Why does AI-written prose put readers off? Martin Fowler: write the draft yourself, writing is a tool for thinking
Oct 30, 2026Original Martin Fowler · LLM Voice ↗
In 3 lines
- Fowler calls LLM-sounding prose an "uncanny valley" that drains readers' desire to read on.
- Even after human editing the LLM voice is hard to remove, so he urges authors to fill the prose themselves.
- Writing is a tool for thinking; handing it off hands off comprehension. Translation and accessibility are exceptions.
"As I read, I get the growing feeling I'm reading the output of an LLM, and my desire to read on atrophies, sometimes to the point of revulsion." That is how Martin Fowler, author of Refactoring, opens "LLM Voice," posted on his site on Oct. 29. He argues the reader's reaction comes before any detection tool, and warns that delegating writing also delegates thinking. This piece summarizes the post and the exceptions he allows.
Sometimes you just feel it's an LLM
Martin Fowler says that while reading some posts he gets a growing feeling that an LLM wrote them, and his desire to read on fades. He calls this the uncanny valley of prose. He cannot define it precisely, but points to tells such as "load-bearing tapestries", and says it is not the em dash.
Detectors miss the point
Fowler distrusts tools that claim to tell AI prose from human prose. Partly he doubts their accuracy, but mostly the reaction is a human one. By the time you reach for a tool, it is already too late, even if it says no LLM is present.
Even a veteran doubts his own ear
He does not fully trust his own judgment either. Phrases like "adopting a new sales motion" or "a sharp round-up of news" bother him, but he has heard people say them, so it may be a generational gap rather than carbon versus silicon. Language changes, and complaining about it is a common sign of age.
Don't let an LLM write the draft
What Fowler urges authors most strongly is not to have an LLM write a draft. Even after human editing, the LLM voice is very hard to remove, and accepting rephrasing suggestions is risky too. He knows one case where an LLM-written draft read cleanly, but it took a professional editor at a major publication. Outlines may be fine; authors should fill the prose themselves.
Writing is a tool for thinking
His deeper worry is thinking, not style. Fowler says it is often only by writing that his thoughts become coherent; he has started essays knowing he did not understand something and resolved the confusion by wrestling with words. "As more people delegate writing to LLMs they will also delegate thinking and comprehension." That applies to design docs and PR descriptions too.
Say your writing, now more than ever
Fowler revisits his old tip, "Say Your Writing": read the draft aloud, or at least move your lips, which engages the part of the brain sensitive to blandness. In a footnote added a year later he calls LLMs the new biggest source of flab, and says the more LLM-speak spreads, the more it is worth sounding like an actual human.
Martin Fowler · martinfowler.com ↗The exceptions are clear
He stresses this is not absolute. He will gladly read the work of his friend Jim Highsmith, who relies on an LLM to write despite Parkinson's disease. Translation for colleagues writing in a second or fourth language is a harder case, though he still prefers a human's clunky prose to smooth but soulless text. Form letters and reports nobody will read are reasonable to delegate.
Disclose how you used AI, and check numbers at the source
Fowler's site policy is that anyone who used AI to help write should say how, usually in the acknowledgments, for transparency and so readers can learn useful practices. He uses AI only as a search tool and verifies at the source. Numbers show the same issue: Trakkr traced 1,285 statistics cited by five AI assistants, and only 40% led to a study saying the same thing.
Martin Fowler · Trakkr · martinfowler.com · trakkr.ai ↗
