Martin Fowler tells writers to stop letting AI draft their prose
His new LLM Voice essay argues that polished machine prose can repel readers, while a Hacker News discussion pushes back on where assistance ends and authorship begins.
Martin Fowler’s advice to authors is blunt: do not start by getting an LLM to write the draft. In his October 29 essay, “LLM Voice,” he argues that even human editing can leave prose with a machine-like quality that drains his interest in reading it.
His concern is not a new test for detecting AI. Fowler explicitly doubts both detection tools and his own ability to recognize generated writing. He distinguishes a reader’s reaction from proof of authorship: by the time he wants a detector, the loss of engagement has already happened.
The essay recommends that writers supply their own prose and be cautious even with suggested rephrasing. Fowler says he has heard of useful AI assistance with outlines, but does not use it himself for prose or structure. His site’s policy asks contributors who use AI in writing to explain how they used it, partly so readers can learn from the practice.
A dispute about what writing is for
Fowler’s larger argument is that writing helps him think. Wrestling an idea into sentences exposes gaps in his understanding. He worries that delegating the sentences also delegates some of that work. This is his assessment of the writing process, not experimental evidence that AI assistance inevitably damages comprehension.
The same-day Hacker News discussion shows why the boundary is contested. Some participants describe generated prose as tiring or low in information. Others argue that writing serves different purposes: a personal essay, an internal account plan and a bureaucratic form do not necessarily call for the same degree of personal expression.
One participant argues that AI help can let people communicate technical knowledge they otherwise could not present fluently. Another challenges the idea that using AI prevents independent thought, emphasizing the user’s choices and context. Those are community arguments and anecdotes, not a representative survey of writers or readers.
The exceptions matter
Fowler does not make his recommendation absolute. He describes AI assistance as valuable to his colleague Jim Highsmith in dealing with Parkinson’s disease, and says he would accept the resulting voice to keep reading his work. He also recognizes administrative writing as a possible use, while remaining uneasy about the trade-off in translation.
That nuance matters for anyone turning the essay into a universal ban. The dispute is about how authors preserve thought, voice and disclosure while using assistance. It does not establish that a particular punctuation mark identifies AI writing, or that every reader rejects an assisted draft.