Field Note
Every working professional who has spent more than twenty minutes delegating routine drafts to a modern large language model knows the uncanny sensation of reading synthetic text. It is rarely ungrammatical, and it rarely makes outright factual blunders on simple tasks. Instead, it suffers from a distinct rhetorical stench: an oily, breathless cadence that immediately signals no living human was involved in putting the sentences together.
The tells are remarkably consistent across model families. There is the dramatic sentence fragment severed by unnecessary punctuation. There is the binary antithesis formula—"It's not just about efficiency; it's about transformation." There are the performative throat-clearing caveats that sound like a nervous corporate communications committee. Most irritating of all is the unprompted self-applause: the conversational habit of dropping a one-sentence paragraph like "And that matters" or "Here is why that changes everything" as if the software were pacing across a TED stage in a headset microphone.
Most users respond to these tics reactively. They get a memo or an email back from the model, cringe at the tone, and type a follow-up prompt: *"Make this sound less like AI,"* or *"Rewrite this so it sounds natural."* This almost never works well. The model lacks an intrinsic model of human casualness; when told to sound "human," it simply substitutes one set of artificial heuristics for another, often veering into forced folksiness or unearned intimacy.
A more durable approach is to treat tone correction as a systems problem rather than a per-chat negotiation. Instead of pleading with the model after the fact, you build an explicit negative style sheet and embed it where the tool permanently references it—whether that is Claude's custom Project instructions, ChatGPT's system prompt and Custom Instructions, or a dedicated workspace configuration.
To build a useful negative style sheet, you must catalog the exact stylistic mannerisms you want barred from your output. A solid starting list includes prohibiting rhetorical questions used as transitions, outlawing the "not X, but Y" construction, banning self-validating summary sentences, and strictly forbidding buzzwords like "delve," "tapestry," "beacon," or "paradigm." When an LLM is given clear negative boundaries (what *not* to do) rather than vague positive aspirations (like "write engagingly"), its output becomes dramatically cleaner.
Some technical writers take this approach further by binding their models to existing controlled language frameworks. One popular reference point is ASD-STE100, the Simplified Technical English standard originally developed for aerospace maintenance documentation. STE enforces strict limits on sentence length, restricts verbs to active voices and specific tenses, and eliminates unapproved synonyms so that complex procedural instructions cannot be misinterpreted. While you probably do not want your daily correspondence to read like a Boeing turbine repair manual, the underlying principle is sound: rigid structural rules produce clearer prose than subjective stylistic requests.
However, even the most rigorously configured system prompt is not a substitute for manual editorial hygiene. Outsourcing the first draft of an internal memo, a summary of meeting transcripts, or an initial technical spec saves genuine time. But treating that output as finished work is an unforced error. AI text generators are pattern-matching engines optimized for plausibility, not taste.
The only reliable filter for machine prose remains the human ear. Before sending any document that carries your name, read the draft aloud. If a sentence feels too theatrical to say to a colleague across a kitchen table without laughing, strike it out. Use the machine to break the tyranny of the blank page, rely on persistent rules to eliminate its worst rhetorical habits, and keep the final editing pen firmly in your own hand.