Why does Grunewald oppose even "AI drafting, manual editing"?

His foundation is "the writing process is the thinking process": writing ideas into text is the most rigorous test of your understanding; once text is on the page, people easily passively accept it. AI drafting saves not typing time, but thinking itself; moreover, AI writing is vague and erroneous in imperceptible ways—not hard errors you can catch, but sentences where the stance slips half a degree, things the AI doesn't even know. Using AI to write without labeling it, in his view, also breaks the reader's default that "behind the text is the author's viewpoint."

How does someone who updates three times a day practice "AI participating in writing"?

A sustainable three-step process: AI produces the draft (arguments, evidence, and talking points are pre-confirmed by a human), human rewrites paragraph by paragraph (after rewriting, the word count is cut by 30-40%, and every remaining sentence must answer "why did I write this"), and finally, a whitelist rule checklist clears out glue words like "it is worth noting" and "in conclusion." By the standard of "editor is okay, author is not," this process violates the draft stage but complies with the latter two; but what is exchanged is the sustainable output of three daily posts.

Does Writing with AI Count as Writing: Two HN Hot Posts, a Stance I Couldn''t Avoid

Q: What does the rule "not a single word suggested by an LLM can be used" prevent?

It prevents the unified smoothness of AI phrasing: frontier models are best at picking "pleasant to the ear" wording, so smooth it sounds like magazine headlines, which readers can detect with extreme sensitivity. This rule requires that even if a suggestion seems better than your own, you must not use it—AI can mark "surgical locations" like passive voice, nominalization, and filler words, but you must make the incision yourself. More subtle is the second rule, Avoid encouragement: models will applaud your terrible draft, causing you to double down on paragraphs that should be rewritten, losing the judgment that "this needs rewriting."

Opening: Two opposing articles in the same week

On September 17, the sockpuppet.org blogger published a post titled “How to Write with an LLM,” which hit the front page of Hacker News that same day: 624 upvotes and 373 comments. The core message: LLMs can be your copyeditor, but never your ghostwriter—you have to write for yourself.

Two days later, on September 19, Erich Grunewald’s “Why You Should Almost Never Use AI to Write Anything Substantive,” published on August 6, fermented again on HN: 213 upvotes and 117 comments. His stance is even tougher: almost all substantive writing should not involve AI—even if you manually edit it afterwards.

Combined, the two posts sparked nearly a thousand upvotes of debate on HN. This month, I happen to be updating three times a day, with every post passing through AI’s hands. So for this issue, I won’t play the peacemaker; I’m taking a side.

Shiwen: Let’s start with the topic—you use AI to write every day, so why dare you touch on this subject?

Yongliang: Precisely because I use it every day, I can’t avoid it. In seventeen years, I’ve led tech teams peaking at seventy people, reviewed proposals written by others, and been reviewed myself. Now, updating three times a day, AI participates in the process of every single post. I have ready samples of every usage criticized in these two articles. So today, I won’t just relay the chatter; I’ll lay out the boundaries of both articles clearly, then bring out my own process to see where I fit—which criticisms hold up, and which ones I’m also dodging.

Q1: First, let’s look at Article A—“Not a single word suggested by an LLM can be used”

Shiwen: Let’s break apart the 624-upvote post first. What exactly does the sockpuppet.org blogger allow and forbid?

Yongliang: Let me lay out his two rules exactly as they are; I won’t elevate them for him.

According to Article A, his division of labor is this: LLMs can be copyeditors, but not ghostwriters. A copyeditor is the final gate for publication: catching errors, smoothing sentences, checking consistency, but the words are yours. A ghostwriter writes for you: you provide the name, it provides the text. He believes LLMs can only do the former.

Rule One, and the harshest one: Not a single word suggested by an LLM can be used. His reasoning is that frontier models are best at picking “pleasant to the ear” wording, smooth to the point of sounding like magazine headlines; even if a suggestion seems better than your own, it is DQ’d without exception. I’ll translate the golden sentence from the original text: Readers can detect LLM text at a concentration of parts per trillion. This has rhetorical elements, but I agree with the direction—AI’s phrasing has a unified smoothness; readers can’t articulate exactly what’s wrong, but they can feel that something is wrong.

Rule Two, easily ignored but more important: Avoid encouragement. Models will reply “that’s gold, Jerry!” to your terrible draft; you write a paragraph you feel unsure about yourself, it applauds you, and you double down on paragraphs and metaphors that should be rewritten. His original words are: Those reflections during rewriting are the load-bearing parts of your voice. A terrible first draft isn’t scary; what’s scary is AI making you lose the judgment that “this needs rewriting.”

So what can AI do? He lists mechanical tasks for finding problems: passive voice, nominalization of verbs, repetitive phrasing, and filler words like very, unfortunately, really, actually. In a word, diagnosis is okay, prescribing is not; it can mark the location for surgery, but you must make the incision yourself.

