Readers can tell, and when they can tell they stop reading. That is Bryan Cantrill's argument in "Your intellectual fly is open", which returned to the Hacker News front page this week with more than 680 points: to anyone who has seen a modicum of LLM output, the tells — emojis, single-sentence paragraphs, "it's not just X, but also Y" constructions — are impossible to ignore, and once a reader suspects generated text they no longer know what is real, so they leave. His prescription is not abstinence. LLMs, he writes, are helpful for brainstorming, invaluable for comprehending text, and astonishingly good editors; they are simply bad at being you. The same week, a paper on arXiv by Ricard Solé, Giulio Ruffini, David Krakauer, Michael Levin and colleagues modelled LLM use as a cognitive virus spreading through a population, with a tipping point past which small increases in adoption produce abrupt losses in competence — and identified the antidote as reversibility: keeping the ability to do the thing without the tool. The workflow below is built on both: you draft, the model edits, and you keep the pen.
The tells, as a checklist
Cantrill names three. Practising editors will add more. Run a draft against all of them before publishing:
Structural tells. Emoji as bullets or section markers. A run of single-sentence paragraphs where none carries a revelation. Every section the same length. A closing paragraph that restates the opening. Lists where prose would do, especially three-item lists of parallel abstract nouns.
Syntactic tells. "It's not just X — it's Y." "Whether you're a beginner or a seasoned professional." Rhetorical questions answered in the next sentence. Colons before every list. Sentences that begin with a participle and end in a moral.
Lexical tells. A specific vocabulary readers have learned to flinch at: delve, tapestry, testament, landscape, robust, elevate, unlock, journey, game-changer, crucial, comprehensive. None of these words is wrong. Their density is the signature.
Absence tells. No specific numbers. No first-hand detail a model could not have known. No opinion that could be wrong. No sentence a lawyer would cut. The absence of anything at stake is the deepest tell, and no vocabulary filter catches it.
What the cognitive-virus paper actually says
The arXiv preprint (2609.03344, submitted 3 September 2026) does not say LLMs make you stupid. It builds a population model with three states — uncoupled users, coupled users who integrate the tool into daily practice, and persistently dependent users with reduced cognitive autonomy — and shows that social transmission, recovery and collective reinforcement interact to produce tipping points and technological lock-in. Past a critical threshold, small increases in adoption tip a population into persistent dependence with abrupt competence loss. The same model identifies the conditions for what the authors call cognitive immunization: reducing transmission and, above all, facilitating reversibility.
Reversibility is the operational word. For a writer it means one thing: can you still produce the first draft yourself? If the answer is yes, the tool is an amplifier. If the answer has quietly become no, you are in the third state, and the paper's model says the way out is to practise the unassisted version regularly, not to swear off the tool.
The workflow
1. Draft cold, badly, alone
Write the whole piece with the model closed. It will be rough. That is the point: the roughness is where your voice, your specific examples and your actual opinion live. Set a timer if you need one. The draft is the part of the process that keeps you reversible.
2. Use the model to comprehend, not to compose
Paste in the source material — the paper, the changelog, the transcript — and ask questions. "What does this paper claim and what does it not claim?" "Which of these numbers is measured and which is modelled?" This is the invaluable-for-comprehension use Cantrill grants, and it produces the specific, checkable details that absence tells are made of.
3. Edit with constrained prompts
Give the model your draft and a narrow brief. The prompt matters; open-ended "improve this" produces the tells.
You are a line editor. Do not add sentences. Do not add lists.
Do not change my claims or my examples. Preserve my sentence rhythm.
Do: cut redundancy, fix grammar, flag any sentence that is vague
or unsupported with [VAGUE] or [CITE?]. Return the edited text
and a short list of what you changed.
Then a second pass, adversarial:
List every phrase in this text that a reader would recognise as
LLM-generated: filler adjectives, parallel triplets, rhetorical
questions, "not just... but also", closing summaries. Quote each
one. Do not rewrite anything.
