The Case Against Letting AI Write Your Prose

Erich Grunewald argues that AI should almost never be used for the actual act of writing—composing the sentences meant to convey an argument, analysis, or substantive idea—even when the model is fed detailed bullet points and a human edits the output afterward. He draws a sharp line between this and other uses he endorses: transcription, data analysis, search, brainstorming, feedback on drafts, and even line or copy editing, provided a human deliberately accepts or rejects each change. The objection is narrow but firm, and scoped to current and near-future models.

Three reasons anchor the case. First, writing is thinking: forcing an argument into sentences and paragraphs exposes gaps in reasoning, weak transitions, and inconvenient counterarguments that an outline handed to an LLM would gloss over. He cites Paul Graham, Clara Collier, and Patrick McKenzie to the effect that much of an essay’s substance is discovered during the act of writing it—and that outsourcing the prose lets you skip the necessary thinking. Second, AI writing is vague and wrong in ways that are hard to catch; he dissects a Claude-generated paragraph on AI chip smuggling, flagging empty truisms (“export controls are only as strong as enforcement”), a misleading emphasis on chips being “compact” when smuggling actually relies on relabeling servers, and a scale estimate whose low end is almost certainly wrong. Third, publishing AI-written text without labeling it is misleading and rude to readers.

Grunewald concedes AI writing is faster and less effortful, so the disadvantages have to be large to outweigh those gains—and he believes they are. He allows that future models may eventually be good enough to delegate writing to, but notes that at that point it would likely make more sense to hand off the entire research-and-writing process, since competent writing depends on the same thinking the models would then be doing anyway.