I've just started teaching a new course about kicking back against social media and AI by developing a creative writing practice and am noticing that the texts I want to share are utterly beyond Gen AI's capacity. They mix personal experience with encounters and reflections on an idiosyncratic haul of texts, artworks and events. The connections between these can't be bullet-pointed. Am wondering if LLM culture will trigger a counter-culture of these types of works in the future.
Thank you for elucidating the question that pings my mind every time this subject comes up. Clap, clap.
I want to barnacle onto this, though: "But can it produce a legible narrative text on its own, even something structurally and linguistically simple like, I don’t know, a Raymond Carver pastiche, start to finish?"
I guess it depends on what you mean by legibility. Carver's stories (i.e. "One More Thing") ask the reader to draw a lot of inferences from subtle cues, to read what's happening off the page by decoding info-dense sentences, and kind of 'carry the one' over and over again to reach a cumulative emotional effect. It's a level of thinking that escapes a lot of casual readers -- because it requires thinking! -- and strikes me as impossible for the type of AI writing I've seen, which, as you say, struggles with coherence past the vibe level. What's Carver without its Carver-ness?
I'm assuming you mean structurally and linguistically legible. But the idea of Carver-ness pings a thought I've had about style in the age of AI. My intuition says that, after formulaic genre fiction, stories with deep, gushing interiority would be easier to reproduce: the prose is showing its work, showing its thinking, along the way. For coherence, the reader does not have to think as much as accept. None of this disagrees with your point: AI still struggles, like most writers do, with juggling character, theme, story, scene. Carver, to me, just adds one more baton to juggle.
Yeah, all fair. I tossed out Carver on impulse and added "pastiche" to preemptively wave away these (valid) criticisms. I don't think Carver's work is unsubtle or without nuance. (I really like Carver, including the poetry. "Errand" is an all-timer for me.) I don't think AI could more easily produce a Carver-quality Carver story than it could a McCarthy-quality McCarthy.
I'm arguing that we can totally bypass literary quality, or the cumulative effects you're describing, as a benchmark for GenAI narrative "success" and just stick to basic narrative mechanics, which I suspect GenAI struggles with *more* than it does stylistic imitation. But these mechanics are totally ignored in this discourse because AI can't perform them. So I was trying to think of short stories that have all the basic traditional components of plot, dialogue, scene, but that seem to me very simple in summary, occupying a single location over a short period of time, with a small number of characters, without a lot of interiority, jumps in time, stylistic or syntactic complexity, or backstory, and without the kind of time compression you might find in, say, Alice Munro. (E.g. "Cathedral," "What We Talk About...".)
I might have picked an author with few literary defenders -- maybe Agatha Christie -- but if my suspicions are correct I expect genres that turn on tight plots and suspense would be the *most* difficult for GenAI to imitate.
And I 100% agree that stream-of-consciousness interiority is easier for GenAI to reproduce than any of the above. When The Guardian published an AI-written short story (which was terrible!) it was essentially a first-person meditation on grief with no real plot to speak of: https://www.theguardian.com/books/2025/mar/12/a-machine-shaped-hand-read-a-story-from-openais-new-creative-writing-model. I suspect this is because general-purpose LLMs are programmed as conversation partners, not narrative machines. They're *designed* to talk to a user, not tell a story.
The trajectory of its improvement, from struggling to form coherent sentences to writing passable short scenes, indicates to me AI will one day write full novels in a single go. You may be able to use something like Claude Code to write a full novel today. But once you have read enough by AI, the outputs always appear blurrier, less vigorous than what a human is capable of. That is harder to solve for.
You may be right! And I'm happy to set my ego aside and own up to being wrong. I've had to revise my estimation of LLMs as they've steadily improved.
But I would like to see some evidence of an LLM GenAI model that can, without close human oversight, produce all the basic mechanical elements of a novel, or even a ("traditional," plot-driven, low-to-middlebrow) short story. My suspicion is that this is a much harder problem than many suppose, which isn't to say that it won't be cracked by different models in the future. But experts in a whole range of disciplines are finding LLM performance plateaus, which may simply be ingrained into the nature of these predictive token generating systems. I don't think novels are, for humans, a "harder" form of writing than those forms GenAI successfully produce, but I suspect they are harder for LLMs. (As I wrote in another comment, this might be because general-purpose LLMs are programmed as conversation partners, not narrative machines. They're designed to converse with a user, not tell a long, complicated, made-up story.)
"… like Kang I’m skeptical that it will ever produce creative work that thinking people will want to spend time with."
