Erich Grunewald은 블로그, 보고서, 메모 등 실질적인 내용을 담은 글을 AI로 쓰는 것을 거의 항상 피해야 한다고 주장한다. 그 이유는 글쓰기 과정 자체가 사고 과정의 핵심이며, AI가 쓴 글은 알아차리기 어려운 방식으로 모호하고 틀리며, AI 작성임을 표시하지 않는 것은 무례하고 오해를 불러일으키기 때문이다. 저자는 AI를 음성 전사, 데이터 분석, 초안 피드백 등에는 활용해도 된다는 입장으로, 순수 '집필' 위임에만 반대한다.
번역된 본문
왜 AI로 중요한 글을 쓰는 일은 거의 멈춰야 하는가 — 하나의 호소. Erich Grunewald, 2026년 8월 6일.
나는 블로그 글, 연구 보고서, 메모, 정성 들인 이메일, 소설 등 아이디어·논증·분석이나 그 외 실질적인 생각을 전달하는 모든 텍스트든, 페이지에 글자를 입력하는 행위 자체를 AI에게 맡기는 일은 거의 하지 말아야 한다고 생각한다. 아주 상세한 핵심 요점이나 구두로 전달한 생각, 다른 맥락을 AI에게 제공하더라도, 그리고 AI가 쓴 텍스트를 직접 수정하더라도 마찬가지라고 본다.
그렇게 생각하는 이유는 세 가지다. (1) 글쓰기 과정은 사고 과정의 본질적인 부분이고, (2) AI가 쓴 글은 알아차리기 어려운 방식으로 모호하고 틀리며, (3) AI를 이용해 쓰고 그렇다고 표시하지 않는 것은 무례하고 오해를 불러일으킨다. 아래에서 자세히 설명하겠지만, 먼저 몇 가지 전제를 밝힌다.
알다시피 나는 반(反)AI가 아니다. 연구와 글쓰기 과정의 다른 많은 부분, 예컨대 음성 전사, 데이터 분석, 정보 검색, 브레인스토밍, 초안에 대한 피드백에는 AI를 쓰는 것이 합리적이라고 생각한다. 또한 문장·표현 수준의 편집이나 문단을 더 명확하고 간결하게 다듬는 용도로 AI를 쓰는 것은, 모든 수정을 사람이 의도적으로 수용하거나 거부하는 한 괜찮다고 본다. 내가 반대하는 것은 텍스트를 '쓰는' 일에 AI를 쓰는 것뿐이다.
물론 AI로 글을 쓰면 힘이 덜 들고 직접 쓰는 것보다 훨씬 빠르다는 등의 장점이 있다. 따라서 AI 글쓰기의 단점이 전체적으로 볼 때 나쁘다고 말하려면 상당히 커야 하는데, 짐작하듯 나는 그렇다고 생각한다.
마지막으로, 이 주장은 현재 존재하고 가까운 미래에 존재할 것으로 예상되는 AI 모델에 대한 것이다. 언젠가 글쓰기를 위임해도 될 만큼 충분히 좋은 모델이 나올 가능성은 있다(다만 그 시점에는 글쓰기뿐 아니라 사고의 전부 또는 대부분도 해야 하므로, 연구나 집필 과정 전체를 처음부터 끝까지 위임하는 편이 더 합리적일 수 있다).
글쓰기 과정은 곧 사고 과정이다
어떤 종류의 연구든 그 목적은 중요한 질문에 대해 정확한 믿음을 형성하고 이를 독자에게 전달하는 것이다. 나는 그 최선의 방법 중 하나가 글쓰기라고 생각한다. Paul Graham은 이렇게 썼다. "무언가에 대해 쓰다 보면, 잘 안다고 생각했던 것조차 실제로는 그만큼 잘 몰랐다는 사실이 드러난다. 생각을 말로 옮기는 것은 혹독한 시험이다. [...] 에세이에 담기는 아이디어의 절반은 글을 쓰는 도중에 떠오른 것이다. 실제로 그래서 나는 글을 쓴다."
