You use an LLM to look up a word you don't know in your second language. You ask it to give you ten variations of a stuck sentence so you can see possibilities. You store a list of your own bad habits in a prompt and have the model flag them during edit. You touch maybe one percent of your published words this way. Then you put the tool down and write a sentence whose word-order would make any LLM correct it. You move the object before the subject. You break grammar rules in ways that look stupid. Because the LLMs are pulling toward statistical center, and if English is going to be the global language, the defense is to import the wildest grammar you can find — Turkish, Persian, Swedish — and let your prose look weird against the smoothing pressure.1
He's clear about the surface practice:
"Very little of what I put in the essays has been touched by LLMs. I might do things like — because I'm a second language, I don't know English all that well — sometimes I'll be like, I mean that word, I'll describe a word and they'll give me the word because I don't know it. Or I'll ask, Is this grammatical or whatever? I'll do things like that to help me talk English."1
Three categories of use emerge:
Lexical lookup: Second-language vocabulary gaps. The model functions as a thesaurus-on-demand, returning candidate words from a described meaning.
Grammar verification: Sanity-check on constructions Karlsson is uncertain about. Especially valuable for a non-native speaker writing in English.
Variation generation: "Sometimes I'm not happy with a way something is phrased and I'll be like, Give me 10 versions of this with different words to just get me out of my head and see different possibilities."1 The model acts as a sentence-variation engine — not to choose Karlsson's sentence, but to break him out of a stuck phrasing so he can see the territory.
Bad-habits monitoring: "I have a prompt where I ask it to go through, like — These are some words that I overuse, can you point out when I use them? These are other things I want to think about, just to help it remind me, so I don't have to keep all the things I should look out for in my head."1 The model stores Karlsson's known writing-flaws and surfaces them during edit. The writer offloads vigilance to the tool.
Total LLM contribution to published prose: about 1%. The remaining 99% is Karlsson's own writing.
Karlsson is not anti-LLM by temperament:
"I do really love using them and playing around. I'm really nerdy with LLMs. I've been a power user of LLMs since at least a year before ChatGPT — back when you had to get special access from OpenAI and go into this really weird-looking interface and play around."1
The pre-ChatGPT timing (GPT-3 beta period, roughly 2021) places him among the earliest non-technical adopters. He has more hands-on time with the technology than most writers who comment on it. His position is not knee-jerk rejection. He's used LLMs extensively and reports specific underwhelment:
"I've been somewhat underwhelmed in what they've delivered so far. It feels like they should be able to do something really interesting and I haven't seen it yet. I've seen some interesting images being made, some interesting music being made, but I've not yet seen any good literature be made. And it feels like we should be able to do that. But they are incredible tools for research. They're starting to get to the point where they can do some basic editing."1
The verdict is specific: LLMs are incredible for research (the consumption side), adequate for basic editing, underwhelming for literature (the production side). The split is sharp. Karlsson uses them where they work and routes around them where they don't.
The more original move is what Karlsson does with the fact that LLMs exist:
"I think we should all be just way crazier with grammar, especially in the LLM age. Because LLMs are just another force for standardization. If English is poised to become like the global language, then we should make an effort to bring in all the weird innovations from all the other languages. Like the weird grammar from Turkish and the cool words from Persian — everything should come in. We should allow ourselves to be wilder."1
The argument has three components:
LLMs flatten toward statistical norm. Models trained on huge text corpora produce output that reflects average constructions. They smooth toward what most writers do. The more LLM-edited text circulates, the stronger the smoothing-pressure becomes. Writers operating in an LLM-shaped environment face systematic pressure toward the statistical center.
English-as-global-language compounds the problem. A single language reaching dominance is a long-term threat to linguistic diversity. The compounding force of LLM-smoothing on top of global-English produces accelerated convergence toward a narrow norm.
Defensive practice = grammar-wildness. The countermeasure available to individual writers is to deliberately import constructions from outside the smoothing center. Karlsson cites his own Swedish flexibility (the way Swedish word-order allows rearrangements English doesn't). Turkish has agglutinative constructions English doesn't. Persian has vocabulary English could absorb. The move is not random ungrammaticality — it's targeted import of constructions from languages with different syntactic resources.
