TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
In a post on Martin Fowler’s website, Fowler describes “LLM voice” as a reader’s sense that prose was generated by a language model, rather than a pattern reliably identified by a detection tool. He advises authors who want to communicate with people to write the prose themselves and disclose AI assistance, while recognizing exceptions such as translation, accessibility and routine documents.
Software author Martin Fowler has argued that writers should be cautious about using large language models to draft prose, saying readers may sense a machine-generated quality he calls “LLM voice.” In an essay on his website, Fowler says the concern is less about detecting AI authorship than about whether writing feels human and communicates a person’s thinking.
Fowler describes a reaction he has when reading some posts: a growing suspicion that a language model produced the text, followed by a reduced desire to continue. He calls the response an “uncanny valley of prose” and says it can arise without a reader being able to identify a specific tell. He dismisses the idea that punctuation alone, such as em dashes, is a dependable sign.
He is wary of tools that claim to distinguish AI-written prose from human writing. Fowler questions their accuracy and says that, by the time a reader turns to a detector, the negative impression may already have formed—even if the tool reports that no AI was used. The essay does not provide test results or evidence comparing detector performance.
Fowler’s practical recommendation is that authors avoid asking an LLM to write a draft if their goal is to communicate with human readers. He says editing a generated draft may not remove its characteristic voice, though he notes it is possible: one example he encountered had been drafted by an LLM and then substantially cleaned up by a professional editor at a major periodical. He says using AI for an outline may be acceptable, but advises authors to write the actual prose themselves.
Why Human-Written Prose Matters
Fowler’s argument concerns more than style. He sees writing as a way to work through ideas: a writer may begin without fully understanding a subject and gain clarity by putting thoughts into words. If people routinely delegate drafting, he warns, they may also delegate part of that process of thinking and comprehension.
The essay also speaks to readers and publishers trying to decide what role AI should play in writing. Fowler’s position is not that every use of language models is harmful, but that polished output may still feel impersonal to some readers. That distinction matters because an AI detector cannot establish whether a piece feels genuine, communicates its author’s thinking or suits its audience.
Fowler proposes transparency as one response. His site’s policy, as described in the essay, asks people who use AI to help write to disclose how they used it, usually in acknowledgments. He says that disclosure can help readers understand the process and learn from particular uses of AI.
Products Worth Considering
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI
Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W
Jackery HomePower 3000 Portable Power Station with Two 200W Solar Panels
As an affiliate, we earn on qualifying purchases.
Fowler’s Boundaries for AI Writing
Fowler distinguishes between generating prose and other uses of AI. He says he uses AI as a search tool, but verifies information by checking the source. He says he does not use it to generate prose or help structure his writing. These are his own practices, not claims about what all writers should or do.
He also gives exceptions to his general advice. Jim Highsmith, a contributor to Fowler’s site, uses an LLM as an essential aid to writing while living with Parkinson’s disease, Fowler says. He accepts that assistance because it enables Highsmith to continue sharing his knowledge. Fowler also describes LLM translation as useful to colleagues who are less comfortable writing in English, while saying he may prefer imperfect human prose to smooth but impersonal machine-generated wording.
For documents that are routine or unlikely to be read by people, Fowler says AI-generated prose may be reasonable. He cites form letters and reports as examples, and raises the possibility that some future documentation may be read mainly by other language models. Michael Nygard, quoted in the essay, describes such communication as being directed to a process rather than to another person.
“There you really aren’t trying to communicate thoughts from one person to other people. You’re communicating with a process rather than a human.”
— Michael Nygard, as quoted by Fowler
Limits of Detecting LLM Voice
Fowler’s essay describes a personal judgment about prose; it does not establish a measurable definition of LLM voice or show how widely readers share his reaction. He also acknowledges that his own judgments may be wrong, and that language changes across generations and communities. Expressions he finds grating may reflect ordinary differences in usage rather than AI authorship.
The essay provides no independent study of AI detectors, no estimate of how often readers correctly identify generated text, and no evidence that editing will always preserve or remove a model’s style. Fowler’s example of an edited AI draft that did not register as machine-written is a single case, not proof of a general rule. The appropriate use of AI will also vary by audience, purpose and accessibility needs.
How Writers May Apply the Advice
Fowler does not announce a policy change, product or formal follow-up. His proposed next step for authors is to make their own prose, use AI selectively where it helps, and disclose writing assistance in acknowledgments. His advice leaves room for tools that support access, translation, research and administrative tasks, while urging care when writing is meant to carry a person’s ideas to another person.
Whether readers and publishers adopt that approach remains uncertain. The practical test Fowler sets out is not whether a detector labels a passage as AI-written, but whether the writing serves its human audience and reflects the author’s own thinking.
Key Questions
What does Martin Fowler mean by “LLM voice”?
He uses the term for a perceived quality in some language-model-generated prose that makes a reader suspect it was written by an LLM. Fowler describes the reaction as an “uncanny valley of prose,” not as a formally defined or measurable category.
Does Fowler say AI detectors can reliably identify generated writing?
No. He says he is wary of tools that claim to distinguish AI prose from human writing, citing doubts about their accuracy. His essay does not present detector tests or performance data.
What does Fowler recommend to authors?
For writing intended to reach people, Fowler recommends that authors write the prose themselves rather than asking an LLM to produce a draft. He says AI-assisted writing should be disclosed and suggests that using a model for an outline may be acceptable, while authors write the text.
Does Fowler oppose every use of AI in writing?
No. He describes exceptions, including AI assistance that helps Jim Highsmith continue writing while living with Parkinson’s disease, translation for colleagues writing in English, and routine documents unlikely to be read by people.
Why does Fowler connect writing with thinking?
He says writing helps him make unclear ideas coherent. He worries that routinely delegating prose generation could also mean delegating some of the thinking and comprehension that writing can develop.
Source: martinfowler.com