Patrick McKenzie의 팟캐스트 한 에피소드에서 Clara Collier은 이렇게 말했다. "실질적인 무언가를 쓸 때, 글쓰기 과정에서 생각하지 않고 마음을 바꾸지 않는 순간은 없다. 개요부터 그것을 텍스트로 옮기는 과정, 그리고 한 문장 한 문장까지 전부 그렇다. 개요를 완성된 글로 만들려고 하면서 문장 연결을 시도하는데 잘 안 풀릴 때, '아, 이 연결이 안 되는 이유는 사실 이 두 가지 요점이 나란히 놓여서는 안 되기 때문이구나. 지금 내가 하려는 게 잘못됐구나' 하고 깨닫곤 한다. 그런데 LLM에 개요를 넣으면, LLM은 멈춰서 '개요 자체가 나쁘지 않을까'를 고려하지 않는다."
Patrick은 이렇게 답한다. "글쓰기 과정이 곧 사고 과정이라는 데 완전히 동의한다. 이제 그것의 실증적 증거가 나왔다고 본다. 글을 쓰는 기계를 만들었더니 사고가 부산물로 튀어나왔으니 말이다."
여러 페이지짜리 글을 쓸 때 논지가 아무리 확고하더라도, 그것을 문장과 문단으로 강제로 풀어내 논증을 구조화하는 행위는 논증의 빈틈이 어디인지 보여준다. 연구 과정이 어디서 부족했는지 드러난다. '아, 실제로 여기에 뭔가 빠진 게 있구나' 같은 것들이 수면 위로 올라온다.
Why You Should Almost Never Use AI to Write Anything Substantive A plea. Erich Grunewald Aug 06, 2026 80 17 20 Share I think you should almost never use AI to write -- that is, to do the thing you’re doing when you type words on a page -- whether for a blog post, a research report, a memo, a thoughtful email, a novel, or any other text aimed at conveying an idea, an argument, an analysis, or other substantive 1 thoughts. I think this is the case even when you give the AI very detailed bullet points, dictated thoughts, or other context, and even when you edit the AI-written text. 2 I think so because (1) the writing process is an essential part of the thinking process, (2) AI writing is vague and wrong in hard-to-notice ways, and (3) writing with AI (and not labeling it as such) is rude and misleading. I’ll explain these points in more detail below, but first, a few throat clearings. As you may know, I’m not anti-AI. I think it makes a lot of sense to use AI for many other parts of the research and writing processes, such as transcribing audio, analyzing data, searching for information, brainstorming, and giving feedback on drafts. I also think using AI for line and copy editing, or for rewriting a passage to make it clearer or tighter, is fine, as long as all the edits are deliberately accepted or rejected by a human. It’s just using AI to write text that I’m against. 3 And yes, there are various advantages to using AI for writing. For example, it’s less effortful and much faster than writing yourself. So the disadvantages of using AI for writing need to be substantial for it to be bad overall. As you may have guessed by now, I think they are. And finally, I’m just making a claim about the AI models that exist now and that I expect to exist in the near future. There will likely exist models at some point that are good enough that it makes sense to delegate the writing to them (although at that point it might make more sense to delegate the entire research or writing process end-to-end, since in addition to the writing they will also need to be doing all or most of the thinking). The Writing Process Is the Thinking Process The point of doing any kind of research is to form accurate beliefs about important questions, which you can then communicate to an audience. One of the best ways of doing that is in my opinion by writing . Paul Graham has written 4 that Writing about something, even something you know well, usually shows you that you didn’t know it as well as you thought. Putting ideas into words is a severe test. [...] Half the ideas that end up in an essay will be ones you thought of while you were writing it. Indeed, that’s why I write them. On an episode of Patrick McKenzie’s podcast, Clara Collier says that When I am writing something, something substantive, there’s no part of that writing process in which I am not thinking and changing my mind. Everything from the outline to turning it into text to just the sentence. Often I’ll have an experience where I’m trying to turn an outline into a finished product, and I’m playing with a transition, and it’s not working, and I realize, oh, the reason this transition isn’t working is because actually these two points should not be juxtaposed. The thing that I’m trying to do here is wrong. And if I feed the outline into an LLM, it is not going to stop and consider maybe the outline is bad. [...] Patrick replies: I absolutely agree that the writing process is the thinking process, and I think we now have empirical demonstration of this, because you make a machine that does writing and thinking just pops out as a side effect. But when writing a multi-page piece, even if you have a strong