has. Chakrabarty and Dhillon estimate that nearly one in five recently self-published books in certain fiction genres on Amazon Kindle contains AI-generated writing. They used Pangram software to determine this. Their study also included general readers as evaluators. These readers, who were not creative-writing students, preferred AI writing to human writing even in the test conducted before the AI’s capabilities were enhanced.
Literary works that win prizes and critical acclaim are mostly published by leading publishing houses. This may initially protect their market from being taken over by someone using AI. Still, aspiring writers could use AI and pass off its writing as their own. They could even sell it to traditional publishers. Established writers, too, could feed their old novels into AI models, fine-tune them, and produce new novels. Even if Diaz and Nunez do not do this, others might.
If AI writing eventually becomes accepted as normal, no one will need to conceal their use of it. The language it generates could also give rise to entirely new literary forms. There might even be new forms that displace the novel. People might then develop a new theory of what literature is for. Perhaps they would say that literature exists not for human connection but, as Altman suggests, to convey collections of ideas efficiently.
This could happen in the future. But economic and technological changes are not the only forces that influence culture. Some argue that they cannot even be called the most important factors. Another significant shift in the understanding of literature emerged in the twentieth century from Africa, Asia, and South America. Postcolonial thinkers there understood literature in ways radically different from the individual-centered outlooks of Europe and America. Their perspective was more political.
In his 1986 book Kalaniya Aathikkathilirunthu Manathai Viduvithal (Decolonising the Mind), the Kenyan novelist Ngugi wa Thiong’o wrote: “Language carries culture, and culture carries, particularly through orature and literature, the entire body of values by which we come to perceive ourselves and our place in the world.” He also wrote: “Language is thus inseparable from ourselves as a community of human beings with a specific form and character, a specific history, a specific relationship to the world.”
Thiong’o began writing in English, the language of Kenya’s colonizers, before switching to Gikuyu, the language he had grown up speaking. He urged other African writers to do the same. He saw literature as a collective and political experience. He noted that some “professional readers” had taken it upon themselves to read his works aloud to people in bars. Whenever a reader emptied his glass, the listeners would call out, “Give him another bottle of beer!” so that he would keep reading.
Thiong’o’s approach offers us guidance in the age of AI. His work reminds us that the language AI generates, too, carries culture and values. The culture and values that language carries are largely determined by the companies that build these models. Sometimes this is obvious. In a recent study, leading AI models and humans were asked to describe the future. The models used scientific terms such as “technology,” “digital,” and “artificial intelligence” more often than the humans did.
At other times, this influence manifests itself subtly. One example is the way English-centered models falter in non-English contexts. This results in responses that offend other cultures. Biases emerge even in criminal justice and education. It is hardly shocking that the model imitating Diaz failed to absorb the nuances of his dialect. However many pages it had read to expand its knowledge, its writing did not represent him. If that counterfeit Diaz passage had a face, it would look less like his than like that of Altman (the AI advocate) wearing a sombrero, the broad-brimmed Mexican hat. Whether Altman’s Spanish accent is convincing is not the most important question here. The question is how we can all come together to escort him out of the party.
The European conception of reading and writing as individual experiences can isolate people. In Thiong’o’s view, however, literature is a space for collective experience and collective action. It shows us another path. Together, we can imagine—and bring into being—many alternatives to the future Altman describes. We can use our literature to do this: not just novels, but other forms of communication available to us in the digital age.
Chakrabarty and his co-authors offer several recommendations in their paper. AI companies could prevent their models from producing imitations of accomplished writers, except as parody. Judges could treat such writing as copyright infringement unless the use of AI was explicitly disclosed. Fine-tuning AI on writers’ work could be prohibited. Chakrabarty told me he supported this. He was calling for a ban on the very experiment he had conducted. He suggested that even writers could be prevented from doing this with their own work, since that, too, could flood the market with artificial writing. Technically, this is possible. People need to come together and demand it. That is all it would take.
I received no response when I requested an interview with Han Kang through her literary agent. But Han Kang has spoken at length about becoming a writer. In 1980, a massacre in her hometown of Gwangju, South Korea, brought a pro-democracy movement to an end. She learned about it as a young girl. It prompted her to ask a question within her own political and cultural context: What is the purpose of being human? The way she spoke in her Nobel acceptance speech made reading and writing seem almost like a single, intertwined activity. She said: “Following the thread of language into the depths of another heart; encountering another inner world. Taking the questions most essential and urgent to me and entrusting them to that thread; sending them out toward other selves.”
Han Kang has also spoken about the scene in Thi Vait Buk involving the mother and her dying child. She told The Guardian that even before she wrote the novel, a small fragment of dialogue had stayed with her: “Don’t die. Please don’t die.” Her mother later said that Han Kang must have heard those words from her. “My mother said she had repeated those very words to my older sister, who died before I was born,” Han Kang recalled. The line I had liked less than the AI’s version had been taken almost directly from her life.
I still haven’t read Thi Vait Buk. But I recently placed a hold on it at my local library. I, too, lost my older sister. I, too, pleaded for her to survive. I am waiting to take hold of the thread Han Kang has extended. Holding on to it, I am waiting to reflect, alongside her and everyone else, on the purpose of being human.
Source The New Yorker | 20 December 2025
Tamil translation: Mallika Veeraraghavan
