Substack writer Eric Grunewald argues that people should almost never use AI to write substantive texts — including blog posts, research reports, memos, emails, and novels — even with detailed prompts and heavy editing afterward.
Grunewald identifies three main problems with AI writing. First, he contends that writing itself is essential to the thinking process. He cites essayist Paul Graham’s observation that “writing about something, even something you know well, usually shows you that you didn’t know it as well as you thought,” and notes that many ideas in essays emerge during the act of writing itself. Similarly, podcast host Patrick McKenzie argues that “the writing process is the thinking process,” and that forcing yourself to structure arguments on the page reveals gaps in research and contradictions in your reasoning.
Second, Grunewald claims that AI-written text contains subtle errors and vagueness that are hard to detect. He demonstrates this with a Claude-generated paragraph about AI chip smuggling. While the paragraph sounds reasonable, Grunewald identifies multiple problems: the statement that “export controls are only as strong as their enforcement” borders on tautological; the description of chips as “compact” misses that servers are often smuggled instead; and the claim that vague “sprawling global supply chains” make smuggling attractive lacks specificity. Most significantly, Grunewald notes that while Claude presents estimates of smuggling volumes ranging from “tens of thousands to perhaps more than a hundred thousand” chips annually, the lower estimates are “almost certainly wrong,” making the range itself misleading.
Third, Grunewald argues that using AI for writing without disclosure is “rude and misleading” to readers.
However, Grunewald clarifies that he is not anti-AI. He supports using AI for other parts of research and writing: transcribing audio, analyzing data, searching for information, brainstorming, and providing feedback on drafts. He also accepts AI for line editing, copy editing, and rewriting passages for clarity — as long as a human deliberately accepts or rejects each change.
Grunewald emphasizes that his position applies only to current and near-future AI models. He acknowledges that sufficiently advanced models might eventually warrant delegating writing tasks, though he suggests that such models might be better used to handle entire research processes end-to-end.
Key facts
- Grunewald argues writing is essential to thinking; the act of structuring ideas reveals gaps in research and reasoning
- AI-generated text contains subtle errors and vague phrasing that are difficult to detect but undermine clarity and accuracy
- Grunewald supports using AI for transcription, data analysis, information search, brainstorming, and editing — just not for generating substantive text
- The argument applies only to current and near-future AI models, not hypothetical advanced systems