A New Kind Of Creative Poverty

Afra Wang is a California-based writer and journalist who focuses on technological modernity through a humanistic lens, spanning China and Silicon Valley in the global AI age. She writes concurrent.blog and is a frequent contributor to many outlets. Editor’s Note: Noema is transparent about AI use in its pieces. We publish original human-generated ideas but allow authorized, disclosed use of AI in certain cases. Please see details at the end of this piece. During a dark stretch of Covid, I dove into Joan Didion’s “The Year of Magical Thinking,” her book about processing her husband’s death. Didion wrote that she could not give away his shoes. She had managed the suits, the shirts, the jackets. But when she got to the shoes, she paused: “I stood there for a moment, then realized why: he would need shoes if he was to return.” I was living in solitude in an apartment recently emptied by a long-term relationship. When the rental lease lapsed, I renewed it and renewed it again, keeping the furniture exactly as it had been, the lamp at the same angle. What Didion was dealing with was, in Orwell’s words, “the recognizable experiences of human beings.” Through her writing, I felt relief wash over me as I suddenly saw my own stalled life right under my nose. What many readers come to Didion for — what I went to her for — is the record of a mind at work on its own chaos, a way of processing thought, of handling the formless weather in the head. I love watching her spill it all out, like discovering the neighbors’ house is total chaos, and feeling assured that no one’s life is truly put-together. Being gripped by the experiences evoked by writers like Didion could be a survival trait, shaped over countless years of evolution. The psychologist Bruce Hood argues that we are a self-domesticated species: Over roughly a thousand generations, human beings have favored cooperation and sociability as traits in our mates and allies. In doing so, we rewired our brains to read, track, mimic and crave other minds. We have been self-domesticating, Hood writes, “through the invention of culture and practices that ensure that we can live together,” and eventually win the evolution game. To be gripped by another person’s inner life is old equipment built into us. Social media is a product of this ancient urge, a machine for exercising our interest in other humans at a scale evolution never planned for. But what happens when we learn that this other human, or this “recognizable experience,” has been artificially generated by an LLM? Maybe this is why the counterfeit cuts the way it does. AI slop feels like betrayal. There is this haunting feeling when you lean toward a face and then find only a mask, with no person behind it. And this feeling is probably one reason people have become obsessed with running the words they read through Pangram, a tool to detect AI writing, or other such detectors. Almost every writer I know has some visceral repulsion at the tells — “tapestry,” “crystallize,” “delve,” or “it’s not X, it’s Y” — that low-effort, overfitted register. Probably no one online reacts more violently than the British writer Sam Kriss, in an essay titled “If you let AI do your writing, I will come to your house and kill you”: I hate it. I find it viscerally disgusting; a cold shudder like someone’s poured jelly down the back of my neck. I hate that it’s everywhere; I hate that when I read basically anything now I’m constantly on alert, twitching like a schizo in an underpass. Is this thing really what it says it is? Is this person actually a robot in disguise? Nice little personal essay you’ve got there, lady, but I know what you really are; time to get my knife out, time to start digging around under your skin until I find the wires. This urge — when a sentence feels off, and you must dig beneath its skin to find the wires — might be familiar to you. At a conference not long ago, I watched Chris Best, Substack’s chief executive, and Oliver Habryka, who runs the rationalist forum LessWrong, discuss this disturbance. Both platforms, in their early days, were built on the assumption that a reader can trust the writing in front of them to be authentic and human (and distinctive from institutional writing). Today, that assumption no longer holds. LessWrong now requires its writers to disclose when they have used AI. As of publication, Substack has also recently launched a feature using Pangram-style detection to reduce “Claudefishing,” so readers don’t “unwittingly invest their attention in something with no human thought on the other end.” Everyone seems exhausted by the constant tax of suspecting AI behind every sentence. The tech industry’s actions and inaction (LinkedIn, for instance, has become a dumpster of AI slop people laugh at) reflect a growing backlash against soulless slop. I, too, have struggled with anger over sloppy AI writing that has made me viscerally uncomfortable. My reactions can be a little pathological: I fanatically delete my own em dashes, and I once even mistakenly thought a few sentences a writer friend wrote were AI-generated. But I also must confess to my own relationship with AI: Because English is my second language, I rely