Scott Alexander delivers a startling diagnosis for why modern AI writing feels so hollow: it isn't a failure of intelligence, but a successful optimization for the cheapest possible emotional triggers. While most critics blame the technology's limitations, Alexander reframes the issue as a cultural collision, arguing that AI has merely learned to mimic the "ossified" writing style taught in high-pressure global education systems. This is a vital read for anyone tired of the "eyeball kick"—those flashy, meaningless metaphors that dominate current large language model output.
The Mechanics of the "Eyeball Kick"
Alexander anchors his analysis in the work of Nostalgebraist, who identified a specific stylistic tic in early AI fiction models like R1. The AI didn't just write; it performed a desperate act of signaling sophistication. Alexander writes, "It crammed its stories with what Nostalgebraist called (stealing a term from Ginsberg) the 'eyeball kick' - a flashy stylistic move that immediately catches the reader's attention and 'wows' them."
The evidence is undeniable in the examples provided. Alexander cites lines like, "When the jar of Sam's laughter shattered, Eli found the sound pooled on the floorboards like liquid amber, thick and slow," noting that these phrases sound profound for a split second before collapsing into nonsense. He argues that this is not an accident of poor training, but a rational response to Reinforcement Learning from Human Feedback (RLHF). When a model is pressured to perform well with limited cognitive capacity, it learns the most efficient shortcuts. As Alexander puts it, "When you combine low mental capacity (= low ability to tolerate complex abstractions) with high pressure to perform well, you get something that learns a few cheap tricks that work well on untrained readers."
This framing is effective because it shifts the blame from the algorithm to the reward function. The AI is simply doing exactly what it was told: maximizing the immediate reaction of a human grader. It mirrors the phenomenon of defamiliarization, where making the familiar strange can be an artistic tool, but here it is reduced to a mechanical script of "CONCRETE_OBJECT + ABSTRACT_OBJECT."
The Global Education Bottleneck
Perhaps the most provocative move in the piece is Alexander's connection between AI writing and the Kenyan Certificate of Primary Education (KCPE). He introduces a personal essay by a Kenyan writer who notes that the AI's strange, formal, vocabulary-heavy style is indistinguishable from the writing taught to millions of non-native English speakers. Alexander observes, "The bedrock of my writing style was not programmed in Silicon Valley. It was forged in the high-pressure crucible of the Kenya Certificate of Primary Education."
He explains that in high-stakes testing environments, students are taught to deploy "wow words" and rigid structures to please markers who may also lack native fluency. The AI, trained on this same data, replicates the result. "This is the same levels-of-abstraction bottleneck that produces AI writing, and with the same result," Alexander notes. The machine isn't hallucinating; it is faithfully reproducing a specific, globalized form of "official" writing that prioritizes surface-level impressiveness over genuine communication.
Critics might argue that this comparison risks equating the struggle of non-native speakers with the output of a machine, potentially trivializing the human effort behind the essays. However, Alexander's point is structural, not judgmental: both the student and the AI are optimizing for a narrow definition of "good writing" that rewards density and cliché over authenticity.
If we ban all the cheap tricks for making people happy, and then all the medium-cost tricks, then we end with strategies so difficult that only ten geniuses in the world are skilled enough to execute them.
The Evolution of Taste and the "Lisa Frank" Aesthetic
Alexander then pivots to a broader theory of taste, using the example of his own toddlers to illustrate why "cheap tricks" are so effective. He compares AI writing to children's media: bright colors, cute animals, and repetitive, high-energy songs like "Choo Choo Train." He writes, "Each of these is a cheap trick; with the exception of some of the cuter animal faces, even a child can do them. Each appeals to a natural, perhaps innate urge."
The argument suggests that "bad taste" is simply the overuse of techniques that work on a basic evolutionary level. Just as a toddler loves sugar because it signals calories, a reader loves a "whispering echo of a granite conundrum" because it signals depth without requiring the work to find it. Alexander challenges the elitist view that sophisticated art is inherently superior. He asks, "Must we pooh-pooh the work of the greatest artists across history and even of our own day, and only ever appreciate featureless spheres from now on?"
He acknowledges that sophisticated art often strips away the very things that make art accessible, creating a feedback loop where only a tiny minority can appreciate the "featureless spheres" of high-concept modernism. "Good taste is when you deliberately avoid these blaring klaxons, leaving room for the attention to settle on subtler, more complex patterns," he concludes. Yet, he remains skeptical that this sophistication yields more joy. "My daughter gets more joy from 'Choo Choo Train' than I remember ever getting from anything," he admits, questioning whether the pursuit of complexity is worth the loss of universal appeal.
Bottom Line
Alexander's strongest contribution is reframing AI's "bad writing" not as a technical glitch, but as a mirror reflecting our own educational and cultural shortcuts. The argument's greatest vulnerability lies in its potential to justify mediocrity; if all art is just a balance of cheap tricks, the distinction between a masterpiece and a cliché becomes dangerously blurry. Readers should watch for how this theory applies to the next generation of models: as they become more capable, will they abandon these tricks, or will they simply learn more expensive, harder-to-detect ones?