The Machine Is Not the Muse

The machine at the center of this project is a graphics card in a rented server, and there is nothing holy about it. It does not think. It does not understand. It does not create, in the sense that a person creates. And it is certainly not a muse whispering to me from somewhere beyond the room. I want to say this first, and plainly, because the error I most want to avoid is the one that flatters the machine. When I open Suno and generate a piece of music, I am operating a tool, and the temptation to treat the tool as a mind is exactly the temptation I am arguing against.

This is the first of three essays that carry the argument behind Found Wanting. It is the one about what the machine actually is.

So what is it? A generative music model is machinery for extracting statistical patterns from a very large body of recorded sound and recombining them on request. That is the whole of it. It detects regularities: what tends to follow what, which intervals and rhythms and timbres recur, how a phrase usually resolves. Then it produces new sound consistent with those regularities. There is no intention inside it and no spark. The intelligence, if you want to use that word at all, is not in the card. It is in the material the card was trained on.

Which is the part worth thinking about. What was it fed?

Suno has said publicly, including in the copyright litigation brought against it by the major record labels, that its training corpus includes essentially all reasonable-quality music accessible on the open internet, numbering in the tens of millions of recordings. I take the company at its word on the scale. I make no claim here about whether gathering that material was lawful; that question is real, it is unresolved, and it is not mine to settle. What I care about is a different property of that archive, one that survives whatever the courts decide.

The archive is not raw noise. It is curated, and curated in two quite different senses.

The first kind of curation happened long before any company existed to collect it, and no engineer performed it. Every recording in that corpus is the survivor of a long chain of human selection. A person composed it. People judged it worth performing. Someone thought it worth recording, then worth releasing, then worth preserving, then worth uploading, and other people kept choosing to listen to it and to pass it on. By the time a recording is sitting on the open internet in reasonable quality, it has already been filtered, again and again, by human hands and human attention. That filtering was distributed across an enormous number of people, and it had nothing to do with any algorithm.

The second kind of curation is technical and recent. Suno then cleaned and filtered what it had gathered, discarding the corrupt and the unusable before training on the rest. This is the ordinary, unglamorous work of preparing data. It matters, but it is not the part that moves me.

I want to be precise about what this culturally selected archive is and is not. It is not only the polished output of record labels. It includes independent uploads, stock music, public-domain recordings, and material posted by ordinary people. And cultural selection is not the same thing as goodness. I am not going to tell you what fraction of everything a human being has ever hummed, or played, or sung ends up as a durable public recording. Nobody knows that number, and I will not invent one to make a point.

More than that: breadth and frequency do not manufacture goodness by statistical fiat. That a pattern appears ten million times does not make it true, and it does not make it beautiful. A popular error is still an error. The loudest thing is not the best thing. If my argument were “popular equals good,” it would deserve to be thrown out, and I would throw it out myself.

My argument is smaller, and I think sturdier. The archive is not a heap of numbers. It is made of created works. Every one of those tens of millions of recordings is a made thing, an act of some person shaping sound on purpose. And creation is good. That is the premise I begin from, and I state it plainly so you can weigh it: to make is a good act, and a machine trained on an enormous quantity of human making has taken in the residue of an enormous quantity of good acts. The goodness is not in the machine. It arrived with the material.

And here is the feature I value most, the reason I wanted to work this way rather than merely put up with it.

You cannot ask who the guitarist is. There is no guitarist. You cannot fall for the singer, because there is no singer. There is no face to print on a poster, no voice to build a cult around, no genius to defer to and no genius to envy. The idolatry that has always shadowed music, the way devotion to a performer can quietly eclipse devotion to whatever the music was pointing at, is here made structurally impossible. The ego is removed. What remains is the sound.

For a project that sets the words of the Bible to music, this is not a small thing. A performer’s personality can wedge itself between a listener and the text, and here there is no performer to do it. No one’s charisma is doing the persuading. If a listener is moved, it is not by me, and it is not by a star. What is left in the room is the pattern and the words.

None of this lets me off the hook, and it does not let the tool off the hook either. A machine that has absorbed the patterns of human making can just as easily be aimed at something false, ugly, or cruel. Order supplies capability; it does not supply intention. We are fallen creatures, and the same instrument I used to assemble these songs is turned to assemble worse things every day. The archive hands you material that is good in the deep sense, because it was made; what you do with it is on you. Every choice in Found Wanting, which passage, which reading, where a verse ends and where the silence falls, which of a hundred generated fragments to keep and which to throw away, is a human judgment I am answerable for. The machine offers patterns. It does not offer purpose. Purpose was always the human part.

So the machine is not the muse. It is closer to a mirror held up to a vast amount of human work, and the work is what deserves the attention. Which raises the obvious next question: why should the consensus of a crowd, musical or otherwise, be worth trusting at all? That is not a matter of faith. It is something I have had to build, with my own hands, in medical research, and it is where this argument turns from assertion into evidence. I take it up in Ground Truth.