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
Martin Fowler has published an essay arguing that AI-generated prose can trigger an “uncanny valley” reaction in readers, even when they cannot identify why. He advises writers who want to communicate with people to draft the prose themselves, while acknowledging uses such as translation, accessibility and routine form letters.
Software writer Martin Fowler has published an essay arguing that prose drafted by a large language model can leave readers with an uncanny sense that the writing is machine-generated, even when they cannot pinpoint the reason. He urges authors writing for people to compose the prose themselves, warning that relying on a model for a first draft can be difficult to undo through later editing.
Fowler describes a personal reaction to some posts: as he suspects they were written by an LLM, his desire to continue reading fades. He says the feeling is not necessarily tied to a single tell, and cautions that commonly cited signals may be misleading. His essay characterizes the response as a human judgment about prose, rather than a classification problem that an automated detector can settle.
He is wary of tools that claim to distinguish AI-written text from human writing, citing doubts about their accuracy and arguing that a detector misses the point. By the time a reader seeks a tool, he says, the writing may already have put them off; a result that says no AI was involved would not necessarily change that reaction. He also doubts his own ability to identify AI prose reliably, noting that language changes over time and that some expressions he dislikes are used by people as well as machines.
Fowler’s practical advice is to avoid asking an LLM to write a draft when the goal is communication with human readers. He says a human editor can sometimes remove the machine-like quality, but describes that as difficult and cites one case that required professional editing. He considers using a model for an outline potentially acceptable, but recommends that authors write the actual sentences themselves and disclose how they used AI.
Why Fowler Favors Human Drafts
The essay shifts attention from whether software can detect AI authorship to how readers experience writing. If prose feels generic or emotionally distant, a detector’s verdict may have little bearing on whether an audience wants to keep reading. That makes reader response, not just authorship labels, part of the practical debate over AI-assisted writing.
Fowler also argues that writing is a method of working through ideas, not simply a way to package them. He says composing an essay can reveal gaps in a writer’s understanding and help bring thoughts into focus. In his view, delegating prose generation can also mean delegating part of the thinking, potentially limiting both communication and the development of writing skills through practice.
Those concerns matter to writers, editors and publishers deciding where AI fits into their work. Fowler’s position is advice grounded in his experience, not evidence that all AI-assisted prose will put readers off or that human writing is always preferable. Still, his distinction between drafting for human readers and producing text for a process offers a concrete way to frame choices about when to use a model.
Products Worth Considering
DJI Osmo Pocket 3 Vlogging Camera with 1-Inch CMOS Sensor and 4K Video
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
As an affiliate, we earn on qualifying purchases.
Where Fowler Draws the Line
Fowler’s essay does not present avoiding AI as an absolute rule. He makes an exception for Jim Highsmith, a contributor to his site who has Parkinson’s disease and finds working with an LLM essential to writing. Fowler says he accepts the machine-like quality that might result because it allows him to continue reading Highsmith’s work. That example makes room for accessibility and individual circumstances in his argument.
He also discusses translation, noting that colleagues who use English as a second or fourth language may rely on LLMs to communicate with English-speaking coworkers. Fowler says he may prefer imperfect human prose to smooth but impersonal machine-generated text, while recognizing that translation can serve a real communication need. For routine form letters or reports that are unlikely to be read by people, he says using an LLM may be reasonable.
For transparency, Fowler says contributors to his site should disclose how they used AI, usually in acknowledgments. He says his own use is limited to searching for information, which he then checks against original sources; he does not use AI to generate prose or structure. The essay therefore distinguishes writing for people from text produced mainly for administrative purposes or machine consumption.
“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 in Fowler’s essay
Limits of Spotting AI Prose
The essay does not establish how often readers experience this reaction, how consistently they can recognize AI-generated writing, or whether particular writing patterns reliably signal model use. Fowler expressly questions his own ability to identify “LLM Voice,” and the source provides no measured accuracy data for AI-detection tools or evidence comparing audience responses to human and AI-assisted writing.
It is also unclear how much editing is needed to remove the quality Fowler describes, or whether the effect varies by model, genre, language or reader. His example of an AI-written first draft successfully edited by a professional editor is an individual case, not proof that editing will reliably solve the problem. The essay is an argument based on observation and experience, not a controlled study.
How Writers May Apply the Advice
Fowler’s essay sets out recommendations rather than announcing a policy change or research program, and it identifies no formal next milestone. Writers using AI can apply his approach by deciding whether a document is meant to reach human readers, limiting assistance to tasks such as search or outline development when appropriate, and disclosing AI use where their publication or workplace asks for transparency.
For readers and editors, the unresolved issue is how to balance fluent, efficient text with voice, accessibility and the value of writing as a thinking process. Fowler’s position leaves that judgment to the circumstances: he favors human-written prose for direct communication with people, but accepts AI help in cases such as translation, disability support and routine documents.
Key Questions
What does Martin Fowler mean by “LLM Voice”?
He uses the phrase for a quality in some prose that can make readers suspect it was generated by a large language model, even when they cannot identify a specific signal. He describes the resulting reaction as an uncanny feeling, not a reliable test of authorship.
Does Fowler say AI detectors can identify machine-written text reliably?
No. Fowler says he is wary of tools that claim to distinguish AI prose from human writing, citing doubts about their accuracy. His essay provides no detector test results or accuracy figures.
What does Fowler recommend to authors?
For writing intended to communicate with people, he recommends composing the prose themselves rather than asking an LLM to produce a draft. He says using a model for an outline may be acceptable, and advises writers to disclose how they used AI.
Does Fowler oppose every use of AI in writing?
No. He describes exceptions, including AI support for a writer with Parkinson’s disease, translation for colleagues who are less comfortable in English, and routine documents unlikely to be read by people. His argument is a preference for human-written prose in communication with human readers, not a blanket ban.
Source: martinfowler.com
