English

PLUS ULTRAErich Grunewald

Writing with AI Risks Losing Thinking Process and Clarity, Argument Says

PLUS ULTRA by Amenoyomi

Erich Grunewald, a researcher at the Institute for AI Policy and Strategy (IAPS), argues that AI should almost never be used to write substantive text, including blog posts, research reports, or memos. While he acknowledges AI's utility in research tasks like data analysis or brainstorming, he claims that using it to generate prose compromises the quality of both the text and the underlying thought process.

Grunewald identifies three primary issues with AI-generated writing. First, he asserts that the act of writing is an essential part of the thinking process; forcing oneself to structure arguments into sentences helps surface logical gaps and contradictions. Second, he notes that AI-written text often contains subtle errors and vague phrasing that can be difficult for humans to detect. Third, he considers writing with AI without disclosure to be misleading to readers.

The argument emphasizes that even when provided with detailed outlines or context, AI lacks the ability to critically evaluate the quality of the ideas being expressed. While AI may serve as a useful tool for copy editing or translation, Grunewald contends that delegating the actual construction of ideas to a model prevents the writer from truly mastering the subject matter.

PLUS ULTRAby Amenoyomi

The act of writing serves as a severe test of a writer's understanding. Forcing ideas into specific sentences and paragraphs often reveals logical holes, contradictory evidence, or inadequate research that were not apparent during the outlining stage. As noted by Paul Graham and Clara Collier, the process of struggling with transitions or word choices frequently exposes that two points should not be juxtaposed or that a central argument is flawed, forcing the writer to change their mind and refine their beliefs in real time.

When AI handles this process, it often produces text that is maximally convincing yet subtly wrong. For instance, in a generated paragraph about AI chip smuggling, a model might describe chips as compact—which is true—while ignoring that the actual smuggled items are often large AI servers. It may use phrases like "sprawling global supply chains" or "enough compute to matter," which sound professional but provide no actual informational value. These "uninformative applause light" expressions create a facade of competence that masks a lack of specific, verifiable detail.

This creates a dangerous cognitive gap between active generation and passive review. There is a significant difference between the effort required to construct a sentence from scratch and the act of nodding along while reading a generated draft. When a human writes, they are forced to resolve every ambiguity to move forward. When reviewing AI text, the writer is prone to let "approximate" wording suffice, bypassing the active thinking required to choose the precise word. This passive acceptance allows the writer to avoid the cognitive load necessary to truly master the subject matter.

Sources

  1. I think you should almost never use AI to write (Hacker News Frontpage, 2026-09-19)
  2. Erich Grunewald