Erich Grunewald contends that the act of writing is fundamentally intertwined with thinking, arguing that producing text forces the writer to examine and refine ideas. He supports this view by quoting Paul Graham’s observation that writing about a topic often reveals what the writer did not truly know, noting that many ideas emerge only during the writing process. He further cites an episode of Patrick McKenzie’s podcast where Clara Collier describes how, while writing something substantive, she continuously reassesses her outline, transitions, and sentences, discovering flaws in her reasoning that would remain hidden if she merely consulted an outline. According to Grunewald, outsourcing this process to an AI bypasses this essential feedback loop, allowing inadequate research or unclear arguments to go unchallenged.
The image shows the author of the article, Erich Grunewald, who presents the argument that writing is tied to thinking and that AI undermines this process.
Grunewald’s second argument is that AI‑generated text tends to be vague and contain subtle errors that are difficult to detect without close reading. To illustrate, he prompted Claude Fable 5 to write a short paragraph on AI chip smuggling. While the output appears reasonable at first glance, he points out multiple issues: the model states that export controls are “only as strong as enforcement” without clarifying what that means, describes AI chips as “compact” when the smuggled items are often whole servers, and presents smuggling estimates that are misleadingly low. He also notes that the model fails to specify a timeframe or define what a “high‑end” chip is, and that its description of the Bureau of Industry and Security’s budget and headcount lacks the concrete numbers needed to assess whether the agency is truly under‑resourced. According to Grunewald, these kinds of inaccuracies are typical of AI writing and require significant effort to spot.
Finally, Grunewald argues that using AI to write text intended for others to read—without disclosing its involvement—violates an implicit contract between writer and reader. He explains that when a piece is shared, the audience expects the text to reflect the writer’s own thought and effort, offering them information or entertainment they could not obtain elsewhere. Referencing the same podcast episode, Clara Collier notes that a text produced by an LLM would always lack something the human writer could contribute, even if the model knows more about the topic. Grunewald concludes that presenting AI‑generated prose as one’s own work is therefore rude and misleading, because it deprives readers of the genuine intellectual contribution they anticipate.