Martin Fowler wants nobody to let an LLM write the first draft

Delegating the draft delegates the thinking

Martin Fowler's note says "LLM Voice" triggers a human reaction he calls an uncanny valley of prose. He doubts his ability to detect it, since some phrasings he dislikes are also used by people and the gap may be generational. His rule: don't have an LLM write a draft, because even heavy editing struggles to remove the voice, and accepting rephrasing suggestions is risky. An outline may be fine. Contributors to his site must declare AI help, and he accepts the voice from Jim Highsmith, who has Parkinson's. Low-stakes forms nobody reads may be fine to delegate.

The thread: cooking, critics, and a prose-over-substance complaint

"I enjoy cooking from scratch, and I really don't like frozen foods, microwave/TV dinners, or even DoorDashing but we're past complaining/arguing/shaming folks who don't cook every meal from scratch" — yawnxyz, who designs and prototypes AI interfaces

A reader who calls themselves aggressively pro-AI agreed with Fowler's symptom anyway: everything sounds identical and rambles, and they stop reading partway (a self-described pro-AI reader).

The practical tip came from tptacek:

"LLMs are exceptionally useful writing critics if you know how to hold them correctly, but you can't use a single word they suggest to you" — tptacek

That splits the issue: Fowler's objection is to LLM output in the draft, while tptacek's workflow keeps the model in the critic's seat and the words human.