Start with what a voice actually is, mechanically. It is not a mystical essence. It is an accumulated stack of deviations: the intervals you overuse, the words you reach for that nobody else would, the grammatical fractures you allow, the subjects you cannot stay away from, the “mistakes” you have repeated until they became signature. Strip the deviations from any distinctive writer and you get competent, anonymous work. The deviations are the voice.
Now consider what a language model is: a system trained to predict the most probable continuation. Probable means typical. Typical means the deviations get smoothed. This is not a criticism of the technology — prediction is the whole engineering achievement — but it means a writer and a model are, in one narrow sense, pulling in opposite directions. The model is a gravity well toward the average. Your voice is everything about you that escapes it.
This does not mean the tool is unusable. It means it is usable the way a loud amp is usable: deliberately, with your hand on the volume, knowing exactly what it does when you stop paying attention.
Watch the smoothing happen
Take a line with some blood in it: “I keep your winter coat because the closet asked.” It is not reasonable. Closets do not ask. That unreasonableness is the entire line — grief displaced onto furniture, said fast, no hand-wringing.
Ask a model to “improve” it and you will get something like: “I keep your winter coat because I can’t let go.” Every metric a model can measure has improved. Clarity: up. Grammar: fine. Sentiment: legible. And the line is dead — it now says what a hundred thousand songs have said, in the way they said it. The closet was the song. The model cannot know that, because the closet was a deviation, and deviations are what it is built to repair.
Once you have seen the smoothing pattern you cannot unsee it, and that is the best protection available. The failure mode is never dramatic. It is a series of small, individually defensible edits, each one making the song slightly more like every other song.
The rules
The first draft is yours, always. Voice lives in first-pass instincts — what you reach for before anything is watching. Let a model write the first draft and everything after is decoration on its bones, not yours. Draft with the chat closed. Bring the draft afterward.
Use it as a mirror, not a pen. The right questions are diagnostic: what is this verse doing, where does the energy sag, which lines are filler. The wrong request is “rewrite this.” A mirror shows you the song so you can change it. A pen changes it for you, at the exchange rate described above.
Interrogate every suggestion. When it proposes a change, demand the reasoning. “The second verse repeats the first verse’s information” is a real insight you can act on in your own words. The specific replacement line it offers is somebody else’s average. Keep the diagnosis, discard the prescription.
When your gut defends a line, the gut wins. The model will flag your strangest lines as problems — reliably, structurally, forever. Some of them are problems. But the model cannot distinguish a flaw from a signature, because from inside the statistics they look identical. Only you have the tiebreaker information: what the line costs to sing, where it came from, whether it is true. Any line you are willing to defend out loud stays.
Never accept language you would not have found. A practical filter: if a suggested phrase impresses you because it does not sound like you, that is precisely the phrase that erodes the voice. Borrowed brilliance reads as borrowed forever. Your listeners will not know why the line rings false. You will.
Recharge between sessions. Voice maintenance is an input problem too. The model’s suggestions are drawn from everything; your voice is drawn from your particular corner — the records you wore out, the way people talk where you are from, the specific life. Time spent there (listening, reading, living, playing with humans) restocks the one thing the model cannot supply.
The bandmate test
All of this compresses into one question you can ask at the end of any writing session: could I defend every line of this song to my band? Not “is every line good” — defend. Explain where it came from, why it is that word and not the obvious one, what it is doing there.
Lines you wrote survive the test even when they are flawed; you know their reasons. Lines you accepted from a machine without interrogation fail it instantly — there is nothing behind them to defend, just a probability and your fatigue on the night you took the suggestion.
A song where every line has a defender is a song with a voice. That is the whole game. The tool is welcome in the room on those terms: it can ask you hard questions, hand you the name of a problem, take the lookup work. It just never gets the pen. The pen is where you live.