LLM Voice

What it claims

Martin Fowler argues, in a bliki post dated 29 October 2026, that readers react to LLM-written prose before they can explain why. He says the reaction itself is the problem. For that reason he is wary of tools that claim to detect LLM text. He doubts their accuracy, but he also says a tool is beside the point: by the time a reader reaches for one, the reader has already lost interest. A "no LLM present" verdict does not undo that.

Fowler says editing an LLM draft rarely removes the voice. Accepting LLM rephrasing suggestions carries the same risk. He knows of one LLM-drafted piece that showed no LLM voice in its final form, and it needed a professional editor at a major periodical to get there. He also doubts his own detector. Some phrasing he dislikes, such as "adopting a new sales motion," he hears from people too. He reads that as a generational difference, not a machine signal.

His second argument is about thinking. Fowler says writing is how he makes his thoughts coherent, and that people who hand writing to LLMs also hand over thinking and comprehension. He finds this more worrying if LLM prose ever becomes indistinguishable from human prose.

How to use it

Fowler recommends "Say Your Writing": speak the words in your head, write them down, then edit.

His site's disclosure policy asks contributors to state how they used AI, usually in the acknowledgments. He gives two reasons: transparency, and letting readers learn useful ways to use AI. His own use is limited to search, and he checks every result against the original source. This post states that no AI was used to write it.

Limits

The exceptions are narrower than "allowed." Fowler accepts LLM voice from Jim Highsmith, who contributes to the site and uses an LLM to write despite Parkinson's disease. Fowler calls translation "awkward": he still prefers a non-native speaker's clunky prose to smooth LLM output. Form letters and reports nobody will read are fine. Michael Nygard's note puts the line here: in those cases the writer is talking to a process, not to people.

The piece is opinion. It offers no data on how often readers notice LLM text or on how accurate detection tools are.