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 titled “LLM Voice,” Martin Fowler argues that readers can react negatively to prose they perceive as AI-generated, even when they cannot identify why. He advises writers to avoid using large language models to draft prose, while allowing that AI can help in some cases, including translation, accessibility and routine communications.
Martin Fowler has published an essay warning that LLM-drafted prose can retain a recognizable “voice” that some readers find alienating, even after a person edits it. The software writer says authors communicating with people should generally write the prose themselves, while acknowledging exceptions such as accessibility needs, translation and routine forms.
Fowler describes a reader’s growing suspicion that a post was generated by a language model, followed in some cases by a loss of interest or an “uncanny valley” reaction. He says this feeling can arise without a reader being able to explain which features triggered it. The essay does not present a test or study measuring how often readers recognize AI-generated writing; it sets out Fowler’s personal observations and advice.
He is skeptical of tools marketed to distinguish AI-written prose from human writing. In his view, readers may already have reacted to the writing by the time they consult a detector, and the tool’s result may not settle the question. He also cautions that a reader’s judgment can be mistaken: language changes, and expressions a person dislikes may reflect generational or regional differences rather than machine authorship.
Fowler urges writers not to ask an LLM to produce a first draft, arguing that a human edit may not fully remove the style he calls “LLM-Voice.” He says outline help might be acceptable, but recommends that authors write the prose themselves. He also says AI use on his own site should be disclosed, and describes using AI as a search tool while checking information against its original source.
The Stakes for Human Readers
Fowler’s argument focuses on a problem that AI-detection scores cannot fully address: reader trust and connection. A text can be accurate and grammatically smooth yet still feel impersonal to someone who suspects a model shaped it. That reaction is subjective, but it may affect whether people keep reading and how they interpret an author’s care or credibility.
He also links writing with thinking. Fowler says working through sentences helps him make sense of ideas, and warns that delegating prose may also mean delegating some of that process. This is his view, not a finding established by the essay. Still, it points to a practical distinction for writers: using a model to assist with a task is not necessarily the same as asking it to formulate the message a person wants to communicate.
The advice has limits. Fowler makes room for cases where AI-generated or assisted language supports a person’s ability to communicate, and for documents intended mainly for a process rather than a human reader. His position is a recommendation about authorship, not a claim that every use of AI writing is harmful.
Products Worth Considering
Samsung Galaxy Tab S10 FE 128GB Wi-Fi Android Tablet, Blue
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
As an affiliate, we earn on qualifying purchases.
Fowler’s Rules for AI Assistance
The essay presents a qualified policy rather than a blanket rejection of language models. Fowler says he does not use AI to generate prose or help structure his writing, though he does use it to search for information and then verifies that information at the source. He recommends disclosing AI assistance in acknowledgments so readers can understand how it was used.
He points to examples that complicate a simple human-versus-machine distinction. Jim Highsmith, a contributor to Fowler’s site, relies on an LLM to continue writing while living with Parkinson’s disease, Fowler says. He also notes that colleagues who are less comfortable writing in English have used LLM translation to communicate. In those situations, he says, enabling someone to share ideas can outweigh concerns about style.
Fowler further allows that an LLM may be suitable for a form letter or a report unlikely to be read by people. A comment cited in the essay from Michael Nygard describes such communication as being directed at a process rather than another person. These examples frame Fowler’s central distinction around the intended reader and purpose, rather than treating AI assistance as uniformly appropriate or inappropriate.
“You’re communicating with a process rather than a human.”
— Michael Nygard, as quoted in Fowler’s essay
Limits of Recognizing AI Prose
The essay does not establish a reliable way to identify LLM-generated writing, nor does it provide evidence that readers can consistently distinguish model-written prose from human writing. Fowler explicitly questions his own ability to recognize it and expresses skepticism about detection tools. His description of “LLM-Voice” is a personal assessment, not a technical category with a stated test.
It is also unclear how much of the perceived style comes from a model, the prompts and editing used, or conventions that are becoming common in human writing. Fowler says one fully AI-drafted piece he encountered did not read as AI-generated after a professional editor revised it, although he presents this as an individual example. The source material supplies no publication date, reader survey or comparative analysis to test the broader claims.
Disclosure and Writing Choices
Fowler’s immediate advice is for writers to decide who will read a document and what they want it to do before choosing how much AI assistance to use. For writing meant to carry a person’s thoughts to other people, he recommends composing the prose directly and disclosing any AI help. He does not set out a new formal policy or announce a planned study.
The essay leaves open whether LLMs will become difficult for readers to distinguish from human writing, and Fowler says that possibility concerns him. For now, the practical questions remain how authors use AI, whether readers are told about that use, and whether assistance helps someone communicate or replaces the work they intend to do themselves.
Key Questions
What does Martin Fowler mean by “LLM-Voice”?
He uses the term for a style of prose that can make a reader suspect an LLM wrote or shaped it. Fowler describes the resulting response as subjective and says he cannot always explain what prompted it.
Does Fowler say AI-detection tools can reliably identify generated writing?
No. He says he is skeptical of those tools’ accuracy and believes a reader’s reaction may already have happened before a tool is consulted. The essay offers no accuracy data or independent evaluation.
What does Fowler recommend to writers?
For writing intended to communicate with people, he recommends that authors write the prose themselves rather than start with an LLM-generated draft. He says outline assistance may be acceptable and recommends disclosing AI use.
Does Fowler oppose every use of LLMs for writing?
No. He describes possible exceptions, including translation, support for a writer with Parkinson’s disease, and routine documents with little expected human readership. His advice depends on the purpose and circumstances.
Source: martinfowler.com