You make the cuts. The model flags; you decide.
4. Put something at stake
Before publishing, find the sentence a reader could disagree with. If there is none, the piece is not finished. Add the number you measured, the thing you got wrong last time, the recommendation you would defend in a meeting. This is the step no model does for you and the one that removes the deepest tell.
5. Read it aloud, then run the mechanical checks
Reading aloud catches rhythm tells — the sameness of sentence length, the paragraph that should be two. Then the mechanical pass: the readability checker will show you sentence-length variance and grade level, and the word counter will tell you whether the "quick" piece became 3,000 words of filler. Neither replaces the ear; both catch what the ear tires of.
The counter-argument, taken seriously
The strongest case for letting the model write is throughput: a marketing team producing forty pieces a month cannot hand-draft all of them, and readers of a product changelog do not care about voice. That is true for text nobody reads as a person. It is false the moment the byline matters — a founder's post, an engineer's write-up, anything where the reader is deciding whether to trust you. Cantrill's point is about exactly that text, and his line stands: have some confidence in your own voice, and write your own content. Where volume genuinely matters, the fix is fewer pieces with a person behind each, not more pieces with none.
Before and after: one paragraph through the workflow
Here is a paragraph of the kind that arrives when a model is asked to "write an intro about Python 3.15", followed by what survived the two prompts above and a human with a pen.
Before: "Python 3.15 isn't just another release — it's a game-changer for developers everywhere. 🚀 Whether you're a seasoned professional or just starting your coding journey, this comprehensive update delivers robust improvements across the board. From lazy imports to enhanced performance, let's dive into what makes 3.15 truly special. Ready to level up? Let's explore!"
After: "Python 3.15 ships on 1 October. Two changes will break code that has worked for a decade: open() now defaults to UTF-8 instead of the system locale, and the garbage collector reverts to the generational design after 3.14's incremental collector caused memory pressure in production. Everything else — lazy imports, frozendict, a faster JIT — is optional. Run your test suite under the release candidate with warnings as errors this week; on our three services that took forty minutes and found 58 call sites."
Count what changed. The emoji, the "not just", the "whether you're", the "journey", "comprehensive", "robust", "game-changer", "dive" and "explore" are gone — that is the lexical and syntactic layer, and a filter could catch most of it. What a filter could not do is what makes the second version readable: a date, two named breaking changes with a reason, a specific instruction, and a number from actual work. The model flagged the first layer in seconds. The second layer took a person who had run the migration, which is the whole argument.
A last honesty check we apply to our own posts, this one included: could a reader tell which sentences came from the source material and which from us? If the answer is no, that is not a compliment; it means the piece has no voice to distinguish. The fix is always the same — say what you did, say what you think, and let one of those be something a reader could push back on.
Quick answers
What are the most recognisable LLM writing tells?
Emoji markers, runs of single-sentence paragraphs, "not just X but also Y", parallel triplets of abstract nouns, and a vocabulary cluster around delve, tapestry, robust and journey. The deeper tell is the absence of anything specific or contestable.
Does the cognitive-virus paper say to stop using LLMs?
No. It models adoption dynamics and identifies reversibility — retaining the ability to work unassisted — as the condition for avoiding lock-in.
What are LLMs good at in a writing workflow?
Per Cantrill: brainstorming, comprehending source material, and editing. Not authoring in your voice.
Write your next piece cold, then run the two prompts above on it and count what they flag; the number will tell you how much of your voice you had been outsourcing. If you publish on a schedule, the Content Calendar Template with AI Prompts is structured around this draft-then-edit split rather than generate-then-post, and the Reel Hooks pack applies the same at-stake rule to short-form openers. Our own take on the adjacent debate is in Is prompt engineering dead?. Every product mentioned is available at wowhow.cloud — pay once, ship forever.
Originally published at wowhow.cloud
Top comments (0)