LLM voice describes the feeling that prose may have been generated by a language model and therefore sounds unnatural. Martin Fowler argues that readers may not be able to point to a specific phrase, yet the impression can still diminish their desire to continue. Detection tools may not solve the problem: a reader's reaction often arrives before any test result.
Why LLM voice is hard to identify
Fowler cautions that language shifts across generations and communities, so expressions people dislike or find unfamiliar do not reliably distinguish human writing from model output. He therefore treats the impression of sounding like AI as an uncertain signal, and does not see detectors as a substitute for readers' judgment. His central concern is less about proving who wrote a passage than whether it carries a person's thinking and character.
Where AI writing can help, and where it falls short
He advises caution about asking a model to draft prose: even after human editing, its original manner of expression can be difficult to remove. Using a model to help shape an outline may be acceptable, he suggests, but authors should write the prose themselves and disclose how AI contributed. There are limits to this rule: Fowler notes one case in which substantial professional editing left an AI-written first draft without an obvious model voice.
Fowler also recognizes cases where AI-generated prose may be reasonable: translation, helping a writer with a disability communicate, routine official correspondence, or reports intended mainly for a process rather than a human reader. Purpose matters. When writing is meant to convey one person's thoughts to another, composing the words can also help the author clarify ideas and keep practicing. Whether future models will become indistinguishable from people, and whether machine-read documentation changes the judgment, remain open questions.