General 21 min read

Never Take a Word the Model Suggests

MMNMNOTE
trendai-writingwriting-craftai-detectioneditingllm

Never take a word the model suggests. Thomas Ptacek's rule is right, and his stated reason for it is wrong. He justifies it by claiming readers will detect the borrowed phrase. Two peer-reviewed studies measured exactly that and found they cannot. A better reason survives: granularity.

The rule comes from "How To Write With An LLM," published on 17 September 2026. Ptacek states it as a heading: "Rule Number One: You may not use a single word an LLM suggests to you."1 Rule Number Two is shorter: "Avoid encouragement."1 The model finds faults. It never supplies phrases.

Three days later the Hacker News thread was past 700 points and approaching 400 comments, and was still being argued with rather than applauded.2 Simon Willison read it as "Thomas Ptacek on using LLMs as copyeditors, not as writing assistants."3 That is the essay's own framing: "It's just you need to use them like a copyeditor rather than a ghostwriter."1

The reason the rule is usually given

The justification offered is detection. The essay's second paragraph opens with it: "Readers can detect LLM words in the parts per trillion."1 The claim is that a borrowed phrase leaks, that the audience senses it, and that the cost is your standing with them. Stated that way, the rule is a defense against getting caught.

It is a compelling mechanism, and the essay is specific about why. "Frontier models are supernaturally good at selecting pleasing turns of phrase," Ptacek writes — which is precisely what makes them dangerous to a draft, because the suggestion that survives your judgment is the one you liked.1

He states the consequence twice, once hard and once soft. The soft version is more careful: "an LLM paragraph will register to much of your audience not as writing but as output."1

"Much of your audience" is a claim you could plausibly defend. "Parts per trillion" is not.

Readers cannot reliably detect LLM words

They cannot. Elizabeth Clark and five co-authors ran the test in 2021: untrained evaluators "distinguished between GPT3- and human-authored text at random chance level" across stories, news and recipes.4 Training helped barely — accuracy "improved up to 55%," and not consistently across the three domains.4 That is a coin with a lean.

The second study is larger and worse for the premise. Across "six experiments, participants (N = 4,600) were unable to detect self-presentations generated by state-of-the-art AI language models in professional, hospitality, and dating contexts," report Jakesch, Hancock and Naaman in PNAS.5

Worse still, the failure is systematic. Human judgment leans on "intuitive but flawed heuristics such as associating first-person pronouns, use of contractions, or family topics with human-written language" — heuristics stable enough that a system can be tuned to produce text perceived as "more human than human."5

Detection is not a skill readers have. It is a story they tell themselves about a text after the fact.

Name the limits before someone else does. Clark tested GPT-2 and GPT-3; Jakesch tested 2022-era models with untrained crowdworkers. Neither measures a 2026 reader who has spent years marinating in model output, and both measure detection in isolation rather than in a feed. What no one has produced is a measurement supporting the parts-per-trillion claim. If one exists, it changes this argument.

He keeps the rule he wrote

The correction is to Ptacek's reason, not to his practice, and the cheapest attack on him does not survive contact with his own text. We counted. The essay body runs 1,259 words. The four filler words he names — "very," "unfortunately," "really," "actually" — appear nine times in it.6

Every one of the nine is deliberate. Four sit inside the sentence that names them as faults. Three sit inside the joke immediately after, where he calls a certain kind of edit "really, actually, very satisfying."1 The last two sit inside a build prompt he pastes in whole. Unmarked uses in his own prose: zero.

He shows his working elsewhere too. "I scrubbed 7 instances of the word 'just' from it," he wrote of this piece, in the thread.7

One commenter read the phrase "Those rethinks are load-bearing parts of your voice" as slop and said so.8 Simon Willison replied that he had "instantly assumed Thomas was making a deliberate point."9 Ptacek confirmed it one reply further down, in three words: "It was a joke."10 The exchange is worth keeping precisely because it is already settled. A take that runs it as a gotcha is repeating a thread it did not finish reading.

The real reason is granularity

The reason that survives is granularity. Every tool for separating your words from the model's works on a unit larger than a phrase. A label marks a note. A watermark scores a passage, readable only by the key-holder. None of them resolve a clause. The distinction exists at phrase granularity for one moment: before you accept the suggestion.

Our own corpus has already conceded both halves of that. The note-level answer is a generated-by field you write yourself, and that post says plainly that "provenance is not recoverable after the fact" and that what you get is "a label that is not a lock."

The machine-level answer is narrower still. A production watermark does exist in AI-generated prose, but the key that reads it belongs to whoever issued it — so a positive detection is evidence and an absent watermark is nothing.

Neither operates below the paragraph. Neither ever will, because the signal a watermark needs is statistical and a clause is too small to carry it.