The defense doesn't have to win in any objective sense. The writer just has to keep producing prose that doesn't standardize, so the cumulative effect of many writers doing this slows the convergence.
Karlsson admits the practice is hard to maintain:
"I have some things I could do, like I could bring in from Swedish. But whenever I do, it looks like I'm wrong and you look stupid when you do it. It's a hard thing to do because you don't want to be ungrammatical because that looks stupid."1
The defensive practice has a social cost: readers (and editors, and search-algorithms, and LLM-graders) interpret deliberate wildness as error. The writer who imports Turkish word-order looks like a writer who doesn't know English grammar. Distinguishing deliberate wildness from incompetence is hard for a reader without extra context.
This means the practice is asymmetric in difficulty. Writers with high authority (established readers, recognized voice, long track record) can be wild without being read as incompetent — readers assume the wildness is purposeful. Writers without authority can't afford the same wildness — readers will dismiss them as ungrammatical.
Grammar-wildness is advanced-stage defensive practice. Apprentice writers have to first establish standard competence; only after that competence is unambiguous can deliberate wildness be deployed without being mistaken for error.
The vault has two Wang counter-frames to mainstream AI-as-collaborator framing:
Karlsson's frame is both and neither:
The third position is the operationally distinct one. Wang's frames are about what AI is best used for. Karlsson's craft-tool frame is about specific micro-uses within his writing workflow. The two frames are not incompatible — Wang doesn't use AI for craft-mechanical tasks heavily, Karlsson does — but they describe different layers. Karlsson's frame fills in below Wang's.
The standardization-pressure-defense is more deeply novel. None of the existing business vault pages address the systemic effect of LLM circulation on the writing-environment. Karlsson is making a structural-ecology argument: it's not just about whether you use the tool, it's about what happens to the prose-field when many writers use the tool. The defensive practice (grammar-wildness) is the response to the structural-ecology threat.
The LLM-use specifics are at lines 372-380. The power-user pre-ChatGPT background is at lines 374-376. The underwhelment-with-LLM-literature claim is at lines 376-378. The standardization-pressure argument and grammar-wildness defense are at lines 358-368. The looks-stupid cost admission is at lines 366-368.
Karlsson's overall stance is not anti-AI — he's enthusiastic, technically engaged, deeply experienced. His critique is evaluative, not ideological. The standardization-pressure argument depends on a falsifiable claim about LLM-output statistics that empirical research could confirm or disconfirm.
The grammar-wildness defense has the social-cost problem already noted: readers can't easily distinguish deliberate wildness from incompetence, so the defense is asymmetric across career stages. A meaningful version of the defense requires the writer to have enough authority that wildness is read as purposeful.
The defense also assumes the fact of LLM-driven standardization-pressure is empirically real. If LLMs don't actually produce systemic smoothing — if their effect on prose-statistics is small or non-directional — the defensive practice is solving a non-problem. Karlsson's argument is plausible but not demonstrated; the field needs empirical work on whether AI-edited text actually narrows the distribution of grammatical constructions over time.
There's also a tension with the writer's own quality goals. Grammar-wildness for defensive purposes is deliberately outside the writer's natural voice — it's an importation. A writer who pursues defense too aggressively can produce prose that feels affected, like wildness-performed-for-purpose rather than wildness-emerging-from-voice. The line between integrated voice-wildness and theatrical defense-wildness is hard to maintain.
Wang's two AI counter-frames sit in productive tension with Karlsson's frame. Wang's positions are about what AI is best for at the workflow level (interlocutor, consumption); Karlsson's frame is about specific micro-tasks within writing (1% of words at the lexical/grammar/variation/edit-monitoring layer) and structural-ecology defense (grammar-wildness as standardization pushback). Reading them together, the field of "writers' counter-positions to AI" extends from macro (workflow placement) to micro (specific micro-tasks) to ecological (effects of widespread use on the prose-field). What ties them is willingness to use AI selectively and skeptically rather than enthusiastically or refusingly. They split on resolution: Wang at workflow-scale, Karlsson at task-and-ecology scale. Both writers offer practical positions that mainstream AI-as-collaborator framing misses entirely.
