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 his website, software writer Martin Fowler describes readers’ growing sense that some prose was produced by a large language model, calling the reaction an “uncanny valley” rather than a problem that detection tools can solve. He advises writers to draft prose themselves and disclose AI assistance, while recognizing uses such as translation, accessibility support and routine forms.
Software writer Martin Fowler has argued that prose produced with large language models can put readers off, even when they cannot identify exactly why, and advises authors to write the prose themselves when they want to communicate with people. In an essay titled “LLM Voice,” Fowler says AI-detection tools miss the point: the reader’s reaction may already have set in before a tool is consulted.
Fowler describes a sense that develops while reading: suspicion that a post came from an LLM, followed by a diminishing desire to continue. He calls the effect an “uncanny valley of prose”. The source essay does not offer a formal test or claim that readers can reliably identify AI-written text. Fowler says he is wary both of detection tools’ accuracy and of treating detection as the central issue.
His main recommendation is that writers should not ask an LLM to produce a draft and then assume editing will remove its stylistic imprint. Fowler says it can be very hard to expel LLM voice from an AI-written draft, though not impossible. He recounts one example in which a fully AI-generated first draft did not register as LLM-like in its final form, but required work by a professional editor at a major periodical.
Fowler allows that AI may help with an outline, but says authors should write the prose themselves. He also recommends disclosing how AI was used, a policy he says applies to contributors to his site. He says his own use of AI for writing is limited to search, with information checked against its original source; he does not use it to generate prose or develop structure.
The Stakes for Human Readers
Fowler’s argument focuses on reader trust and connection, rather than on whether software can label a passage as AI-generated. For people publishing essays, commentary or other writing intended for a human audience, his point is that a polished piece can still feel impersonal. If readers sense that quality, they may disengage regardless of how the text was produced or edited.
He also links the choice of writing process to thinking and understanding. Fowler says writing helps him make ideas coherent, including when he begins an essay unsure of his own view. He warns that delegating prose to an LLM could also mean delegating some of that thinking. That is his concern and interpretation, not a measured finding about what happens to all writers who use AI.
The essay does not call for a blanket ban. Fowler recognizes cases in which AI assistance can help people communicate or continue working. His distinction is between tasks directed at a human reader, where he values an individual voice, and routine or process-oriented writing where that aim may matter less.
Products Worth Considering
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Logitech MX Keys S Wireless Keyboard, Graphite
As an affiliate, we earn on qualifying purchases.
Fowler’s Exceptions to the Rule
Fowler notes that patterns readers associate with AI are not conclusive evidence of AI authorship. He says some expressions he finds grating are also used by people, and suggests that language shifts across generations and regions. That makes his own ability to recognize an LLM’s style uncertain, even as he describes a strong personal reaction to it.
He gives accessibility and translation as reasons to use language models. His colleague Jim Highsmith, who has Parkinson’s disease, relies on an LLM to support his writing, Fowler says. He also describes colleagues who use LLM translation to communicate in English, a language they may speak as a second or fourth language. Fowler says he would accept some AI-like prose to keep receiving Highsmith’s ideas, and recognizes the value of translation for colleagues.
Fowler considers routine letters and reports unlikely to be read by people other cases where AI-written prose may be reasonable. He quotes Michael Nygard’s comment that a form letter can amount to communication with a process rather than another person. He also raises the possibility that some documentation may eventually be read mainly by LLMs. These are examples and possibilities in the essay, not evidence that such uses are already standard.
“You’re communicating with a process rather than a human.”
— Michael Nygard, as quoted by Fowler
Limits of Style Detection
The essay does not establish how common “LLM voice” is, whether readers share Fowler’s reaction, or how accurately people can identify AI-generated writing. Fowler explicitly says he is suspicious of his own recognition and wary of detection tools, but provides no accuracy figures or study results.
It is also unclear how much editing is needed to change the feel of an AI-generated draft: Fowler cites one professionally edited example but gives no broader comparison. His concerns about the effects of delegating writing on people’s thinking are presented as a personal view, not as a demonstrated general outcome. The source material gives no publication date, so the timing of the essay cannot be specified more precisely.
A Writer’s Choice of Tools
Fowler sets out advice rather than announcing a policy change or a forthcoming study. Writers who follow his approach can use AI for limited purposes such as searching, translation or outlining, while drafting the final prose themselves and stating how AI contributed. His website’s stated policy is that writers who use AI to help should declare that use, usually in acknowledgments.
The essay does not name a next milestone, propose a test for identifying LLM voice or say that Fowler plans to evaluate detection products. The immediate question for readers and authors is how they want to balance assistance, transparency and personal expression, especially when writing is meant to carry a person’s ideas to another person.
Key Questions
What does Fowler mean by “LLM voice”?
He uses the term for prose that can make a reader suspect an LLM wrote it and feel less interested in continuing. He describes the response as an “uncanny valley”, but does not provide a formal definition or test.
Does Fowler say AI-writing detectors can identify it reliably?
No. Fowler says he is wary of detector accuracy and argues that detection misses the point: a reader may already have disengaged by the time they seek a tool’s verdict.
Does Fowler oppose all AI help with writing?
No. He sees possible uses in outlining, translation, searching, accessibility support and routine writing. For prose aimed at human readers, however, he recommends that authors write the actual text themselves.
Why does Fowler connect writing with thinking?
He says writing often helps him clarify ideas and work through confusion. He worries that delegating prose may also mean delegating some of the thinking, but presents this as his own concern rather than a measured conclusion.
Source: martinfowler.com
