Essay on martinfowler.com urges writers not to draft prose with LLMs

A post dated 29 October 2026 on martinfowler.com, titled "LLM Voice", is an opinion essay about how LLM-written prose feels to read. The author says that sometimes, partway through a post, he gets a growing feeling he is reading LLM output. He cannot always say why, though there are common tells: in his words, "it's not the em-dashes, it's the load-bearing tapestries". His desire to keep reading withers, sometimes to the point of revulsion.

He is wary of tools that claim to tell LLM-written prose from the human kind, partly because he doubts their accuracy and partly because they miss the point. The reaction to what he calls LLM-Voice is a human one, a sense of being in an uncanny valley of prose. By the time he would reach for a detector, he says, it is already too late, even if the tool confidently reports no LLM is present. He also distrusts his own ear. Language changes, and complaining about the language of the young is a sign of age. He bristles at phrases like "adopting a new sales motion" or "a sharp round-up of news", but he has heard such constructs from human mouths, so that is more a generational difference than one between carbon and silicon.

His advice to authors is to be extra-wary of LLM-Voice. In particular, he strongly recommends not getting an LLM to write a draft: even when a human edits it, the LLM-Voice is very hard to expel. Even accepting rephrasing suggestions is dangerous. He has heard of people using LLMs for outline and structure, which may be ok, but only as far as an outline; authors should fill the prose themselves. A footnote adds that avoiding the voice after an LLM first draft is very hard but not impossible. He knows one case where the first draft was entirely LLM-written and he detected no LLM Voice in the final piece, but it took a professional editor at a major periodical publication to clean it up.

On his own site, he says, anyone who uses AI to help them write should declare how, usually in the acknowledgments, partly for transparency and partly so readers can learn useful ways to use AI. He does not use AI to generate prose or for structure; he uses it only as a search tool and verifies any information at the source. He states that he used no AI or LLMs in any way while writing this post.

He also gives a practical tip: speak the words in your head, write them down, then edit what you wrote. Good writing takes lots of practice, he says, and surrendering to an LLM will undermine any progress. His larger worry is thinking. He finds writing an almost-essential tool for thinking, often starting an essay knowing he does not really understand something and resolving his confusion by wrestling with the words. As more people delegate writing to LLMs, he argues, they will also delegate thinking and comprehension. He adds that LLMs may in time get so good that nobody can tell the difference, which he finds in many ways more worrying.

The rule is general, not absolute. The first exception is Jim Highsmith, who contributes to the site and is the author's old friend. Highsmith finds working with an LLM essential to his writing because of Parkinson's disease, and the author says he will gladly deal with LLM-Voice to keep reading his wisdom. Translation is a more awkward case: many colleagues have English as a second or fourth language, and some find LLM translation important for communicating. Still, the author says he would prefer a genuine human's clunky prose to the smooth but soulless LLM. The last exception is text where it is reasonable to let an LLM write, such as a form letter to an official or a report he suspects no-one will read. Maybe most future documentation will be read only by LLMs, he writes, and they will not share his uncanny-valley worries. Michael Nygard commented on the form-letter case: "You're communicating with a process rather than a human. A human may open the envelope but that person is acting as a component in a cybernetic organism." The essay closes on the point that to communicate with humans, one's humanity has to shine, or at least ooze, out of the prose.

Key facts

• The author says he can often sense LLM-written prose without defining why; the common tells include "load-bearing tapestries", not em-dashes, and the reaction is an uncanny-valley feeling, not something a detector tool settles.
• He strongly recommends that authors not get an LLM to write a draft, since the LLM-Voice is very hard to remove even after human editing; rephrasing suggestions are also called dangerous, and LLM help is acceptable only up to an outline.
• Exceptions: Jim Highsmith, who uses an LLM to write because of Parkinson's disease; translation, which he calls a more awkward case; and text such as a form letter or a report nobody will read.
• His site's policy is that anyone using AI to help them write should declare how, usually in the acknowledgments; he says he used no AI or LLMs in writing the post.
• His deeper concern is that writing is a tool for thinking, so delegating writing to LLMs means delegating thinking and comprehension.

Why it matters

The essay puts a name on a reaction many readers report: text that reads as machine-made loses the reader's goodwill quickly. The author frames this as a human response, not a detection problem, so the cost falls on writers whose prose triggers it. His larger claim goes beyond style. Because he treats writing as a tool for thinking, he argues that handing drafts to an LLM hands over thinking and comprehension too.

Who it affects

Authors, bloggers, documentation writers and anyone who communicates regularly in writing are the main audience; he says his advice on practising and speaking your words matters most to them. It also touches contributors to his site, who are asked to declare how they used AI. Non-native English speakers who rely on LLM translation, and writers with disabilities such as Jim Highsmith, are named as cases where the general rule bends.

How to use it

The practical advice is short. Write the draft yourself. If you use an LLM at all, keep it to an outline and fill in the prose yourself, and treat even rephrasing suggestions with caution. Speak the words in your head, write them down, then edit. If AI helped, say how, usually in the acknowledgments. The author says his own use is limited to search, with every fact checked at the source. For form letters or reports nobody will read, he considers LLM drafting reasonable.

How solid is it

This is a personal opinion essay. The argument rests on the author's own experience and ear, not on studies or measurements. No AI-detection tools are named and no accuracy figures are given. He concedes his own recognition of LLM voice may be unreliable, since some phrases he bristles at come from humans. The one anecdote of an LLM first draft that read cleanly is not identified by piece or periodical. The post says nine people commented on drafts, and it states that no AI was used in writing it.

Risks and caveats

The author himself flags the limits. His sense of what sounds like an LLM may reflect generational taste, not a real boundary between human and machine writing. Translation is a hard case where he admits LLM help has real value for colleagues. He also warns that LLMs may eventually get so good that the difference cannot be heard, which he finds more worrying, not less. The post gives no enforcement rule for contributors beyond the disclosure policy.

“But if I want to communicate with humans, my humanity has to shine (or at least ooze) out of my prose.”

— From the post "LLM Voice" on martinfowler.com, 29 October 2026