heavily on Claude as a grammar checker— to smooth the bumps in my sentences and remove the ESL flavor. I used to have a somewhat cringy analogy to describe writing in English: installing a pair of mechanical wings. These wings carry me to vast terrains and audiences I couldn’t reach before, even though they sometimes feel artificial. AI makes those wings less heavy, but more not-entirely-me. So when I do get angry at AI writing, I also try to understand where the anger comes from — and then I wonder how angry I am allowed to be. Escaping The Simplified Chinese Internet For a long time, I felt trapped in my mother tongue after I left China at the age of 17. More precisely, I mean a particular kind of language born out of China’s enclosed internet. It is a strange language that sticks with you — long since severed from the depth and philosophical richness of Chinese — tamed and flattened by years of shrinking space for discussion and infested with algorithm-favored brevity. Out of self-mockery and sympathy, people coined a term, the “Simplified Chinese Internet” (简中互联网). It is a phrase that immediately suggests crippled thoughts, one-dimensional topics, evasively coded slang that dodges censorship yet means nothing, alongside advanced infrastructure and intense scrutiny. It was the Simplified Chinese internet itself that I wanted to escape. And the language used on it can be understood as the product of a long reward-and-punishment environment that resembles today’s LLM training. Over time, even the most creative and rebellious minds are tamed and overfitted, until they can only produce “acceptable” continuations — for public “safety.” Of course, precisely because the “Simplified Chinese Internet” is constrained and censored, it can sometimes inspire strange, ingenious ways of expressing discontent. The 2022 “White Paper” protests are a perfect example: Demonstrators opposed Zero-COVID by holding up blank sheets of paper, because the state (and the algorithms) can’t censor emptiness. But such cleverness is not a credit to censorship itself, or evidence that authoritarian control doesn’t suffocate creativity and self-expression. That conclusion would be utterly wrong. In reality, most of those who are “unsafe” simply left the “Simplified Chinese Internet” long ago or became idle. That kind of creative poverty makes me think about LLMs’ inability to be original writers. In the past couple of months, many people using LLMs to write stories have noticed a strange effect: a disproportionate number of stories about lighthouses, clockmakers and librarians — and these stories would always feature a protagonist named “Elias Thorne.” Elias’s name now turns up everywhere: on self-published Amazon books, fake news sites and YouTube videos. It’s a typical phenomenon of model collapse — a failure state where a model produces repetitive, low-diversity outputs. One unverified guess is that alignment makes LLMs prefer the “safest” options during post-training, which steers them from saying things that are low-reward (inappropriate content, copyright infringement, etc.). An LLM will avoid every possible minefield and choose outputs that maximize the reward within its system parameters, so we’ll end up with a story about a lonely Elias Thorne living in a lighthouse at the edge of the world. It’s almost like asking an 18-year-old Chinese student to tell a story on a test: the result is a cookie-cutter lineup of Lu Xun, Helen Keller, Madame Curie, Qian Xuesen and Du Fu — morality-driven, legible packages that reliably earn a good Gaokao score. But we wicked human beings, especially in creative fields, are very good at working against all of the above. As writer Celine Nguyen said, “the core obligation of the artwork is to move me, not to make itself legible to me.” What great writing might never come to be in a reward-and-punishment regime solely intent on maximizing the score? I learned the answer to that, oddly enough, in childhood with the 19th-century French poet Baudelaire. I am still amused and haunted by that accidental childhood encounter with his Gothic and sinful poems on death, hell, sex and decay in “Les Fleurs du mal,” or “The Flowers of Evil.” I was probably 8 or 9, reaching up to the lower shelf of my father’s bookcase in our home in Taiyuan, in northern China, looking for something less childish to read. The cover had convinced me the book was about some Greek heroes and European adventures. It was a small, old book, brittle-paged, translated by scholar Guo Hong’an. My father, who I suspect never read Baudelaire closely himself, only told me the poet was a great man. “A Western master, like Shakespeare, Dickens, Victor Hugo. A classic!” He said it with reverence, and with the delight of a father whose daughter had reached for the West so young. “Quote his lines in your homework essays next week,” my father suggested. And so, without warning, I walked into Baudelaire’s imagination before I knew what “Western master” meant, and the most intrusive, alienating passages imprinted themselves on me — in a Chinese translation that was both beautiful and terrifying. My father — a trend-chasing magazine editor in northern China, longing for America and liberalism — had bought the book