Erich Grunewald thinks you should almost never use AI to write — and he is explicit that this is not an anti-AI position. Transcribing audio, analyzing data, searching, brainstorming and commenting on your drafts are all fine, as is line editing, provided “all the edits are deliberately accepted or rejected by a human.” What he objects to is the writing itself: the part where you type words on a page to convey an argument. The 94-comment thread on Hacker News takes the argument seriously enough to attack it from both ends.

His three reasons

  • The writing process is the thinking process. You cannot know what you want to say until you have to say it. He quotes Paul Graham (“Half the ideas that end up in an essay will be ones you thought of while you were writing it”) and Clara Collier, who describes fighting a transition that will not work and realising the reason is that the two points should not be adjacent at all. An LLM fed that outline will not stop and tell you the outline is bad.
  • AI writing is vague and wrong in hard-to-notice ways. To show it, he asks Claude Fable 5 for a short paragraph on AI chip smuggling and annotates the result. It reads fine, and it is full of holes: “export controls are only as strong as enforcement” says something obvious or nothing; chips are called “compact” though the smuggled goods are servers, and they get relabelled and shipped in plain sight rather than hidden; a widely varying estimate range is quoted without noting the low end is probably wrong; “enough compute to matter for training and deploying capable AI systems” carries no information at all.
  • Unlabeled AI writing is rude and misleading. When you publish or send a text, the reader pays attention and expects it to be your thoughts. Grunewald calls that the contract between reader and writer, and says AI-written prose — vague in a way that sounds authoritative — violates it. Labelling the output fixes the rudeness; he thinks very few people want to do that.

The failure mode is not obvious error, it is confident vagueness. He cites Eric Schwitzgebel’s point that reviewing a text is a different cognitive act from generating one — you nod along instead of working out the right word — and adds that a model asked about something it does not know will rarely say so: it “will write something that’s vague and meaningless enough to be true or something that sounds true but isn’t.”

He is careful with the exceptions, which suggests the argument is a craft claim rather than a ban. Short formulaic coordination emails: okay. Translating a text you wrote in your native language: acceptable, better disclosed. An entire substantive text by AI: almost never, even labelled, because people will not read it and you lose the thinking. The “high stakes, so I need the speed” case gets short shrift — policy writing already drowns in briefs nobody reads, and what is scarce is accuracy.

What the thread adds

  • bko — the sharpest dissent, arguing the rudeness runs the other way: “I think it’s rude to not use the best tools to convey the message most clearly and most respectfully of my time.” They use AI to tighten emails, and note that human writing is also vague and wrong — just “more likely to be wrong in easy to notice ways, which is … better?” Two replies push back: elmomle describes a partner company’s AI-written doc whose “basic reasoning errors” cost a team time to discover in good faith, and mathgeek reports the opposite experience on length — AI writes more than they do, not less.
  • atdt — the strongest counterexample, and a direct refusal of the premise: “LLMs have made me a better writer of technical reports.” Their claim is that iteration with a patient model breaks you out of a linguistic frame you cannot escape alone — “The marble is the agent’s, but the chisel and mallet are in my hands.” Replied to with a blunt “it makes sense why AI writing resonates with you” from semanticc.
  • patrickmay — corroboration from the other direction, with a number behind it: two weeks spent reviewing and improving an LLM-drafted work document, and the conclusion that “writing it manually would have been faster and resulted in a higher quality product.”
  • jameshart — a rule that maps where the line falls: “Use AI to write things for you to read that you wish someone else had written… Don’t use AI to write things that you are producing for someone else to consume.”
  • vanschelven — extends the whole argument to code: “Writing code is thinking, AI code is often vague and wrong in hard-to-notice ways, and the (human) reader of code is the one that pays the cost for this later.” perrygeo disagrees on the economics, because code can be run and checked objectively, which prose cannot.
  • rectang and gyulai — two countermeasures that stay inside the author’s carve-outs. rectang: ask the model to critique your writing instead and “use your own judgement as to which bits of advice to follow and which to ignore,” since it will always offer a full rewrite. gyulai: have the LLM draft in a different language, use that only as a blueprint, then write the target-language draft yourself so “every word choice” is yours.
  • tzumaoli — the observation that covers a cluster of near-identical comments (gyulai, a-dub, pigpop, sklargh all describe a personal workflow that supposedly dodges the problem): “people always think their way of using LLMs is the best (and other people are using LLMs in the wrong way).”
  • erwald — the author, in his own thread, after vikrantrathore argued the post must be AI-written because it uses double dashes instead of em-dashes. in_absentia guessed a Substack paste artifact; Grunewald replied that he writes in org-mode: “the article is 100% human-written.”

On reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering above is HN’s own ranking, not a vote. This is a slice of the thread, not a consensus.