Glamour Lessons (2)
Into the hall of mirrors that is the realm of literary writing produced by LLMs
This post is the second of five on some literary discourses I’ve been mulling over, and making correspondences between, throughout the winter of 2025-26.
You can read the first post in the series here. For the conceptual grounding that makes subsequent posts meaningful, you’ll need to begin at the beginning.
By coincidence, twenty-four hours after Brady Brickner-Wood’s article on performative reading appeared online, The New York Times published something else that seemed to be in conversation with it. This was an essay by Sam Kriss on the stylistic quirks of LLMs like ChatGPT, Claude, Gemini, and so on. It’s easy to sense that the prose these things produce isn’t quite human enough to pass for human. What’s harder is to say why that’s the case, to really pinpoint the sources of the LLMs’ oddities. Kriss was up to the task, though, and set about subjecting chatbot outputs to keen rhetorical analysis, to spotlight the most recognisable tics and account for their predominance.
The result, to my mind, is a barnstormer of a polemic; it’s as hilarious as it is despair-inducing. Substantively, Kriss makes three main moves. First he identifies the root problem with LLMs, then he surveys the knock-on effects of developers’ efforts to compensate for this problem. Finally, he takes the measure of popular attitudes to these effects.
Fundamentally, Kriss argues, LLMs are prone to “overfitting” when generating text in response to user prompts. Overfitting amounts to operating according to the logic of a skewed syllogism: the LLM takes the inclusion of a keyword in the prompt as premise x, then extrapolates a second premise y from its own dataset-derived associations with x, and finally generates an output z that tries to respond to x by doubling down on y. Blatant absurdities were the inevitable byproduct of this associative logic in the earliest publicly-accessible versions of ChatGPT, as Kriss illustrates with a particularly memorable example. “One of the [initial] tasks I gave the machine,” he says,
was to write a screenplay for a classic episode of The Simpsons. I wanted to see if it could be funny; it could not. (Still can’t.) So I specified: I wanted an extremely funny episode of The Simpsons, with lots of jokes. It did not deliver jokes. Instead, its screenplay consisted of the Simpsons tickling one another. First Homer tickles Bart, and Bart laughs, and then Bart tickles Lisa, and Lisa laughs, and then Lisa tickles Marge. …
It’s not hard to work out what probably happened here. Somewhere in its web of associations, the machine had made a connection: jokes are what make people laugh, tickling makes people laugh, therefore talking about tickling is the equivalent of telling a joke.
“Tickling the Simpsons” therefore serves as Kriss’ term for the “overfitting” tendencies of LLMs. And although Kriss does concede that refinements in the design of ChatGPT now allow it to avoid the simplistic overfitting of those early days, he goes on to suggest that this isn’t evidence of the elimination of overfitting. Rather, he sees overfitting as a persistent problem, albeit one that today shows up in more varied and sophisticated forms. Since “the same basic structure governs essentially everything [LLMs] write,” their current absurdities “make sense when you understand that [their structure is] constantly tickling the Simpsons.”
This is the second move in Kriss’ polemic, and he defends it by casting an eye over the LLM-generated prose that displays pretensions to eloquence. This kind of prose has become notorious for, among other things, the frequency with which it employs the emdash. “You used to find [the emdash] only in self-consciously literary prose,” Kriss writes. “Not anymore.” The emdash is the supposedly highbrow equivalent of tickling the Simpsons:
Within the A.I.’s training data, the em dash is more likely to appear in texts that have been marked as well-formed, high-quality prose. A.I. works by statistics. If this punctuation mark appears with increased frequency in high-quality writing, then one way to produce your own high-quality writing is to absolutely drench it with the punctuation mark in question.
“The A.I. is trying to write well,” Kriss adds. “It knows that good writing involves subtlety,” that “good writing is complex,” that “good writing takes you on a journey.” It knows, also, that subtlety correlates with quietude and implication, that complexity correlates with craftwork of a textile variety, and that journeying correlates with pathfinding and exploration: so, as well as favouring the emdash, it favours language that evokes shadows and hushed voices, tapestries and the intricacies of weaving, thresholds to adventure and the invitation to “delve in.” “All of this contributes to the very particular tone of A.I.-generated text,” Kriss writes,
always slightly wide-eyed, overeager, insipid but also on the verge of some kind of hysteria. … A.I. does still try to work sensory language into its writing, presumably because it correlates with good prose. But without any anchor in the real world, all of its sensory language ends up getting attached to the immaterial.
The common result? “Mixed metaphors and empty sincerity. Impersonal and overwrought.” And although Kriss is obviously playing for laughs when he mimics the voice of an LLM—“We are unearthing the echo of loneliness. We are unfolding the brushstrokes of regret. ... We are weaving a coffee outlet into our daily rhythm”—the sentence I’ve plucked from the space those ellipses now occupy seems to me perfectly true, perfectly on-point:
We are saying the words that mean meaning.
Kriss’ third and final move in his polemic, and by far the most dispiriting, follows on from this observation. It’s the contention that many, many people today prompt LLM-generated text outputs and claim the results as their own. These people are content to speak words that simply “mean meaning,” executing speech acts that are vapid in themselves but ape meaning (or, more properly, meaningfulness) by playing on associations with what meaningful rhetoric sounds like.
In short, many people, while wanting to write literarily, are outsourcing the task to mechanised systems that derive their notions of “the literary” from literariness: not from the textual nuances of individual works of literature, but from qualitative generalities about literature that are rooted in the designated status (high/low) of a range of works. And yet, of course, the very outsourcing of the task puts the lie to the sincerity of the want. We’re entering a hall of mirrors here. Those who outsource their writing to LLMs don’t in fact want to write literarily: they want to be seen as having written literarily—to perform it—in order to wear the glamour of literariness, which, in its production by LLMs, issues only from a prior designation of literariness.
And, for Kriss, what’s worse than this sort of ersatz literary activity is the fact that, culturally, a critical mass of people seems prepared to accept the output as instances of meaningful expression. “Every day,” he writes, “another major corporation or elected official or distant family member is choosing to speak to you in [the LLMs’] particular voice. This is just what the world sounds like now. This is how everything has chosen to speak.” And, he adds,
A lot of people don’t seem to mind this. … Researchers found that most people vastly prefer A.I.-generated poetry to the actual works of Shakespeare, T.S. Eliot and Emily Dickinson. It’s more beautiful. It’s more emotive. It’s more likely to mention deep, touching things, like quietness or echoes. It’s more of what poetry ought to be.
Surely, though, “a lot of people” doesn’t include identifiably literary people. I mean people of a genuinely literary disposition, with genuinely literary interests, who find value in actually reading literature written by authors with literary bonafides. These people are the sort who cherish the inwardness of the activity of reading and the privacy of the experience. They’re the engaged, demanding readers that LLMs can’t keep up with, the ones who can see the LLMs’ associative logic for the junk it is and who’ll bin the output. Right?