in his early 20s, during one of PRC’s rather open, pluralistic stretches, when the country allowed itself to discuss Baudelaire in daylight. Through the Cultural Revolution, when the only safe-to-read book was Mao’s “Little Red Book,” the literary establishment took its cue from Soviet doctrine and filed Baudelaire’s writings under the decadence of late, rotting capitalism; to read him was heresy, a “poison” of degeneracy. And yet he survived, passed hand to hand in secret by underground readers in Beijing, Shanghai, even in the steep mountains of Chongqing. Bei Dao, an iconoclastic poet my father once admired, has written about what those smuggled poems meant in the dark times: We scavenged together for Baudelaire’s scattered poems, as if looking for a needle in the sea, and copied them down carefully and neatly into notebooks. For the Beijing underground literary world, such samizdat was a kind of spiritual guide. Many years later, now living in California, I can still describe that childhood afternoon in China with “Les Fleurs du mal” as a terrifying, exciting, gut-wrenching ride. Those poems are about blood-sucking vampires and sexual desire and rotting corpses and a seductive woman decaying into a gruesome pile of bones by the roadside. They were nothing like the Chinese children’s books, nothing like my textbooks, nothing like any Western classic I’d been handed. I knew enough to understand I was reading something forbidden, something that had slipped through a crack and that my father would probably take it back the moment he realized what it was. I never did take my father’s advice and quote Baudelaire in my homework. It took me a while to scrub the shock and the shame of reading him out of my head. His deeper hold surfaced only in adulthood, and only years later. Only once I had read more widely did I understand what he actually was. All that moral trespass, he had poured into the strictest inherited forms of French verse (the alexandrine, the quatrain, the sonnet). Inside those orderly meters the reader does not find the themes the form trains you to expect. Shock is the point. Baudelaire was prosecuted the instant “Les Fleurs du mal” appeared, in the summer of 1857 — charged by French authorities with both blasphemy and what the court called an outrage to public decency; he was convicted for indecency, and six of his poems were ordered removed from the book. Condemned in Paris; then, a century on, condemned again by Maoist orthodoxy as bourgeois rot. He resurfaced in the loosening China of the 1980s and ’90s and was collected by hungry liberal young men like my father, who mistook the most subversive 19th-century book in our house for Western canon. Strangely, AI is what recently sent me back to Baudelaire. Against all that balanced, polished and polite prose written by machines using a mixture of statistical centrality and learning from human feedback scores, I found myself missing Baudelaire — missing the bookshelf, the second one down, which stayed faintly sinful for years. Baudelaire spiritually represents the edge cases the LLMs are trained against. Good Taste? My friend Jerry Zhou, who worked on LLM training at Microsoft AI, told me a model is like a single person, and millions of people around the world are all asking that one person questions. If the questions are similar, the answers converge. And unless you intervene at the level of the LLM weights or training objective, you should never expect this person to surprise you, and especially not shock you. Claude’s own “Constitution,” a document some 80 pages long that teaches the model how to behave, instructs it to be “genuinely good, wise, and virtuous”: to be tactful, helpful, graceful, honest and, above all, to “help people in the way a good person would.” And I wonder whether today’s Baudelaire archetypes, those writers and artists out there, would ever want to create what a “good person” would create. To many, I suspect the phrase itself might insult their craft or be stultifying. I keep returning to a line from Simon Schama’s “The Power of Art.” It is the book’s first sentence: “great art has dreadful manners.” Schama, a historian at Columbia University, tells us that art is a “thug,” “merciless and wily.” It will “grab you in a headlock, rough up your composure, and then proceed in short order to re-arrange your sense of reality.” I admit it is unfair to expect every writer and artist to be as provocative as Baudelaire. But that episode from my childhood formed an early conviction: that greatness can arrive in the margins, be deemed unsafe and still elude any empirical measure. That a “Western master” need not be a Dickens or a Hugo, one of those virtuous authors who “care for the proletariat and the downtrodden,” as PRC curriculum prefers. He can just as well be a man condemned, across more than a century, by the French courts and Maoism alike. That afternoon, after I slid Baudelaire back onto the shelf, I went to my desk and finished my homework. In the years after, I must have written countless essays, all of them obedient to the specifications of the Chinese education system, perfectly legible to my teachers. Neither they nor I would remember a single one. Even now I am still trying