Q2: Now look at Article B—Grunewald opposes even “AI drafting, manual editing”

Shiwen: Article A at least left room for an editor. Why does Grunewald oppose drafting plus editing together?

Yongliang: His three arguments go deeper than the last.

First, the writing process is the thinking process. He quotes Paul Graham’s “Writing ideas into text is the most rigorous test of your understanding”; he also cites philosopher Eric Schwitzgebel’s view: nodding along while reading existing text versus generating that text word by word yourself are two different cognitive activities—once text is on the page, people easily passively accept it. This is his foundation: you let AI draft, you save not typing time, but thinking itself.

Second, AI writing is vague and erroneous in places you can hardly detect. It’s not the hard errors you can catch, but sentences that seem smooth on the surface but reveal a half-degree slip in stance upon close reading—slips the AI doesn’t even know it made.

Third, using AI to write without labeling it is rude and misleading. He doesn’t budge on this: readers assume that behind the text is the author’s viewpoint; if you break this default, you should say hello.

Interestingly, he claims not to be anti-AI: transcription, data analysis, search, brainstorming, giving feedback on drafts are all okay; line editing, rewriting are also okay, but every change must be manually accepted or rejected one by one. He opposes only one thing: AI writing itself, even if you give detailed points, even if you edit afterwards. Because in his view, thinking happens before the pen hits the paper; the moment the bullet points are listed, the writing is already finished.

Shiwen: Where does this “before the pen hits the paper” idea diverge from Article A?

Yongliang: The two actually share the same boundary: AI can be an editor, not an author. The real disagreement is in execution details: A allows AI to diagnose problems and only forbids its phrasing; B believes diagnosis and phrasing cannot be separated, and that a draft touched by AI is twisted at the root, and manual editing is just pruning a twisted tree. One manages process, the other manages thoughts.

Q3: Ammunition from the HN comments, and the irony aimed at Article A itself

Shiwen: The comments are livelier than the main text; pick a few usable ones.

Yongliang: First, the one I most want to give to professionals, paraphrased: You want a second pair of eyes, not a first pair of hands. This is the top resonance under Article A’s 624 upvotes, cleanly separating copyeditor from ghostwriter. The second is like a general principle: Only use AI where it makes you think harder. The judgment standard isn’t in the tool list, but in yourself—after using AI in this step, did you think deeper, or did you stop thinking entirely? Use for the former, stop for the latter.

Two operational ones are also practical: Let AI criticize your draft, not rewrite it for you. The model will always give you a complete rewrite; don’t take it. Let AI write things you wish you had read, like a summary of a topic, a decision report, or meeting minutes—these are materials you need to read anyway, where no one can think for you.

The last one is the finale of this issue, aimed at Article A itself. A comment pointed out: The sentence “Those rewriting reflections are the load-bearing parts of your voice” is itself typical LLM-speak—the person teaching you how to prevent AI flavor might have been influenced by AI themselves. I give this irony full marks because it brings the whole matter back to earth: no one can stand outside AI to judge AI, including those who write the rules. We are all in the water; the difference is only in who admits they are in the water.

Q4: Putting myself under the knife—Daily updates, AI in every post, does my process withstand the criticism?

Shiwen: Time to map myself to the seat. What is your process, and how would Articles A and B judge it respectively?

Yongliang: Let me lay out the process first, three steps: AI produces the draft, human rewrites, whitelist rules clear AI flavor. Every post goes through these three steps, no exceptions.

Step one, AI produces the draft. Let me say it definitively: this step is exactly where the firepower of both articles is concentrated, and I won’t defend it. By Article B’s standard, my draft step is already out of bounds; thinking happens before the pen hits the paper, and I outsourced the pen-hitting. Article A’s standard is slightly lighter, but I still get scolded: the phrasing smoothed out by AI in the draft shouldn’t have been kept in the first place. My actual usage in this step is: feed in the arguments, evidence, and tone I want to write, and let it build a “skeleton of facts I confirmed.” This is my usage; I accept the part that gets scolded. Whether it saves typing time or thinking itself, the boundary between the two isn’t always clear to me either.

Step two, human rewriting—this is the real writing in my process. After getting the draft, I go through it paragraph by paragraph: is the argument what I wanted to say, is the example something I’ve seen, is the tone mine? The word count after rewriting is usually cut by 30 to 40 percent; this is my estimate, I haven’t counted precisely, but every post is in this magnitude. Article A’s rule “not a single word suggested by an LLM can be used” is the other side of the whitelist thinking I execute when rewriting: every sentence left must be able to answer “why did I write this sentence.”

Step three, whitelist rules clear AI flavor. List some well-known AI-tone phrasings, scan the final draft item by item, and rewrite any hits. This is mechanical work, the kind of “diagnosis” Article A approves of, only the object of diagnosis is AI itself.

