A songwriter at her kitchen table in Brooklyn, late on a Tuesday, types a prompt into a music generation tool — americana ballad, three-quarter time, patient, sad. Forty seconds later a song arrives. Three verses, a chorus, a credible voice singing words she did not write, a steel guitar swelling in the right places. She listens twice. Then she closes the laptop and writes a different song into her notebook — the one that had been asking her to write it before she ever opened the machine.
The honest position on artificial intelligence in the creative process, the one that gets applause from neither side, is that the tools are astonishingly good at what music looks like and astonishingly bad at what music is for. The current models produce material that resembles songs the way wax fruit resembles an apple. The color is right. The shape is right. The thing is not alive. A song is alive because someone meant it.
There is real work the machine is already doing well in serious rooms. Stem separation that used to take an afternoon takes thirty seconds and sounds better than it should. Vocal repair that would have cost a week of editing happens in the background while you eat lunch. Mastering assistants get you ninety percent of the way to a competent master on a Tuesday afternoon. A songwriter who is genuinely stuck can ask a model for a chord she would not have thought to play and find the door she had been pressing on. None of that should embarrass anyone. The work of removing tape hiss has never been the work that makes a record matter.
What the model cannot do is mean anything by what it produces. It cannot have driven home in the rain from a friend's funeral last Thursday. It cannot have stood in the doorway of the room where a marriage ended and then sat down and tried to put four lines around it. It can simulate the sound of those experiences, beautifully sometimes, because it has been trained on a million records made by people who did live them. But the song the machine generates is a memorial to other people's lives, produced quickly. The song you write at three in the morning is a transmission from your own.
Holly Herndon has been honest about this in a way more songwriters could borrow from. Her work with Spawn, the vocal model she and Mat Dryhurst trained on her own voice, is openly collaborative — not a tool pretending to be a partner but a partner being treated as a tool. The disclosure is the point. The audience knows what they are hearing and consents to it, and the work is stronger for the clarity. The collapse, in this moment, will not come from honest collaboration. It will come from artists who try to launder machine work as their own and slowly stop being able to tell the difference between what they made and what was made for them.
The competitive question for every working musician now is the same one cabinetmakers and translators and illustrators have already been asked: which part of your work is the part the machine cannot do, and how much of your day is being spent there. Every artist still has access to the one resource the model does not have, which is a body in time, with a history, that has loved and lost and stood in specific rooms with specific people. The work has not stopped being possible. The work has stopped being optional.
A song that sounds like a song is now cheap and getting cheaper. A song that someone needed to write has not been automated and probably cannot be. The difference is small in the data and enormous in the room.