So the boundary between your sentence and the model's is not a property of the finished document. It is an event, and it happens once, at the keyboard, in the half-second where you either take the phrase or don't.

That is why the rule has to be absolute rather than tasteful. Ptacek writes it as a disqualification, not a preference: "Even if you like the words, even if you're sure they're better than what you already have, LLM-generated phrases are DQ'd."1 A preference negotiates. A disqualification does not need to.

Be precise about what this buys. It does not make your writing legibly yours to a reader; the studies above settle that. It makes the composition of your draft knowable to you, which is a smaller claim and the only one the evidence supports.

The strongest objection: exposure, not copying

The rule polices copying. The commenter SupremumLimit argues contamination does not travel that way: "while you're reading their output, the slop makes its way directly into your brain anyway - and then it drips out into your writing."8 If that is right, a rule about which words you paste cannot contain it.

This objection has an experiment behind it. Jakesch and four co-authors asked 1,506 participants to write a post on whether social media is good for society, some of them with a writing assistant tuned to argue one side. "Using the opinionated language model affected the opinions expressed in participants' writing and shifted their opinions in the subsequent attitude survey."11

Read that honestly in both directions. The study measured opinions, not phrasing, and its assistant suggested completions rather than flagging faults, a stronger exposure than the fault-finding workflow under discussion. It does not prove that reading a critique borrows you a cadence. It does establish that the influence channel runs through reading, not only through pasting.

A second commenter pushes the same wedge from the other side. "You'll anchor to the suggestions made by the LLM and it will turn your writing into LLMese," writes in_absentia.12

Before that line comes the circularity. "To look at the style suggestions produced by an LLM and decide that you agree with some of them in spirit, but want to put it in your own words, you already need how to know how to write well," the same comment argues; the grammar slip is in the original.12 To use the model safely you must already write well enough not to need it.

Both objections land. Neither defeats the rule; they resize it.

The rule is a floor, not a fence. It closes the one channel you can actually close — the phrase you consciously accept — and leaves open the one you can't, short of not reading the output at all. Ptacek concedes that option generously: "You can just not copyedit. That's fine. More than fine, lately: rough edges are a signifier of authenticity."13

Half of this was automated in 1981

Ptacek answered the linter objection with a challenge: name "the actual human-written program that you're thinking of that detects verb nominalization, verifies that the subjects of sentences are actual characters in the action, that new information is placed in the stress position of sentences, and that each paragraph has a clear topic sentence."14 Part of that program shipped in 1981.

The objection came from the commenter Semiapies: "Actual human-written programs exist for that purpose, things you don't have to bludgeon into not praising you or semantically ablating your work."15

Semiapies is right about the first item on that list. The style command, still packaged today, takes -p/--print-passive to "print all sentences phrased in the passive voice" and -N/--print-nom to "print all sentences containing nominalizations."16 Both faults Ptacek names in the essay — passive voice, nominalized verbs — have been one flag away for four decades.

The lineage is documented in the manual itself, which points back to Cherry and Vesterman's Writing Tools — The STYLE and DICTION programs, Bell Laboratories Computer Science Technical Report 91, 1981.16 The peer-reviewed write-up is Macdonald, Frase, Gingrich and Keenan, "The Writer's Workbench: Computer Aids for Text Analysis," IEEE Transactions on Communications 30(1), January 1982, pages 105–110.17

He is wrong about the other three. Nobody in the thread named a program that verifies a subject is a character taking action, that new information sits in the stress position, or that a paragraph has a topic sentence, and the style and diction documentation disclaims the job outright: "These programs cannot help you structure a document well, but they can help to avoid poor wording."18

Half wrong, half right.

The half he is right about is the half that matters. The three unautomated clauses all require modeling a reader's mind moving through a sentence, which is exactly the work no regular expression has ever done.

Ptacek's defense of why a model can do it at all is the strongest paragraph he wrote in the whole thread: "More importantly: they're technical. They're not really judgement calls. The subject of a sentence either is or is not a character taking actions in the story you're trying to tell."19

The criteria are his, chosen in advance, out of Joseph Williams's Style: Lessons in Clarity and Grace — a book the essay names outright and credits Richard Gabriel for.1 The model is only the thing that never gets bored of checking.

What to do at the keyboard

The discipline is five moves, and none of them are about output quality. They are about keeping the seam between your language and the model's visible at the only moment it is visible. The cost is real: Ptacek reports "8-12 passes on any piece" and responding to roughly 60% of suggestions.20

  1. Ask for faults, never for phrasing. His standing instruction begins "Workshop a piece with me. NO ENCOURAGEMENT. Encouragement is useless; the only useful things are suggested corrections," and continues "DO NOT WRITE COPY FOR ME. Any words you provide will be disqualified."20 Encouragement is the default failure — hand a draft to a model and it tells you the thing is gold.
  2. Disqualify the phrase even when it is better. The moment you grade suggestions on quality, the rule stops being a rule.
  3. Run the 1981 tools first. style -p and style -N are offline and deterministic, and they will flag the mechanical faults before a model ever sees the draft.16
  4. Budget passes, not minutes — and refuse a share of them. Ptacek declines about 40% with "nah, I wanted it to sound that way."20 A pass you accept wholesale is a pass you didn't read.
  5. Keep the draft somewhere you can diff it. A plain text file under version control is the one phrase-level record of your own composition that actually exists, and it is the only one you hold yourself.