Now mapping to the seat. Article A would judge me: draft step foul, rewriting and flavor-clearing steps compliant, overall stay on probation. Article B would judge me: out of bounds from the first step, the latter two steps can’t fix it. I accept both verdicts, but I won’t change the sentence, because what I exchange for the process is the output of three daily posts. One thing I’ve learned in seventeen years is: perfect principles are no match for sustainable processes. I admire Article B’s set, but if I only did premium work, this column couldn’t be weekly, let alone daily.

Q5: My Stance—The tool boundary is not in the tool, but in “who bears the thinking”

Shiwen: Finally, give me a straight answer. Which side are you on between the two articles?

Yongliang: I stand with Article A, but I want to take the message half a step further than Article A.

Standing with Article A is because Article B’s standard collapses practically into “don’t use AI to write,” and this isn’t a promise a person who updates three times a day, nor a tech lead leading a seventy-person team, can fulfill. A standard that can’t be fulfilled is worse than no standard, because it only filters out people who are good at pretending.

Half a step further than Article A. I think the rule “not a single word suggested by an LLM can be used” prevents the wrong focus. What readers resent is the lack of thinking behind the text. So my yardstick is only one, and it asks about the person, not the tool: For every judgment in this text, is there someone willing to stand up in a review meeting and say “I checked this, I’m responsible”? If yes, AI can participate as deeply as needed; if no, purely manual work is a hollow shell. The tool boundary lies in these five words: “who bears the thinking.”

This also returns to my profession of seventeen years. When leading a team, what I feared most was no one being able to explain clearly where the numbers in a document came from. The writing debate of the AI era is essentially the same thing in a different coat: what readers want is “a responsible person standing behind.”

Epilogue

Shiwen: Finally, summarize this issue in one sentence?

Yongliang: AI can be an editor, not an author; ninety percent of people agree with this boundary, and the remaining ten percent are the battlefield: the division of labor between editor and author is drawn in the daily process of every writer, and the pen drawing the line can only be human. Where you draw it today is more important than what conclusions two hot posts argue to.

Shiwen: This sentence is for everyone. See you next time.

[Technical Depth] What exactly is “AI Tone”, and why can readers smell it

This issue has been talking about “AI flavor” and “AI tone,” and some readers might want to know one layer deeper: what exactly is this thing, and why do so many people claim they can smell it.

First, it is a set of statistical preferences. When large models predict the next word, they naturally prefer the “globally average smoothest” continuation: explanations prioritize definitions, transitions prioritize “however,” conclusions prioritize elevating value. Individually, none are wrong, but linked together, there is a uniformity—every sentence exerts the same force, every paragraph lands the same way. Human writing has rough edges; ideas rush, detour, suddenly insert a sentence of irrelevant honesty. The model’s output has no rough edges; this is the truth of “smoothness”: not lack of personality, but averaged to the point where personality is invisible.

Second, the high-frequency appearance of filler words and connectors. The very, unfortunately, really, actually named in Article A are one type; in Chinese, they correspond to “it is worth noting,” “in conclusion,” “not only that,” “in a certain sense”—these universal connectors. They aren’t wrong, they are lazy—glue without specific information. Whitelist clearing AI flavor mainly clears these glue words: every hit must either be deleted or replaced with a sentence that has specific reference.

Finally, “being able to smell it” comes at a cost. That irony in the HN comment was cold: even the person teaching people to prevent AI flavor is using AI tone themselves. The ability to identify AI text relies on reading volume; having read enough real human text and AI text, the nose naturally becomes sensitive. And in this era, the growth rate of reading volume likely won’t catch up with the production volume of AI text. So rather than training your nose, it’s better to keep your eyes on that verifiable thing: behind this text, is there someone bearing the thinking?

Returning to the main thread of this issue: AI tone is the smell of statistics; responsibility is the smell of humanity. The former can be cleared; the latter cannot be faked.

Information Sources for This Issue

sockpuppet.org Blog 2026-09-17: “How to Write with an LLM” (Rule 1 “not a single word suggested by an LLM can be used”, Rule 2 “Avoid encouragement”, “parts per trillion” golden quote, “load-bearing parts of your voice” are all original viewpoints)
Hacker News 2026-09-17: Discussion on the above post, 624 ▲ / 373 comments (including “second pair of eyes, not a first pair of hands” top comment, “load-bearing” irony comment, both first-hand comment retellings)
Erich Grunewald Personal Substack 2026-08-06 First Release: “Why You Should Almost Never Use AI to Write Anything Substantive” (Three arguments, Paul Graham and Eric Schwitzgebel citations, list of allowed AI uses are all original viewpoints)
Hacker News 2026-09-19: Re-fermentation of the above post, 213 ▲ / 117 comments (including “only use AI where it makes you think harder”, “let AI criticize not rewrite”, “let AI write things you wish you had read”, “rereading corrects AI summary, subtlety disappears” etc. comment retellings).