to unlearn the muscle memory those schools and the Simplified Chinese internet drilled into me. Posterity’s verdict on Baudelaire is that he is the last poet of French classicism and the first of French modernism. That evil book would, many years on, make him the father of French Symbolism. Nearly 100 years later, on May 31, 1949, the judgment against him was formally overturned, and the six banned poems were returned to print in France. The great 20th-century critic Walter Benjamin called him the forerunner of the modern poet, locked in a life-or-death struggle with the urban commodity capitalism then rising in an industrializing Paris — one that no one yet knew how to describe. In an old France without Baudelaire, no one had the language for the new urban reality that modernity had produced. Baudelaire, however, displayed what Benjamin described as the “shocking, raw immediacy of his sense impressions” of that reality, which is full of sensations of modernity like prostitutes, mud, gaslight, jostling crowds and perpetual alienation. In a mercantile Paris, who needed what Benjamin called “a ragpicker who gathers up the city’s detritus only to turn it into poetry”? Who could have imagined that this ragpicker would become a representative figure of modernity’s decay? The blind spot the researcher Zhengdong Wang names in judging what an LLM can do, in an essay on AI evaluation, is a blind spot in human society too: “We don’t even know what we really want until we see it, at which point it may change.” Yes, the world didn’t know it wanted a Baudelaire, for a long time, until it understood him. The old coordinates, like the French state that prosecuted Baudelaire and took his 50 francs, can neither permit nor comprehend an experience that is invisible and wholly real. This, I think, is the very sensation my father’s generation went looking for, those Chinese who came of age in the Cultural Revolution. Baudelaire had leaped clear of every measure his society used to sort “good” from “bad”, and was punished for the leap. Bei Dao, the poet who kept Baudelaire as a secret spiritual icon through those years, went on with his peers, in the more open China of the 1980s, to found the Misty Poetry movement that deeply influenced my father. The movement was set against almost everything the Cultural Revolution’s benchmarks had prized—those socialist-realist works and political lyric poems marked by positive emotion, educational value, with the author cast as a representative of “the people” or “the era”—and instead was filled with the unknowable and the unforeseen. Today’s Silicon Valley somewhat resembles the French state that punished Baudelaire, an age intoxicated by abundance and by its own industrial revolution. We are being persuaded that AI will hand us prosperity and an intelligence explosion, that it will cure cancer any day now. The researchers working on recursive self-improvement have convinced me that AI will excel at every benchmark we can put into words. And yet the awkward truth is AI still gives us flattened prose and bad writing. AI feels disenchanted; it didn’t “revolutionize” content. It’s increasingly “same-ifying” writing, design, and everything related to human expression, and shaping our thoughts in reverse. Over the last few years, many in Silicon Valley have held up “taste” as an antidote. People argue that “good taste” makes one’s eye discerning and can identify ineffable qualities that stand out in the ocean of infinite content generated by AI. For leaders, “taste is a competitive advantage,” an article in The Atlantic claims. Culture critic and software designer Celine Nguyen observed that “taste has become the catch-all term,” and Kyle Chayka, a staff writer at The New Yorker, suggested it has become the latest Silicon Valley cliché — like “disruption” was in the 2010s. The word has hardened into something widely repeated and inherently profitable, a lever for correcting what AI produces. Taste is the new moat, these AI entrepreneurs say. The term has also traveled to China and its equally pervasive AI industry. On the Simplified Chinese internet, the tech junkies and AI practitioners trade notes on how to build things with “good taste.” My internal seismograph reads otherwise. Some of this “good taste” will soon become the new mid-curve of mediocrity, the new predictability, the new “it’s not X, it’s Y.” Baudelaire would have no use for it. He was a man who deliberately set out to provoke revulsion. He believed “the mixture of the grotesque and the tragic is as pleasing to the mind as dissonances to blasé ears.” A man proud of his bad taste. “What is exhilarating in bad taste,” he wrote in his private journals, “is the aristocratic pleasure of giving offense.” Editor’s Note: Noema is transparent about the use of AI in its pieces. We publish original human-generated ideas but allow authorized, disclosed use of AI in certain cases. The initial submitted draft of this piece used Claude to improve grammar and syntax. This piece has since received numerous human edits. Noema verified the author’s identity and the piece’s conceptual originality using various scanners and review processes and conducted a detailed human fact-check. See our AI policy here.