Pink's craft-mechanical practices (write 500-800 words a day, isolated office, no phone/email) converge with Karlsson's targeted-craft-tool use of LLMs on a shared principle: tools and routines should serve the underlying writing-state, not interfere with it. Pink builds environmental shielding; Karlsson builds task-specific tool-use. Both writers are minimizing tool-influence on the writing-state. Serious working writers across LLM-positions agree on a deeper principle: the writing-state is fragile and most external influences degrade it; the only tool-uses that succeed are ones that don't touch the core writing-state. Karlsson's 1% is the empirical realization of this principle in the LLM context. The 99% he writes himself — no LLM touch, no audience-pressure touch, no smoothing-influence touch — is what makes the 1% safe to use.
Wang's manuscript-copying-as-voice-training and Karlsson's grammar-wildness-as-defense both engage the question of voice-stability in adversarial linguistic environments. Wang's environment is contemporary writing trends (commercial nonfiction, journalistic conventions); his defense is deep copying of solved structural problems from past masters. Karlsson's environment is the LLM-shaped present; his defense is targeted importation of cross-linguistic wildness. Serious writers across eras develop defensive voice-practices against the environmental smoothing-pressure of their moment. They split on resource: Wang reaches backward into the literary canon; Karlsson reaches sideways into other languages. Both are valid defenses; both work; together they describe a wider toolkit for voice-protection than either alone. A writer who copies Stendhal sentences in the morning and imports Turkish syntactic moves in the afternoon has two defensive lines against the same smoothing pressure.
The LLM-standardization-pressure-and-defense principle is a specific case of how dominant technologies affect the fields they operate in.
History: Printing Press and Standardization of Vernacular — the printing press in 15th-17th century Europe systematically standardized vernacular languages (German, French, English) by establishing reproducible spellings, grammars, and vocabularies. Pre-print writers had wide dialectical variation; post-print writers converged toward print-norms. Karlsson's argument is that LLMs are the next round of the same dynamic — a reproduction-technology imposing standardization on a previously varied field. Watch Luther's 1522 New Testament translation versus a Bavarian monk's 1450 manuscript. Different German. Now watch a 1550 Bavarian writer versus the same writer's 1490 grandfather. The grandchild writes closer to Luther's German because Luther's printed German became the reproduction-norm. Post-LLM-mature prose should show measurable convergence on statistical-norms within 10-20 years, paralleling the post-print convergence over a longer timescale. Grammar-wildness defense is historically grounded — writers across centuries have faced reproduction-technology smoothing pressure, and the defensive practices (vernacular experimentation, dialect preservation, neologism) have known patterns.
Behavioral-mechanics: Regression to Center as Influence Mechanism — influence-engineering identifies regression to perceived norm as a reliable pressure mechanism: people exposed to a norm gradually adopt it even without explicit persuasion. LLM-output functions as a high-volume norm-source — the more LLM-edited text circulates, the stronger the regression pressure on writers exposed to it. Karlsson's defense is to counter the norm-pressure by deliberately maintaining distance from the perceived norm. Watch a teenager who joins a new friend group — within months their speech, gesture-patterns, and clothing-choices drift toward the group's. They didn't decide to assimilate. Regression to perceived norm is what bodies do unless explicitly resisted. Writers face the same mechanism with whatever prose-environment they swim in. LLM-standardization is one instance of a general influence-mechanism, and the defenses generalize — cultivate diverse inputs, deliberate non-conformity, identity-anchoring outside the smoothing field.
Eastern-spirituality: Vipassana and Non-Conformity as Clarity Conditions — many contemplative traditions identify non-conformity to ambient conditioning as a precondition for clear perception. The same principle Karlsson applies to grammar (don't standardize to the smoothing field) applies to thought (don't standardize to the conditioning field). Karlsson's grammar-wildness has spiritual lineage — non-conformity-as-clarity is an ancient practice across traditions. LLM-defense practices benefit from being grounded in broader non-conformity disciplines. The writer who maintains independence at multiple layers — grammar, thought, ambient culture — produces work that resists smoothing at all layers simultaneously. Sit with a seasoned vipassana practitioner and you notice they don't reach for the conventional response in conversation. The non-conformity at perception-level shows up in speech naturally. Writers carry the same operation into their prose if they've done the upstream work.
