The pitch for AI songwriting tools usually goes one way: describe a song, receive a song. That version of the product exists, works as advertised, and produces songs nobody needs. If a machine can generate the song from a one-line description, the description was the whole idea, and it was not much of an idea.
There is a different way to use a language model in a writing session, and it looks less like a vending machine and more like the friend who sits in the room while you work. Not writing. Listening, asking, occasionally naming the thing you already half-knew. That version is genuinely useful, and it is worth being precise about why.
What a companion is good at
Naming what you already suspect. Most feedback a songwriter needs is confirmation of a suspicion. You know the second verse sags. You know the bridge changes the song’s temperature. What you want is someone to say it out loud so you can stop negotiating with yourself. A companion that has read your draft can do this instantly, at 1 a.m., without the social cost of asking a friend to care about your demo again.
Diagnosis, not prescription. “Why does my chorus feel like a different song?” is a real question with real answers: the melody’s range jumped, the harmonic rhythm doubled, the lyric switched from concrete to abstract. A model is good at spotting these mechanical discontinuities because they are patterns, and patterns are the one thing it actually has. You keep the judgment; it supplies the vocabulary.
Generating options when you are stuck on mechanics. Sometimes the block is not inspiration, it is inventory. You need a rhyme that is not “fire/desire,” a chord to get from F back to A minor, a word with two syllables and a hard consonant. This is lookup work wearing a creative costume. Delegate it freely. Nobody’s voice ever lived in their rhyme lookup.
Being a rubber duck with opinions. Explaining your own song to someone — what it is about, why the second verse exists — is diagnostic even if they say nothing back. A chat forces the explanation into words. Half the time you find the problem while typing the question.
What it is bad at, structurally
A language model’s sense of quality is an average of everything it has read. Your song’s job is to not be an average. This is not a flaw that better models will fix; it is what “model” means. Three consequences follow.
First, its generated lines will be competent and dead. Competent because the patterns are right. Dead because a line lives on its deviation from pattern, and deviation is exactly what got sanded off.
Second, it will flag your best lines as problems. The strange image, the grammatical fracture, the word that should not work — these read as errors to a system trained on regularity. If you accept every suggestion, you will watch your song get smoother and worse in real time.
Third, it cannot hear. Every judgment it offers about your melody is a judgment about your description of your melody. The gap between those two things is where songs actually live.
A workflow that keeps the song yours
The order of operations matters more than the tool.
Draft first, alone. The first pass is where the voice is. Do it with the chat closed. Bad first drafts are not a problem a companion should solve; they are the raw material everything else needs.
Bring the draft, ask diagnostic questions. Paste what you have. Ask what is working, what is filler, where the energy drops. Argue with the answers. The argument is the useful part.
Interrogate, do not accept. When it suggests a change, ask why. The reasoning is often more valuable than the suggestion — “the verse is all abstract nouns” is a lesson you keep; the rewritten line is a transplant that may not take.
Keep the weird lines. If the model flags a line and your gut defends it, the gut wins. It has information the model does not: what the line costs you to sing.
Close the tab and finish. The song gets finished on an instrument, not in a chat. If the conversation is running long, that is the tool optimizing for conversation. Yours should be trying to get rid of you.
The test
There is a simple check for whether a companion is helping: after the session, is there more of you in the song or less? More — sharper intentions, defended choices, a verse you rewrote yourself because the chat helped you see what it was missing — means the tool is working. Less — accepted lines you cannot explain, a lyric that would embarrass no one and move no one — means you have been ghostwritten by a very polite average.
The machine is a good bandmate and a bad ghostwriter. Book it accordingly.