thesis, the act of forcing yourself to write that down into sentences and paragraphs to structure the argument will show you where the holes are in the argument. It will show you where your research process has been inadequate. It will surface things like: oh, actually, there is something that is contradictory to the thing that I want to say. And, being a reasonably ethical person, I think I will acknowledge that there are counterarguments, or have to add that the full claimable conclusion is less than the one that I would, quote-unquote, “like to write”. I basically agree with these takes. When I write something, I must first know what I want to say, i.e., I must think about evidence and arguments and arrive at some beliefs. So writing forces me to do those things, which are necessary for any writing or research project. For any such project, someone has to have the entire thing in their head -- all the bits of evidence, considerations, arguments, and how these fit together. Writing with AI makes it easier to avoid the necessary thinking. AI Writing Is Vague and Wrong in Hard-to-Notice Ways When I write something, I make a lot of small decisions about specific wording and phrasing. Together, these small decisions contribute to the overall clarity and correctness of the text. For example, if I ask Claude Fable 5 to “write a short paragraph introducing the problem of AI chip smuggling”, a topic that’s been written about and reported on extensively, it outputs: Since October 2022, the United States has restricted exports of advanced AI chips to China, aiming to slow Chinese progress at the frontier of AI. But export controls are only as strong as their enforcement, and AI chips are compact, extraordinarily valuable, and sold through sprawling global supply chains — properties that make them attractive and feasible to smuggle. Diversion networks route controlled GPUs through intermediaries in places like Singapore, Malaysia, and the UAE, using shell companies and falsified paperwork to move them into China. Estimates of the scale vary widely, from tens of thousands to perhaps more than a hundred thousand high-end chips per year, and even the lower end represents enough compute to matter for training and deploying capable AI systems. Meanwhile, the Bureau of Industry and Security, tasked with policing a multi-hundred-billion-dollar trade, operates with a budget and headcount better suited to a far smaller problem. Chip smuggling thus threatens to quietly hollow out one of the central pillars of US AI policy. That’s not terrible, and perhaps even quite reasonable, but is that how I would write it? No, in fact, Claude made a lot of choices that I find subtly wrong or bad: Claude writes that “export controls are only as strong as enforcement”, but what does this mean? It either says something obvious (of course policies that are not enforced or poorly enforced are less effective) or nothing at all. 5 Claude writes that AI chips are “compact”, which is true, but what is usually smuggled are AI servers, which are not compact. Anyway, more importantly, this doesn’t matter, because AI chip smuggling rarely involves hiding products to get through customs; usually the products are just relabeled as some other kind of good and shipped in plain sight, so to speak. Claude writes that being “sold through sprawling global supply chains” makes AI chips “attractive and feasible to smuggle”. What does this mean? Is it that smugglers can more easily buy chips from companies outside the US? (Until recently, smugglers seem to have been able to procure AI chips from US-headquartered companies with relatively little difficulty.) Is it that it makes smugglers buying a lot of AI chips in countries such as Malaysia less conspicuous? (This is closer to being true, I think.) Or is it something else? Claude writes that estimates of the scale of smuggling “vary widely, from tens of thousands to perhaps more than a hundred thousand high-end chips per year”. This is literally true, but the low estimates are almost certainly wrong, and the true number is probably much closer to the higher end mentioned by Claude, i.e., hundreds of thousands. 6 So this is misleading. Also, Claude doesn’t specify a year, but smuggling volumes have fluctuated widely since Oc