There's a particular sensation that comes over a reader partway through a post: the growing suspicion that the prose isn't human. It resists precise definition. The usual tells circulate online — not the em-dashes, someone quips, but the load-bearing tapestries. Whatever the marker, the effect is real: the urge to keep reading withers, sometimes into outright revulsion.

Why Detection Tools Miss the Point

Tools that claim to separate LLM prose from "the old-fashioned biological product" are a doubtful remedy. Accuracy is one concern. The deeper problem is that the reaction to LLM-Voice is human and precedes any tool: it's the sense of sitting in an uncanny valley of prose. By the time a reader reaches for a detector, the damage is done — even if the verdict comes back clean.

Self-suspicion cuts the other way, too. Language shifts between generations, and griping about the speech of the young is a familiar symptom of age. Phrases like "adopting a new sales motion" or "a sharp round-up of news" can irritate, but they come out of human mouths as well; that's a generational gap, not a carbon-versus-silicon one. (The transplanted Brit has the same reaction to Americanisms, and expects to get used to them momentarily.)

Drafting Is Where Authors Get Trapped

Authors should treat LLM-Voice with extra caution right now, and the specific trap is letting an LLM produce a draft. Human editing afterward doesn't reliably expel the voice, and the uncanny-valley reaction follows.

Very hard, but not impossible. I've come across one case where I do know the first draft was entirely done by LLM, but I detected no LLM Voice when I read the final piece. It did, however, need a professional editor at a major periodical publication to clean it up.

Accepting suggested rephrasings is similarly risky. Using an LLM to develop an outline or structure may be acceptable, but only that far: the prose itself should be the author's.

Where AI has a hand in a piece, the author's declared policy is transparency — contributors should state how they used it, usually in the acknowledgments, both to be open with readers and to model useful practice. The author's own usage stays narrow: no prose generation and no structural help, only search, with every fact verified against the source.

Writing as Practice — and as Thinking

Advice to say your writing aloud grows more valuable in the LLM era, not less. The habit is simple: speak the words in your head, write them down, then edit. Anyone who wants to communicate through writing regularly should take it seriously, because good writing is built on lots of practice — a skill worth the time, and one that handing over to an LLM will undermine.

There's a second worry past the questions of voice and skill. Should LLMs get good enough that the difference becomes undetectable, the situation turns more troubling, not less. For most writers, writing is an almost-essential tool for thinking; coherence often arrives only in the act of putting words down. Essays frequently begin in admitted confusion and end in understanding because the wrestling with language forces the confusion out. Delegate the writing and you delegate the thinking and comprehension along with it.

Declining to delegate has its own reward: where's the fun of writing otherwise?

When LLM-Voice Is Worth Accepting

The rule against LLM-Voice is general, not absolute.

  • Disability. Jim Highsmith (who contributes to the same site) finds working with an LLM essential to his writing in order to overcome the bane of Parkinson's disease. Trading some LLM-Voice for an old friend's continued wisdom is an easy bargain.
  • Translation. Many colleagues work in English as a second or fourth language, and some have found LLM translation an important part of communicating with those who know rather less Marathi. Still, clunky genuine human prose is often preferable to the smooth but soulless alternative.
  • Prose addressed to no one. A form letter to an official, or a report no one will ever read, is a reasonable place for even a native Englishman to let an LLM write. Michael Nygard's framing: "There you really aren't trying to communicate thoughts from one person to other people. 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." Perhaps most future documentation will be read only by LLMs, which have no uncanny-valley worries. But prose meant for human readers has to carry its humanity — shining, or at least oozing, off the page.

Drafts of this piece drew comments from Alessio Ferri, Caer Sanders, Ian Cartwright, Jie Jessie Xia, Joshua Kerievsky, Michael Feathers, Michael Nygard, Scott Davies, and Wei Zhang. No AI or LLMs were used in writing it.