Then stop short of finishing.

Ptacek's last line about his own essay is the corollary to the whole method: "it said the whole thing was 20% too long. It's probably right. But I'm not fixing it."1

Frequently asked questions

The first seven are questions the Hacker News thread actually asked, quoted as they were written; the last is the one this post exists to answer. Each answer points at the evidence above rather than restating it. Where a question belongs to a different argument than this one, the answer says so instead of stretching.

"If you can't spend the time to write it, why should anyone read it?"

This was the thread's centre of gravity, and Ptacek agrees more than the question expects. "You can just not copyedit. That's fine," he replies.13 The method assumes you already wrote the thing. It speeds up refinement, not composition. The essay's own word for the role is copyeditor.1

"Does anyone know about 'Style: Lessons in Clarity and Grace'?"

It is the book the whole method rests on, by Joseph Williams. The essay names it and credits Richard Gabriel for the recommendation.1 In the thread Ptacek puts it plainly: "The model didn't tell me to do that; Joseph Williams did, decades ago."19 The criteria come from the book. The model only applies them.

"What is the actual human-written program that you're thinking of that detects verb nominalization…?"

Partly, style. The -N flag prints sentences containing nominalizations and -p prints passive-voice sentences, and both have shipped since 1981.16 The other three things Ptacek asks for — subjects as characters, stress position, topic sentences — have no such program, and the tools concede it.18

"Why are you giving yourself all this busy work?"

Because the mechanical pass is tedious and the machine is not. "You can spot them mechanically, but that's tedious and exhausting work. The models don't get tired," as the essay puts it.1 If you do not run editing passes at all, the method has nothing to offer you, and that is a legitimate way to write.

"How do you know it's not his/her thought?"

Largely, you don't. Untrained readers separate machine text from human text at chance, and trained ones reach about 55%.4 That is the whole case for controlling what enters the draft instead of auditing what came out of it — the input is the only side of the process you can observe.

"What's preventing the editor from using one of your prompts/tooling and replacing you entirely?"

The criteria, which are chosen by a person in advance. "They're technical. They're not really judgement calls," Ptacek argues; the model flags, the writer decides what counts as a fault.19 A prompt without a writer's prior commitments produces critique nobody asked for.

"Could we set up a pipeline that leverages the voice models to clean up text that reeks of AI?"

That is the opposite discipline — repair after the fact — and it has a measurable cost. Automated rewriting is what produces tortured phrases, where multi-word terms are destroyed and single words survive. Prevention at the keyboard leaves nothing to launder.

"Can readers really detect words an LLM suggested?"

The best available measurements say no. Clark and colleagues found untrained evaluators performed at random chance, rising only to about 55% with training.4 Jakesch, Hancock and Naaman found that across six experiments with 4,600 participants, people could not detect AI-generated self-presentations.5 No measurement supports a stronger claim.

The seam is only visible once

Two rules, one of which is right for a reason its author did not give. Readers will not catch the borrowed phrase. Watermarks will not surface it. A note-level label cannot reach down to a clause. The record does not exist below the paragraph, which is why the only place it ever existed is the instant before you press accept.

We have argued before that writing it yourself is what makes it yours to recall. This is the mechanical half of the same claim.

Refuse the phrase you liked. Not because someone will notice. Because you are the last one who can still tell.

The rule and the method are Thomas Ptacek's; only the reason has been corrected here.

Two rules like these are easiest to keep when the draft is plain Markdown on your own device, where every version stays yours to compare — which is what mnmnote.com is for.

Footnotes

  1. Ptacek, T. "How To Write With An LLM." A Final Ward, sockpuppet.org, 17 September 2026. https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/ — archived at https://web.archive.org/web/20260920162906/https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/. Accessed 2026-09-20. ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13

  2. Hacker News item 49747070, "How to Write with an LLM," submitted by joeriddles, 17 September 2026 21:48:38 UTC. 704 points and 393 comments in the thread tree as retrieved via the HN Algolia API at 2026-09-20T17:12Z; both counters are live and still climbing. https://news.ycombinator.com/item?id=49747070. Retrieved 2026-09-20. ↩

  3. Willison, S. "How to Write with an LLM." simonwillison.net, 17 September 2026. https://simonwillison.net/2026/Sep/17/how-to-write-with-an-llm/. Accessed 2026-09-20. ↩

