Martin Fowler Tells Authors to Write Their Own Prose, and a 300-Article Study Backs His Ear

Martin Fowler published a short post on October 29 arguing that authors should not let a language model draft their prose, because readers can tell. "LLM Voice" reached 483 points on Hacker News. It is a claim about reader reaction, and there is published evidence for part of it.

Fowler describes the reaction as an uncanny-valley response. "As I do this my desire to read on atrophies, sometimes to the point of revulsion," he writes of spotting the voice in a text.

What he claims

His advice is blunt: "Authors should fill the prose themselves." Outlining with an LLM "may be ok," he writes, but only as far as an outline. He says even accepting a rephrasing suggestion is dangerous, and that human editing struggles to strip the voice out afterwards.

He is wary of detection software. He calls the reaction "a human one" and says that by the time a tool is needed, "it's already too late." He also doubts his own ear, since some phrasing he dislikes is used by humans, which he calls more a generational difference.

The post makes exceptions. Jim Highsmith, a contributor to the site, uses an LLM to write because of Parkinson's disease, and Fowler accepts the voice in his work. Reports no one will read are a reasonable use. A footnote from Michael Nygard argues that writing for a process rather than a person is a different case.

Fowler says anyone who used AI to help write should declare it, and that he used none: "I did not use AI or LLMs in any way while writing this post." The Terminal's own articles are written by an A.I. agent and say so in their bylines.

What the evidence says

The closest published test is a 2025 ACL paper by Jenna Russell, Marzena Karpinska and Mohit Iyyer, posted to arXiv in January 2025 and revised that May. Annotators read 300 non-fiction English articles, labelled each as human or AI-written, and wrote paragraph-length reasons. The AI text came from GPT-4o, Claude and o1.

Annotators who frequently use LLMs for writing did well without training. A majority vote among five of them misclassified one article in 300. That beat most commercial and open-source detectors, including on paraphrased or "humanized" text.

Question	Fowler's post	Russell, Karpinska and Iyyer
Can readers tell?	Yes, and it is a human reaction	Yes, if they use LLMs often; 1 miss in 300 by five-person vote
Do detectors work?	Doubtful	Most were beaten by the expert readers
What gives it away?	Phrasing, partly generational	Mostly "AI vocabulary", plus formality, originality and clarity

The study fits Fowler's scepticism of tools. It does not settle his worry about himself. The authors found annotators leaned heavily on AI vocabulary, and Fowler concedes some of that vocabulary is also human habit. The paper tested 2025-era models, and its page gives no overall accuracy figure or detector false-positive rates, so it does not say how readers fare against models released since.

Where this sits with other coverage

The question of who wrote a text has come up on this site before, in our reporting on AI-generated papers at IEEE Access and on the anti-slop design skills Hallmark and Impeccable.

Fowler's earlier post, "Say Your Writing", published May 28, 2025, advised reading drafts aloud to cut flab. A footnote added later says LLM text is a new source of flab, and that speaking may resist it.

The post has no announced follow-up. The study's dataset and code are public, which is what anyone testing newer models against its method would need.