There is a voice you already recognize without meaning to. It almost always opens the same way, with a promise that here you will discover something; it strings together lists of three; it leans on "moreover" and "in short"; and it closes with a line that reads like the moral of a brochure. It is the factory voice of AI when you ask it to "write about X" and let it run loose. It is not bad. It is average. And in writing, the average is the one thing no one remembers.
It helps to understand why this happens before trying to fix it. A language model does not reach for the truest sentence, or the one most yours: it reaches for the most probable. It was trained on oceans of text and then tuned to please human raters, a process that rewards the correct, the harmless, and the agreeable [1]. The result is prose that gravitates toward the center, like a marble rolling to the lowest point of a bowl. Asking it to "write well" and nothing more is asking it to roll toward that bottom. And the bottom, by definition, is crowded.
01 · the reflexWhy factory text sounds like a factory
Think of the model as an instrument tuned to a single note: the note of the middle ground. When you give it no voice, it adopts the voice of the entire corpus, which is no one's in particular. That is why its texts share recognizable tics: hollow transitions, filler adjectives, that insistent politeness that sounds like a customer service manual. Researchers who studied how to detect generated prose found regular statistical patterns (a certain uniformity in rhythm and word choice) precisely because the model avoids risk [2].
AI does not write badly. It writes to the center. And the center is where everyone already is.
From here comes the first principle, the one that governs all the rest: if you give it no voice, it hands you back everyone's voice. The task, then, is not to ask it to write better. It is to deny it the average. And to deny it the average you have to give it something concrete to write against: a voice, a text of your own, an uncomfortable constraint.
02 · the craftEditor and sparring partner, not ghostwriter
The original mistake is asking the AI to write for you. That is where you lose the one thing that sets you apart, your voice, in exchange for speed. The use that does keep your voice is a different one: using it as an editor and as a sparring partner. The editor takes what you already wrote and tightens it; the sparring partner throws punches back so your idea either defends itself or falls.
As an editor, the productive instruction is never "improve it." That word gives it license to rewrite to its own taste, and its taste is the average. The productive instruction points to a specific flaw and forbids touching anything else: "Cut this paragraph in half without dropping any of the three ideas." "Flag the five weakest sentences and explain why, but do not rewrite them." "Find where I repeat the same idea in other words." Notice the pattern: you ask for a diagnosis or a bounded surgery, not a new version.
More than a century ago, William Strunk Jr. distilled the craft of editing into an order that still holds: omit needless words [3]. A model obeys that command better than almost any other, because removing is a bounded, verifiable operation. Ask it to cut, not to embellish, and suddenly it works in your favor instead of diluting you.
As a sparring partner, the value is that it never tires of contradicting you. "Give me the three strongest arguments against this thesis." "What would a skeptical reader object to in the second paragraph?" "If you had to demolish this idea in one sentence, which would it be?" You do not use its prose: you use its resistance. You return to your text knowing where it is fragile, and you rewrite it yourself, in your own words, now armored.
Before accepting any rewrite, paste it next to your original and read them aloud, one after the other. If the AI version is more "correct" but could have been written by anyone, discard it even though it sounds better. Correctness is not the goal; the loss of voice is the price, and it is almost never worth paying for a smoother paragraph.
03 · the voiceGive it a pattern to write against
The most direct way to pull the model out of the average is to give it a sample of how you sound. Do not describe it with adjectives ("warm but professional tone" means nothing operational); show it to the model. Paste two or three paragraphs you have written that represent you, and ask it to edit while respecting that register: same cadences, same sentence length, same vocabulary. Learning from a few examples inside the message itself is, in fact, one of the core capabilities of these models, demonstrated in the earliest work on large models [4].
That sample works like a flavor template. Without it, the model averages the world; with it, it averages toward you. The difference in the result is the difference between a text that could belong to anyone and one that sounds like you wrote it on a good day.
| Goal | Request that flattens | Request that keeps voice |
|---|---|---|
| Shorten | "Make it more concise." | "Cut 30 % without removing any of my three ideas or changing my vocabulary." |
| Polish | "Improve the writing." | "Flag the sentences where the rhythm breaks; do not rewrite them, just point them out." |
| Strengthen | "Make it more convincing." | "Give me the two strongest objections to my central argument." |
| Register | "Use a professional tone." | "Edit by imitating the rhythm and vocabulary of this paragraph of mine that I am pasting." |
There is a limit worth stating plainly, without dressing it up: none of this guarantees a good text. A model can sharpen a sentence, but it cannot have something to say on your behalf. The idea, the angle, what you risk when you assert something: that remains your work, and it is precisely the part no average knows how to do.
04 · the habitYou write, the machine corrects
The flow that keeps your voice inverts the usual order. First you write, even if it comes out rough: the ugly draft has something the factory text will never have, which is intention. Then the AI steps in, and only for bounded tasks: spotting repetitions, flagging weak sentences, proposing a cut, arguing an argument. The last hand is yours again, always, because you are the only one who knows how you want to sound.
Write first, yourself. The machine sharpens; it should not be the one deciding what to say.
Sounding like AI is not a flaw in the tool: it is what happens when you hand over the decisions only you can make. Give it back its right role (a demanding editor, a tireless sparring partner, never the author) and the same model that used to flatten you starts putting an edge on you. The average prose will still be there, waiting at the bottom of the bowl, for whoever wants it. Your job is not to roll all the way down to it.
Fuentes
- Ouyang, L. et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (NeurIPS) 35. arXiv:2203.02155. (Describes tuning by human feedback, RLHF, which steers the model toward agreeable and safe responses.)
- Gehrmann, S., Strobelt, H. & Rush, A. M. (2019). GLTR: Statistical Detection and Visualization of Generated Text. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL): System Demonstrations, pp. 111-116. ACL Anthology: aclanthology.org/P19-3019. (Shows that generated text exhibits detectable statistical regularities.)
- Strunk, W. Jr. (1918). The Elements of Style. Classic rule of style: "Omit needless words." (Public domain; expanded edition by E. B. White in 1959.)
- Brown, T. et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems (NeurIPS) 33. arXiv:2005.14165. (Demonstrates learning from few examples inside the prompt itself.)