AImartin-fowlerllmwritinggenerative-aiediting
Martin Fowler Tells Writers Not to Let AI Draft Even a First Version, Because Editing Can't Get the 'LLM Voice' Back Out
In an October 29 post, software author Martin Fowler says a human edit rarely removes the telltale voice of an LLM-written draft, that even accepting an AI's rephrasing is dangerous, and that an outline is as far as he would let one go.
Martin Fowler, author of the book Refactoring, has a name for the thing readers notice in machine-made prose: LLM Voice. His one-line description of it is “it’s not the em-dashes, it’s the load-bearing tapestries.”
He doesn’t want a detector for it. Asking a tool whether a human wrote something, he wrote, “rather misses the point,” because LLM Voice is an “uncanny valley of prose” that a reader feels before any tool weighs in. “By the time I’d reach for a tool to tell me what I’m reading, it’s already too late, even if it confidently tells me there’s no LLM present.”
He also doubts his own ear. Phrases like “adopting a new sales motion” or “a sharp round-up of news” strike him as LLM tells, and he allows that some of that may be “more a generational difference rather than between carbon and silicon.”
The advice is the part with teeth. “I strongly recommend that authors should not fall into the trap of getting the LLM to write a draft. Even if it’s edited by a human, it’s very hard to expel the LLM-Voice.” He goes further than drafting: “Even accepting suggestions of rephrasing is dangerous.” Asking a model for an outline “may be ok, but only as far as an outline.”
In a footnote, he gives the one counterexample he knows of, a piece that began as an LLM draft and showed no LLM Voice. “It did, however, need a professional editor at a major periodical publication to clean it up.”
Fowler names the cases where he’d take the voice anyway. Jim Highsmith, who contributes to Fowler’s site, works with an LLM to write despite Parkinson’s disease, and Fowler says, “I’ll gladly deal with LLM-Voice” to keep reading him. Translation gets a pass, and so do form letters, where, as Michael Nygard put it in a comment Fowler quotes, “You’re communicating with a process rather than a human.”
His own site has a rule now. Anyone who uses AI to help write for martinfowler.com “should declare how they used it, usually in the acknowledgments.” The post’s own acknowledgments say he “did not use AI or LLMs in any way while writing this post,” and he uses AI only “as a search tool (but I verify any information by going to the source).”
Fowler’s line between an outline and a draft is a line about who decides what gets said. The outline only hands over the structure, while the draft hands over the actual wording, which is the part readers read. Where the line sits for code review comments, commit messages and pull request descriptions, all places developers already lean on models, is still an open question.