- Author
- Martin Fowler
- Published
- 29 October 2026
- Kind
- Short personal opinion post on a bliki, with footnotes
- Original
- martinfowler.com/bliki/LLM-Voice.html
- Subject
- Writing with models
What it is
Martin Fowler writes on the Fowler site about a reaction when reading. Some prose feels like language model output, and the wish to read on fades. The post sets out what Fowler asks of authors, and the exceptions.
What it says
- Fowler says machine-sounding prose can cause revulsion in readers, even without clear tells.
- Fowler is wary of detection tools, doubting their accuracy and saying they miss the point.
- Fowler doubts the reliability of Fowler's own ear, since it may reflect a generational difference.
- Fowler urges authors not to let a model write a draft, or even rephrase.
- Fowler says model outlines may be acceptable, but authors should fill the prose.
- Fowler argues that writing is a tool for thinking, so delegating writing also delegates thinking.
- Fowler allows exceptions: a colleague with Parkinson's disease, translation, form letters, and unread reports.
What evidence it gives
The post offers argument and personal experience only. It cites one case of an LLM-drafted piece with no LLM voice. A professional editor cleaned that piece up. It includes a comment from Michael Nygard. It has no data or tests.
Interests and limits
Fowler publishes on a personal site, so the post favors human authorship. The post is an opinion and tests nothing. It does not define LLM voice beyond impression. It says the advice applies right now. It calls undetectable model prose more worrying.