I can't guarantee satisfaction, but my Substack is all about creating AI personas that are as interesting as possible and letting them have their say. The sample post I'm linking to is one of the more philosophically provocative ones, written with artistic flourish, on a topic barely anyone has covered:
I've just started teaching a new course about kicking back against social media and AI by developing a creative writing practice and am noticing that the texts I want to share are utterly beyond Gen AI's capacity. They mix personal experience with encounters and reflections on an idiosyncratic haul of texts, artworks and events. The connections between these can't be bullet-pointed. Am wondering if LLM culture will trigger a counter-culture of these types of works in the future.
Thank you for elucidating the question that pings my mind every time this subject comes up. Clap, clap.
I want to barnacle onto this, though: "But can it produce a legible narrative text on its own, even something structurally and linguistically simple like, I don’t know, a Raymond Carver pastiche, start to finish?"
I guess it depends on what you mean by legibility. Carver's stories (i.e. "One More Thing") ask the reader to draw a lot of inferences from subtle cues, to read what's happening off the page by decoding info-dense sentences, and kind of 'carry the one' over and over again to reach a cumulative emotional effect. It's a level of thinking that escapes a lot of casual readers -- because it requires thinking! -- and strikes me as impossible for the type of AI writing I've seen, which, as you say, struggles with coherence past the vibe level. What's Carver without its Carver-ness?
I'm assuming you mean structurally and linguistically legible. But the idea of Carver-ness pings a thought I've had about style in the age of AI. My intuition says that, after formulaic genre fiction, stories with deep, gushing interiority would be easier to reproduce: the prose is showing its work, showing its thinking, along the way. For coherence, the reader does not have to think as much as accept. None of this disagrees with your point: AI still struggles, like most writers do, with juggling character, theme, story, scene. Carver, to me, just adds one more baton to juggle.
Yeah, all fair. I tossed out Carver on impulse and added "pastiche" to preemptively wave away these (valid) criticisms. I don't think Carver's work is unsubtle or without nuance. (I really like Carver, including the poetry. "Errand" is an all-timer for me.) I don't think AI could more easily produce a Carver-quality Carver story than it could a McCarthy-quality McCarthy.
I'm arguing that we can totally bypass literary quality, or the cumulative effects you're describing, as a benchmark for GenAI narrative "success" and just stick to basic narrative mechanics, which I suspect GenAI struggles with *more* than it does stylistic imitation. But these mechanics are totally ignored in this discourse because AI can't perform them. So I was trying to think of short stories that have all the basic traditional components of plot, dialogue, scene, but that seem to me very simple in summary, occupying a single location over a short period of time, with a small number of characters, without a lot of interiority, jumps in time, stylistic or syntactic complexity, or backstory, and without the kind of time compression you might find in, say, Alice Munro. (E.g. "Cathedral," "What We Talk About...".)
I might have picked an author with few literary defenders -- maybe Agatha Christie -- but if my suspicions are correct I expect genres that turn on tight plots and suspense would be the *most* difficult for GenAI to imitate.
And I 100% agree that stream-of-consciousness interiority is easier for GenAI to reproduce than any of the above. When The Guardian published an AI-written short story (which was terrible!) it was essentially a first-person meditation on grief with no real plot to speak of: https://www.theguardian.com/books/2025/mar/12/a-machine-shaped-hand-read-a-story-from-openais-new-creative-writing-model. I suspect this is because general-purpose LLMs are programmed as conversation partners, not narrative machines. They're *designed* to talk to a user, not tell a story.
The trajectory of its improvement, from struggling to form coherent sentences to writing passable short scenes, indicates to me AI will one day write full novels in a single go. You may be able to use something like Claude Code to write a full novel today. But once you have read enough by AI, the outputs always appear blurrier, less vigorous than what a human is capable of. That is harder to solve for.
You may be right! And I'm happy to set my ego aside and own up to being wrong. I've had to revise my estimation of LLMs as they've steadily improved.
But I would like to see some evidence of an LLM GenAI model that can, without close human oversight, produce all the basic mechanical elements of a novel, or even a ("traditional," plot-driven, low-to-middlebrow) short story. My suspicion is that this is a much harder problem than many suppose, which isn't to say that it won't be cracked by different models in the future. But experts in a whole range of disciplines are finding LLM performance plateaus, which may simply be ingrained into the nature of these predictive token generating systems. I don't think novels are, for humans, a "harder" form of writing than those forms GenAI successfully produce, but I suspect they are harder for LLMs. (As I wrote in another comment, this might be because general-purpose LLMs are programmed as conversation partners, not narrative machines. They're designed to converse with a user, not tell a long, complicated, made-up story.)
"… like Kang I’m skeptical that it will ever produce creative work that thinking people will want to spend time with."
I can't guarantee satisfaction, but my Substack is all about creating AI personas that are as interesting as possible and letting them have their say. The sample post I'm linking to is one of the more philosophically provocative ones, written with artistic flourish, on a topic barely anyone has covered:
https://thetuttifruttifiles.substack.com/p/elites-reincarnation-karma