  4. Clark, E., August, T., Serrano, S., Haduong, N., Gururangan, S., & Smith, N. A. "All That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated Text." Proceedings of ACL-IJCNLP 2021, pp. 7282–7296. DOI 10.18653/v1/2021.acl-long.565. https://aclanthology.org/2021.acl-long.565/. Accessed 2026-09-20. ↩ ↩2 ↩3 ↩4

  5. Jakesch, M., Hancock, J. T., & Naaman, M. "Human heuristics for AI-generated language are flawed." Proceedings of the National Academy of Sciences 120(11):e2208839120, 7 March 2023. DOI 10.1073/pnas.2208839120. Abstract verified via Europe PMC (PMID 36881628). Accessed 2026-09-20. ↩ ↩2 ↩3

  6. MNMNOTE measurement of the essay body, 20 September 2026. Body span from "It's tricky to write about writing" to the copyright footer, HTML tags stripped, tokenized on letters, apostrophes and hyphens: 1,259 words; very ×2, unfortunately ×1, really ×4, actually ×2 (nine total, each use-mention, joke, or inside a quoted prompt); just ×3; load-bearing ×1. Source: https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/. ↩

  7. Ptacek, T. (tptacek), Hacker News comment 49748579, 18 September 2026 00:15:02 UTC. https://news.ycombinator.com/item?id=49748579. Accessed 2026-09-20. ↩

  8. SupremumLimit, Hacker News comment 49748006, 17 September 2026 23:12:56 UTC. https://news.ycombinator.com/item?id=49748006. Accessed 2026-09-20. ↩ ↩2

  9. Willison, S. (simonw), Hacker News comment 49748255, 17 September 2026 23:36:11 UTC, in reply to comment 49748006 and the parent of comment 49748296. https://news.ycombinator.com/item?id=49748255. Accessed 2026-09-20. ↩

  10. Ptacek, T. (tptacek), Hacker News comment 49748296, 17 September 2026 23:39:07 UTC. https://news.ycombinator.com/item?id=49748296. Accessed 2026-09-20. ↩

  11. Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L., & Naaman, M. "Co-Writing with Opinionated Language Models Affects Users' Views." CHI '23: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pp. 1–15, 19 April 2023. DOI 10.1145/3544548.3581196. Abstract verified against the authors' deposit at https://arxiv.org/abs/2302.00560. Accessed 2026-09-20. ↩

  12. in_absentia, Hacker News comment 49757059, 18 September 2026 16:54:58 UTC. https://news.ycombinator.com/item?id=49757059. Accessed 2026-09-20. ↩ ↩2

  13. Ptacek, T. (tptacek), Hacker News comment 49753453, 18 September 2026 12:30:31 UTC. https://news.ycombinator.com/item?id=49753453. Accessed 2026-09-20. ↩ ↩2

  14. Ptacek, T. (tptacek), Hacker News comment 49757183, 18 September 2026 17:03:38 UTC, in reply to comment 49757000. https://news.ycombinator.com/item?id=49757183. Accessed 2026-09-20. ↩

  15. Semiapies, Hacker News comment 49757000, 18 September 2026 16:50:26 UTC. https://news.ycombinator.com/item?id=49757000. Accessed 2026-09-20. ↩

  16. style(1) manual page, GNU diction package (Debian bookworm, 1.11-1+b1). Flags -p/--print-passive and -N/--print-nom; SEE ALSO cites Cherry, L. L. & Vesterman, W., "Writing Tools — The STYLE and DICTION programs," Computer Science Technical Report 91, Bell Laboratories, Murray Hill, N.J., 1981. https://manpages.debian.org/bookworm/diction/style.1.en.html. Accessed 2026-09-20. ↩ ↩2 ↩3 ↩4

  17. Macdonald, N. H., Frase, L. T., Gingrich, P. S., & Keenan, S. A. "The Writer's Workbench: Computer Aids for Text Analysis." IEEE Transactions on Communications 30(1), January 1982, pp. 105–110. DOI 10.1109/TCOM.1982.1095380. Bibliographic record verified via Crossref. Accessed 2026-09-20. ↩

  18. GNU Project. "Style and Diction." https://www.gnu.org/software/diction/. Accessed 2026-09-20. ↩ ↩2

  19. Ptacek, T. (tptacek), Hacker News comment 49761565, 18 September 2026 23:19:21 UTC. https://news.ycombinator.com/item?id=49761565. Accessed 2026-09-20. ↩ ↩2 ↩3

  20. Ptacek, T. (tptacek), Hacker News comment 49753616, 18 September 2026 12:43:44 UTC. https://news.ycombinator.com/item?id=49753616. Accessed 2026-09-20. ↩ ↩2 ↩3