Martin Fowler has been writing about software craft for three decades, and this bliki post turns that same precision on the act of writing itself. His claim is simple and firm: when you read prose generated by an LLM, something registers wrong before you can name it. The uncanny valley isn't a technical problem for detection tools to solve — it's a human response, and by the time you're reaching for a classifier, the trust is already gone. The practical advice is blunt. Don't let the LLM write a draft. Don't accept its rephrasing suggestions. Maybe use it for an outline, but fill in every sentence yourself. Fowler notes he's seen exactly one case where an LLM-drafted piece escaped detection — and it required a professional editor at a major publication to scrub the voice clean. That's not a workflow; that's an organ transplant. What elevates this beyond style policing is Fowler's argument about writing as thinking. He describes starting essays without understanding his subject, then using the act of wrestling with sentences to force coherence. Delegating prose to an LLM doesn't just produce worse text — it delegates the thinking. The muscle atrophies. For anyone who writes regularly, this is the real stakes: not whether readers can tell, but whether the writer still has anything to say. Fowler is honest about edge cases and doesn't pretend the rule is absolute. His colleague Jim Highsmith uses LLM assistance to overcome Parkinson's disease. Non-native English speakers use LLM translation to communicate across language barriers. Form letters to bureaucracies don't need a human soul. Michael Nygard's observation — that such communication addresses a process, not a person — is the cleanest framing of when LLM prose is fine: when no human is really on the other end. The self-aware moments are what keep this from becoming a lecture. Fowler admits he can't always tell LLM prose from human prose, acknowledges that language evolves and older writers always grumble about younger ones' diction, and cops to his own transplanted-Brit sensitivities. He declares his own AI policy transparently: search tool only, always verified against sources, no prose generation, no structural help. The piece lands its weight through a simple structural move: it opens with the reader's experience (that creeping revulsion), connects it to a craft argument (writing is thinking), then softens the edges with genuine compassion for people whose circumstances differ. It's a short essay that practices what it preaches — you can hear Fowler's actual voice in every sentence, which is the whole point. Whether you agree that LLM drafts are irredeemable or think Fowler overstates the danger, the core observation is durable: the act of writing is not separable from the act of thinking, and tools that make writing easier may make thinking rarer. That's worth sitting with.
