Almost Never Use AI to Write Anything Substantive

A clear argument against using AI to generate any substantive writing - essays, reports, memos, novels, or analytical pieces - because doing the writing is itself a key part of thinking. Three core objections are offered: writing forces the author to discover holes in evidence and argument; current AI outputs are densely filled with vague, misleading, or subtly wrong phrasing that’s hard to spot; and presenting AI-written work as if it were authored by a human is deceptive and discourteous. Useful AI roles are acknowledged (transcription, data analysis, searching, brainstorming, commenting, and line editing where humans accept or reject edits), but the claim is that delegating the actual composition short-circuits necessary cognitive work. Future models could change the calculus, but not the present ones.

Concrete examples illustrate the problem: a short AI paragraph on AI chip smuggling contains many small but consequential defects - ambiguous causal claims about enforcement, incorrect emphasis on chip compactness when servers are often moved, unclear or unsupported claims about supply chains and smuggling scale, missing numeric context for enforcement agencies, and empty phrases about “compute” and “capable” systems. Experts’ word choices carry nuanced signals that LLMs don’t reproduce, and passively approving machine text reduces the active, error-finding thinking humans do when authoring prose. The net effect is plausibly persuasive but potentially misleading output that undermines careful reasoning.