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, not a reliably detectable technical signature. He advises writers who want to communicate with people to write their own prose, while recognizing exceptions such as accessibility needs and translation.
Software author Martin Fowler has argued that writers should be cautious about using large language models to draft prose, saying readers may sense an impersonal quality he calls “LLM voice.” In an essay on his website, Fowler presents the reaction as a matter of human perception rather than a reliable test of whether a machine wrote a piece, and says writers seeking to communicate with people should generally put the prose into their own words.
Fowler says he sometimes begins to suspect that a post was generated by an LLM but cannot always explain why. He describes a growing reluctance to continue reading, which can reach what he calls an “uncanny valley of prose.” He cautions that recognizable language habits are not proof of AI authorship: some phrases readers find artificial may also reflect generational or regional differences.
He is skeptical of tools that claim to distinguish machine-written prose from human writing, citing doubts about their accuracy and arguing that detection misses the point. By the time a reader seeks a verdict from a tool, he says, the negative response to the writing may already have occurred. The source provides no test results or independent evidence establishing how accurately such tools identify AI-generated text.
Fowler recommends that authors not ask an LLM to produce a first draft if their goal is to communicate with human readers. Even after human editing, he says, the model’s style can be difficult to remove. He allows that using AI to develop an outline may be acceptable, but says authors should write the prose themselves. He also acknowledges that an entirely AI-generated first draft can be successfully edited; his example required a professional editor at a major periodical.
Why Human-Drafted Prose Matters
Fowler’s essay addresses a practical question for writers and publishers as language models become part of everyday writing workflows: whether an AI-assisted draft still sounds like its human author. His argument is that the issue is not limited to whether readers can identify a model’s work. A reader’s response to the language can affect whether they continue, regardless of any detection result.
He also links writing to thinking. Fowler says that putting thoughts into words helps him clarify what he understands, and warns that delegating prose may also mean delegating some of that work of comprehension. This is his view, not a finding supported in the post by a study. It nevertheless frames the stakes as broader than style: for writers, drafting can be part of working out an idea, not simply packaging it for an audience.
The argument has limits in Fowler’s own account. He recognizes that model assistance can enable people to communicate when they face barriers, and that some documents are directed at processes rather than human readers. His position is a recommendation about purpose and authorship, not a universal ban on using AI.
Products Worth Considering
Jackery HomePower 3000 Portable Power Station with Two 200W Solar Panels
DJI Osmo Pocket 3 Vlogging Camera with 1-Inch CMOS Sensor and 4K Video
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Fowler’s Exceptions to the Rule
Fowler distinguishes between writing meant to convey a person’s thoughts to other people and text whose primary audience may be a system. He says AI may be reasonable for routine form letters or reports he expects few people to read. He also relays a comment from Michael Nygard: in some such cases, the writer is communicating with a process rather than another person.
He identifies accessibility and language as further exceptions. Fellow contributor Jim Highsmith, Fowler says, relies on working with an LLM to continue writing while living with Parkinson’s disease. Fowler says he would accept the possibility of “LLM voice” in order to keep reading Highsmith’s work. He also notes that colleagues who are not comfortable writing in English have used LLM translation to communicate with English speakers.
Fowler says his own site asks contributors to disclose how they used AI, usually in acknowledgments, both for transparency and to help readers understand possible uses. He says he uses AI as a search tool for writing, then checks information against its source, and does not use it to generate prose or structure his work.
“You’re communicating with a process rather than a human.”
— Michael Nygard, as quoted by Fowler
Limits of Detecting AI Style
Fowler’s essay is a personal account and recommendation, not a controlled study of reader responses or AI-detection accuracy. It does not identify a consistent set of language features that proves a text was written by a model, and Fowler explicitly doubts his own ability to recognize AI style reliably.
The source also does not quantify how often readers experience the reaction he describes, how much editing is needed to change a model-generated draft, or whether particular writing methods reliably avoid it. The boundary between AI assistance and authorship remains dependent on the task and the writer’s choices in the examples Fowler discusses. His post does not claim that all AI-assisted writing is alienating or that every reader responds in the same way.
How Writers May Apply the Advice
Fowler’s immediate advice is for writers to decide whether they are communicating with people or preparing text mainly for a system, then choose their tools accordingly. For human-facing work, he recommends drafting the prose themselves and disclosing AI assistance when it has been used, following the policy he describes for his site.
The source announces no new policy, study or follow-up deadline. The questions left for writers and publishers include how to describe different forms of AI assistance, how much editing can make a model-drafted text sound like its author, and how to balance stylistic concerns with access needs such as translation and disability support.
Key Questions
What does Martin Fowler mean by “LLM voice”?
He uses “LLM voice” for a reader’s impression that prose has an impersonal quality associated with language-model output. Fowler presents it as a subjective response, not a definitive way to identify who or what wrote a text.
Does Fowler say AI-detection tools can reliably identify model-written text?
No. Fowler says he is wary of those tools, both because he doubts their accuracy and because a reader’s reaction may happen before a tool is consulted. His essay does not provide comparative test data.
Does Fowler oppose all uses of AI in writing?
No. He discusses exceptions including translation, disability access, routine letters and text intended mainly for an automated process. His main recommendation is that authors write the prose themselves when they want to communicate with human readers.
What does Fowler say about using an LLM for an outline?
He says using a model to help develop an outline or structure may be acceptable, but advises writers to produce the prose themselves. He does not present this as a universal rule for every writing task.
Source: martinfowler.com
