Why one writer says you should almost never let AI write for you

Erich Grunewald argues in an essay published in August that people should almost never use AI to write anything substantive — a blog post, a research report, a memo, a thoughtful email or a novel. His case rests on three claims: the writing process is part of the thinking process, AI prose is vague and wrong in ways that are hard to notice, and passing off AI-written text as one's own is misleading.

Writing is how the thinking happens

Grunewald is not against AI in general. He considers it sensible for transcribing audio, analysing data, searching for information, brainstorming and giving feedback on drafts, and he accepts line and copy editing if a human deliberately accepts or rejects each change. What he objects to is using a model to produce the text itself, even when the writer supplies detailed bullet points or dictated thoughts and edits the result afterwards.

His first reason is that writing is where the thinking happens. He quotes Paul Graham's observation that putting ideas into words is a severe test, and that half the ideas in an essay are ones the author had while writing it. The writer Clara Collier, speaking on Patrick McKenzie's podcast, described how every stage of drafting changes her mind — and noted that an AI given an outline will not stop to consider that the outline itself may be wrong. McKenzie's summary: you make a machine that does the writing and the thinking just pops out as a side effect.

Vague, subtly wrong and very convincing

Grunewald's second argument concerns the texture of machine prose. Asked to write a paragraph introducing AI chip smuggling, the model produced something "not terrible", but he catalogues the small failures: a sentence about export controls that is either obvious or empty; a description of chips as "compact" when what is usually smuggled are servers; and a closing line he calls an uninformative applause light. Each choice, he argues, would have been made differently, and more precisely, by someone who knew the subject.

He cites the philosopher Eric Schwitzgebel, who writes that experts' word choices reflect sensitivities they may not themselves be aware of. Grunewald adds the asymmetry that makes the problem worse: when a model does not know something, it rarely says so, and instead writes something vague enough to be true or convincing enough to pass.

The reader's side of the bargain

His third argument is about trust. Readers expect a text to represent the author's thought, and Grunewald treats that as an implicit contract: the reader gives attention, the writer returns something of value. Unlabelled AI writing breaks it, and he says his trust in a text falls immediately when he suspects it. He accepts that clearly labelled AI output can be fine, and that translating a text you wrote yourself is a reasonable use, preferably disclosed. Working on high-stakes writing with AI, he adds, mostly speeds things up by lowering quality.