Most arguments about AI and writing collapse into a simple productivity question: does it save time? A blog post from researcher Erich Grunewald skipped that question entirely and went viral anyway. Its thesis, laid out under the blunt title "Why I Think You Should Almost Never Use AI to Write Anything Substantive," is that AI-drafted prose is bad for three separate reasons — and on September 20, 2026, it climbed to 288 points and 141 comments on Hacker News, cross-posted to LessWrong along the way.
The Core Claim: Writing Is Thinking, Not Just Output
Grunewald's first argument is the one doing the most work. He leans on Paul Graham's old observation that "putting ideas into words is a severe test" — that half of what ends up in a good essay is discovered during the act of writing it, not before. Hand that drafting step to a model, Grunewald argues, and you skip the exact process that would have caught the holes in your own thinking. The essay isn't objecting to AI as a tool; it's objecting to what gets lost when a writer stops doing the cognitive work that writing is supposed to force.
A Contested Premise
That claim didn't go unchallenged. In the comments, one reader pushed back that "writing is thinking" generalizes from Grunewald's own experience — a version of the typical-mind fallacy, since some people reason perfectly well without writing it out first. Another commenter split the difference, suggesting the effect is real for open-ended, worldly analysis but weaker for something like mathematics, where the thinking can happen elsewhere.
Where AI Prose Goes Subtly Wrong
The essay's second argument is more concrete: Grunewald walks through a paragraph Claude generated about AI chip smuggling and picks it apart line by line, cataloguing vague phrasing and misleading framing — including a claim that estimates "vary widely" when, in his reading, the lower figures are almost certainly wrong. His point isn't that the AI lied outright. It's that the errors are hard to notice, exactly because the sentences read smoothly and confidently. A reader without domain expertise has no way to catch what's off.
The Disclosure Problem
The third leg of the argument is about trust rather than accuracy. Grunewald cites writer Clara Collier's framing: readers approach an authored piece expecting something they couldn't get anywhere else — a person's actual thinking, not a plausible-sounding average of it. Publishing AI-drafted text without saying so, the essay argues, breaks that implicit contract, even if every fact in the piece happens to check out.
Hacker News Draws Its Own Line
What made the thread notable wasn't agreement with Grunewald across the board — plenty of commenters pushed back on the "writing is thinking" claim, and one pointed out that treating AI writing as inherently detectable is shakier than people assume, after a comment accused of "sounding AI-written" turned out, per a Pangram check, to be entirely human. What stood out instead was how many commenters who use AI tools daily still landed close to Grunewald's conclusion.
Edit, Don't Draft
The most cited comment came from tptacek, a well-known security researcher who is no AI skeptic in his day job. His proposed rule: use AI for copyediting only, never generation — and specifically, "any word an LLM suggests to you is disqualified, even if it's better." In his framing, a model should flag weak verbs, passive voice, and overused words, but never supply replacement text. Another commenter, rectang, offered a lighter version of the same idea: ask the AI to critique a draft, then apply your own judgment about which suggestions to actually take.
The Productivity Counterargument
Not everyone accepted the line. Some commenters argued that refusing AI drafting is a competitive liability, warning that if a colleague producing three times your output starts using AI while you don't, the market decides who keeps their job — output, not process purity, is what gets measured. Others countered that AI drafts read as verbose and impersonal, and that the time spent hunting for a model's subtle inaccuracies can cost more than it saves.
What to Watch Next
The essay didn't convince Hacker News that AI has no place in writing — if anything, the thread confirms that most serious users already have a personal rule for where the line sits. But the fact that people who build and champion AI tools for a living keep arriving at some version of "let it edit, never let it draft" is itself the more interesting signal. As AI writing gets harder to detect and easier to reach for, that distinction — between a model correcting your words and a model choosing them — looks like the one actually holding.
-EditorZ
Photo by Vadim Bozhko on Unsplash
Post a Comment