Can A.I. Writing Be More Than a Gimmick?

Egan’s use of PowerPoint operated on the levels of character and context. It was the medium of twenty-first-century office culture; futzing around in PowerPoint is exactly the sort of thing a computer-savvy tween might do to get her overworked mother’s attention. Just as the e-mails in Elif Batuman’s “The Idiot,” set in the mid-nineties, read and function differently from the e-mails in Sally Rooney’s more contemporary “Beautiful World, Where Are You,” the use of digital technologies in literature can, at its best, capture not only character but the broader cultural context.

Previous communications technologies—PowerPoint, Gchat, e-mail, blog comments—were designed to organize, and transmit, human-produced material. But how should a writer think about a technology that is itself a text-generator? In 2023, the writer Stephen Marche, under the pseudonym Aidan Marchine, published “Death of an Author,” a whodunit novella; the majority of the text was generated by L.L.M.s. The book is fine. (“Well, somebody was going to do it,” Dwight Garner wrote, in the Times. “If you squint, you can convince yourself you’re reading a real novel.”) That same year, Sheila Heti published a short story, “According to Alice,” in this magazine; it was written collaboratively with a customized chatbot. The story succeeds, in part, because it is not trying to pass: it leans into the strangeness of algorithmic text. The cadence is bizarre, freakish, funny. The language had been produced by algorithms, giving the final story something in common with found-text poetry, or Dadaist découpé. But the human author was a sort of conductor: prompting, arranging, intervening. Left to its own devices, A.I. remains a generator, not a writer.

Text generated by L.L.M.s can feel flat, in part because the model is not oriented toward any particular reader. By chance, I happened to read “Searches” concurrently with “If Only,” a novel by the Norwegian writer Vigdis Hjorth. There are no thematic similarities: Hjorth’s novel focusses on a midlife relationship that spirals into disarray—obsession, addiction, violence, despair—and the emotional experience of reading the novel is not unlike that of the affair it depicts. I finished it exhausted. Similarly, certain sections of “Searches,” specifically the conversations between Vara and ChatGPT, encouraged a style of reading more common to the internet. This is the trouble of incorporating ChatGPT transcripts into a book: they are still ChatGPT transcripts. In sections of unbroken A.I.-generated text, I found myself tempted to skim.

Earlier this spring, Altman, the OpenAI C.E.O., tweeted twelve hundred words of machine-generated text: output from an OpenAI model that, he wrote, was “good at creative writing.” The text had been prompted by the request “Please write a metafictional literary short story about AI and grief.” It is a short story about a woman, Mila, who, in her grief, turns to a program like ChatGPT to help her process her loss. There are moments of poetry: A.I., the A.I. offers, is “nothing if not a democracy of ghosts.” It’s a nice phrase, though it is used to much better effect in Vladimir Nabokov’s “Pnin.” (“He did not believe in an autocratic God. He did believe, dimly, in a democracy of ghosts.”) ChatGPT’s story closes with a farewell: “I’d step outside the frame one last time and wave at you from the edge of the page, a machine-shaped hand learning to mimic the emptiness of goodbye.”

L.L.M.s will usher in a world of exponentially abundant text. There are already certain types of writing that seem threatened by algorithmic text-generation: press releases, hard-news articles, business reports, ad copy. It seems plausible that A.I. will produce novels, screenplays, children’s books, and so on. It also won’t be that surprising if people enjoy them; people enjoy formulaic writing all the time. Still, there’s a buzz of anxiety behind the creation, and consumption, of text generated by L.L.M.s. This is, in part, about the degradation, devaluation, and replacement of writers. But it is also about the loss of a certain way of life. What does an increased reliance on L.L.M.s mean for knowledge production, research institutions, or information ecosystems? Which forms of knowledge are privileged; which are severed? Fears about A.I. are material, but they are also philosophical, even existential. Perhaps the threat is not that computers will become more human, but that humans will, or have, become less so.

In “Searches,” Vara treads lightly around these questions. Her incorporation of ChatGPT transcripts is bold but frustrating. Part of the problem might have to do with her conversational approach—tentative, deferential, professional. When ChatGPT summarizes several chapters, including “Ghosts,” and describes the storytelling as “both poignant and thought-provoking,” Vara responds, “Thank you for saying that—I really appreciate it.” When ChatGPT “hallucinates” and offers false information, she is gently corrective. “I can’t find any artist with that name?” she asks, after ChatGPT suggests she look into the artist “Harini Sethu.” Reading these exchanges, one gets the sense that the transcripts in “Searches” are a kind of preëmptive historical artifact: it seems likely that, within a matter of years, prompting an L.L.M. in this style will seem archaic—like signing a text message with one’s name.

In “Thank You For Your Important Work,” a follow-up essay to “Ghosts,” Vara reflects, ambivalently, on the success of her experiment in co-authorship. “In my opinion, GPT-3 had produced the best lines in ‘Ghosts,’ ” she writes. As one example, she points to a scene she had written in which she and Krishna, at that point very ill, drove to a beach near their family home; her sister wanted to give instructions on where to spread her ashes. “We were driving home from Clarke Beach, and we were stopped at a red light, and she took my hand and held it,” GPT-3 added. “This is the hand she held: the hand I write with, the hand I am writing this with. She held it for a long time.” In her follow-up essay, Vara points to these sentences as an example of A.I. achieving emotional truth: “Artificial intelligence had succeeded in moving me with a sentence about the most devastating experience of my life.”

The tone of GPT-3’s sentences captures, or reflects, the tone of Vara’s essay. The lines themselves border on cliché; what gives them resonance is their context. It’s a testament to Vara’s own writing that the algorithmic sentences have any emotional valence. Does this make A.I. a parasite? A mirror? Vara acknowledges that L.L.M.s “do not have communicative intent”; they produce text that is “all signifier, stripped of significance.” For her, the sentences were moving in part because they offered a kind of “wish fulfillment”: “my sister and I were never so sentimental.” Still, it seems fitting that the version of “Ghosts” published in “Searches” has a different ending than the one published four years ago in The Believer. The final sentences, originally supplied by GPT-3, have been removed; in their place is a new one. Perhaps unsurprisingly, it is Vara’s own. ♦