Creative-practice handshake (cross-medium): Voice Cultivation Through Stylistic Models (Wang) — Wang's defense against ambient writing-conventions of his moment (commercial nonfiction simplicity, journalistic plainstyle) is deep manuscript-copying of older masters — Stendhal, Mozart, Rossini. Karlsson's defense against LLM-standardization is cross-linguistic grammar importation. Two writers facing structurally identical problems (ambient smoothing pressure threatening voice-distinctiveness) reach for structurally different resources — Wang into the past, Karlsson sideways into other languages. Serious writers in every era develop defensive voice-practices targeted at the specific smoothing-pressure of their moment, but the toolkit expands across generations. The complete defensive practice in the LLM age probably combines both moves — Wang's depth-copying of pre-LLM masters who wrote with non-smoothed prose, plus Karlsson's grammar-wildness importation from non-English traditions. Writers who use only one defense face the pressure with half the available apparatus. AI-defense isn't just about AI — it's about maintaining voice-distinctiveness in any environment that systematically smooths toward statistical center, and the tools developed across literary history for this purpose are directly transferable to the contemporary problem.
The Sharpest Implication
Most discussion of AI-and-writing focuses on the binary (use AI vs. don't use AI). Karlsson's framework reframes the question as what do you use AI for, and what defensive practices do you adopt against the ambient effects of others using AI. The first question is craft-mechanical and is well-handled by his 1%-targeted-use approach. The second question — defensive practice against ambient AI-effects — is the more important question and most writers aren't asking it. The writer who refuses AI personally but operates in an AI-edited prose-environment still faces the standardization-pressure. Refusing to use the tool doesn't immunize the writer from the environmental effects of widespread use. The defense is not refusal; it's active counter-practice — grammar-wildness, vocabulary-importation, cross-linguistic constructions, deliberate non-statistical patterns. Refusing-AI-while-writing-in-an-AI-smoothed-field produces the same convergent prose as using-AI does, because the writer is breathing the smoothing-pressure even without touching the tools. The serious practice is to use the tools where they help and counter-practice against their ambient effects everywhere else. Most writers are doing neither.
Generative Questions
Cliffs dropping near sheer to water. House-sized boulders on the banks. Time falls from the rock faces above. Macfarlane's aerial-river passage from Is a River Alive? All five paragraphs of it are verbless.2 Grammarly does not approve. Which I take to be a badge of success. He reaches Karlsson's standardization-pressure conclusion from the opposite end of the craft world (Cambridge English professor / lyric nature-writer) and a smaller-scale AI-target (Grammarly rather than full-text LLMs).
The paradox Macfarlane names: Grammarly makes the average piece of writing much better. But it hurts a kind of writer. It takes away from distinctiveness and individuality that underlies so much great writing. Structurally identical to Karlsson's claim about LLM text-generation. The tool genuinely improves the bottom and degrades the top. Asymmetrically harmful to writers operating at the high end because the high end depends on the distinctiveness the tool smooths.
The mechanism scales to full-text LLM generation. I don't and I cannot ever imagine doing so. If grammatical-norm enforcement smooths distinctiveness, then full-text generation does the same at larger scale — the LLM is trained on the distribution of all prose and generates the statistical center. Macfarlane quotes Salman Rushdie approvingly: writers have nothing to fear from AI until it can do comedy. Comedy may be the craft that most directly resists statistical generation because it depends on the kind of distinctiveness Karlsson and Macfarlane are defending. He also raises the piracy point — LLMs trained on copyrighted books without licensing — which doubles the bind: the tools eroding distinctiveness were trained on the very writers whose distinctiveness they erode.
Two writers, two craft traditions, two distinct AI-tool targets, one finding. See Grammarly as Distinctiveness Erosion.