Why You Should Almost Never Use AI to Write Anything Substantive
-
The Core Thesis:
- Generative AI should almost never be used to produce the actual prose of substantive writing—such as essays, analytical memos, research reports, or thoughtful emails.
- The author grounds this argument in three central reasons:
- Writing is an indispensable mechanism of thinking.
- AI text is chronically vague and subtly incorrect in hard-to-spot ways.
- Publishing or sending uncredited AI prose to others is disrespectful and misleading.
-
Writing as Cognition, Not Just Communication:
- Substantive work requires a human author to hold the entirety of an argument, evidentiary nuance, and logical links in their mind.
- Outsourcing sentence generation short-circuits this cognitive process, allowing writers to dodge the difficult, necessary thinking that only occurs while struggling to put ideas into words.
- Reviewing or editing an AI draft creates a passive illusion of competence; writers reflexively nod along to plausible-sounding text rather than rigorously probing concepts.
-
The Danger of Plausible, Subtly Flawed Prose:
- AI-written text rarely fails loudly; instead, it is packed with pseudo-profound platitudes, tautologies (e.g., "export controls are only as strong as enforcement"), uninformative filler, and misleading approximations.
- Even when fed extensive notes or outlines, models distort domain-specific context (e.g., mischaracterizing chip smuggling methods or confusing components with servers) in ways that require excessive cognitive effort to detect and fix.
-
Social Trust, Signal, and Deception:
- Presenting AI-generated text as your own signals to the reader that you understand the material and stand behind every stylistic and substantive choice when you do not.
- Readers invest time expecting an authentic human synthesis; offloading the writing betrays that mutual social contract.
Hacker News Discussion
-
Communication as a Proof-of-Understanding Signal:
- A central theme among commenters was that human writing serves not merely as data transmission, but as proof that the sender has deeply understood and synthesized the problem.
- Sending AI-generated summaries or memos undermines trust because it masks the sender's actual competence, leaving readers unsure if anyone truly comprehends the underlying message.
-
Editing AI Slop vs. Writing from Scratch:
- Multiple commenters cited Eric Schwitzgebel's observation that reading and tweaking existing text is cognitively passive compared to de novo composition.
- Practitioners noted that debugging, fact-checking, and stripping out the verbose, generic "texture" of AI drafts often requires more time and effort than writing manually from an outline.
-
Prose vs. Code and Objective Evaluation:
- Several developers contrasted using AI for coding versus writing: code comes with a deterministic test harness and runtime feedback to verify correctness objectively.
- In contrast, prose lacks an automated test suite, meaning subtle logical drifts, hollow phrasing, and false nuances easily slip past casual human inspection.
-
Nuance for Low-Stakes & Non-Native Writing:
- While agreeing on substantive essays, some participants defended LLMs for tedious bureaucratic tasks (e.g., Jira tickets, boilerplate forms) where creative thought is absent.
- Commenters also highlighted the legitimate value of LLMs for non-native speakers seeking grammar refinement, though many cautioned that models frequently flatten authentic personal voice into sterile corporate jargon.

