In a Substack essay, Eric Grunewald argues that people should “almost never” use AI to write substantive content, including blog posts, research reports, memos, and analyses—even with detailed prompts or heavy editing.
Grunewald outlines three main concerns. First, he contends that writing itself is an integral part of thinking. He cites Paul Graham’s observation that writing about something reveals gaps in understanding, with many ideas emerging during the writing process itself. Writer Clara Collier notes that substantive writing involves constant thinking and revision at every stage—outline, draft, and sentence level—and that feeding an outline to an AI won’t prompt reconsideration of the outline’s validity. Grunewald agrees that forcing oneself to structure arguments in writing surfaces logical holes, inadequate research, and contradictions that writers must then ethically address.
Second, Grunewald argues that AI-generated text contains subtle errors and vagueness that are difficult to detect. He demonstrates this with an AI-generated paragraph on chip smuggling, showing how Claude made numerous problematic choices: writing that “export controls are only as strong as enforcement” (either obvious or meaningless), calling AI chips “compact” when servers are typically smuggled (and this doesn’t match how smuggling usually occurs), using vague language about supply chains, and presenting widely varying estimates in a way that masks that lower estimates are “almost certainly wrong.”
Third, Grunewald raises an ethical concern, arguing that using AI to write without disclosing it is misleading to readers.
Grunewald notes he is not anti-AI and supports using AI for research and writing support tasks: transcribing audio, analyzing data, searching for information, brainstorming, and providing feedback on drafts. He also endorses using AI for copy editing and line editing when a human deliberately accepts or rejects each change.
He acknowledges the advantages of AI writing—it requires less effort and is faster than manual writing. However, he argues the disadvantages are substantial enough to outweigh these benefits for current and near-future AI models. He speculates that sufficiently advanced future models might change this calculus, though at that point end-to-end delegation of research and thinking might become relevant.
Key facts
- Grunewald argues writing forces necessary thinking by surfacing logical gaps, contradictions, and research inadequacies
- AI-generated text contains subtle errors and vague phrasing that readers may not notice, as demonstrated in an AI paragraph on chip smuggling
- Grunewald supports using AI for supporting tasks like transcription, data analysis, brainstorming, and copy editing with human review
- The author does not oppose all AI use in writing, but specifically opposes delegating substantive text generation to AI
- Grunewald acknowledges potential future AI models might justify different conclusions