Why generative AI will never create truly great music
In a spirit of pushing the envelope in a countercultural direction, my wife and I recently decided to host a “listening party” with some good friends here in our little town of Colwich, KS. The plan is to share some dessert together with our families and then go down into the basement where the stereo is and listen together to a recording of Rodrigo’s Concierto de Aranjuez on vinyl record. (An amazing piece of music if you haven’t heard it.)
Beforehand, I’ve sketched out a little talk, introducing the piece and giving some background about it and its composer, helping to identify the sound of the various orchestral instruments, explaining exactly what a concierto is, etc, in order to better enjoy the music together. I also plan to talk a little bit about taste: what it is when applied to things other than food, why it is worth cultivating, how it can be cultivated, and some thoughts about why sapientia, the Latin word for wisdom, originally referred to it.
Fundamentally, the idea is that to be wise is to be capable of discernment and judgement, aware of what is present, accurately sensing and feeling even on the most basic bodily level what is good and bad. This requires a certain degree of refinement. The discernible strain of poison (literally in food, metaphorically and morally in music) is sometimes subtle; on the other hand without refinement what is truly fine and good can often be lost to the senses as well.
What exactly is refinement, and how does it come about? One could think of it as a kind of purity, a habit of attunement to what is fine, subtle, and delicate, an ability to look past what is predominant, obtuse, chaotic, and seemingly overwhelming. One of the truly remarkable things about classical music is that the immense amount of structure inherent in it – sometimes as many as 100 musicians told exactly what notes to play, when, and for how long, even where to sit – which might seem to stifle and prevent true human freedom and artistic expression, seems somehow on the contrary to actually provide occasion for deeply human existence. Within the regimentation, regularity, and predictability of the classical scale, rhythm, harmonies, and musical forms, the possibility for significance of small and subtle things arises. A passage taken a little faster, a note played a little louder, the tone of the strings a little sharper than usual, these things can all have significance, and because there is relatively little chaos and unpredictability in classical music as compared to, say, a style in which the musicians might even be improvising in real time, this significance can actually be heard by an attentive listener and not just ascribed to chance. In a word, classical music is refined.
As an exercise in cultivating refinement for our party, I intend to present our guests not only with what I consider to be a truly great piece of classical music, but also with a slight dilemma as well; shall we listen to the recording of the Concierto with Pepe Romero as the soloist, an early recording endorsed by none other than the composer himself, or shall we listen to a later recording with John Williams (the guitarist, not the writer of the Star Wars soundtrack) as the soloist, a guitarist who I think perhaps slightly edges out Romero in terms of pure skill and talent? My hope is that the very slight pain of this decision serves to drive home to our guests the point above – even between two recordings of the same exact piece of music, both of them truly excellent (let alone between various pieces, composers, or eras of classical music) there are differences which can be heard (with practice) and really felt in their meaning.
And this brings me to the title of this article: why generative AI will never create truly great music. It is no doubt risky to make a statement like this, as AI has proven capable of many things (including the rendering of “original” music at all) which had at one time seemed absolutely out of reach for machines. As computers have, like something of a tidal wave, gradually overcome barrier after barrier which they were formerly predicted to never wash over, why would great music stand as a final stopping point? We already have “pretty good” AI music, isn’t it safe to assume we’ll get all the way to truly great music pretty soon?
In order to explain why the answer is no, we’ll need a brief acquaintance with what AI is and how it works. Without needing to go over all the technical details, one can think of generative AI as essentially an averaging machine. This is actually not a new concept at all. One reference to it I love is from the 1950’s post-apocalyptic novel A Canticle for Leibowitz:
Why is Brother Sarl able to complete these works? The reason is because language, and in fact nearly all that can be known, possesses a certain degree of structure. One word tends to follow another, at least in a given context. Of course it does take considerable arithmetic, which is why there is now an estimated three trillion dollar graphics card buildout happening. Three trillion dollars gets you the chips to do an absolutely ludicrously large amount of arithmetic, and now we can complete whole books like this in seconds.
And that’s really all our generative AI is doing. Lots of arithmetic to determine what “fits” in a given context. And this is just as true for AI generated music as it is for AI generated text. This is nothing to be scoffed at, by the way, and I am a huge believer that gaining insight into this underlying structure of the world this way will enable things we can only yet begin to imagine. But the point, for the present discussion, is that AI does not yield what falls outside of the pattern, it yields what falls within it.
At this point you’re likely starting to put together the argument for my opening claim, that generative AI will never create truly great music.
Truly great as opposed to merely decent music is distinguished from the rest precisely by how it diverges from the pattern, from what is expected, what is statistically likely given the existing body of work. Not necessarily because it is shocking or gaudy or out of order. It might be a passage that is played so very evenly, or perfectly in tune. It might be the utter ferocity of an opening note. It might be the absolutely un-reproduced softness of an oboe solo, the clarity of a clarinet, the brashness of a trombone. More than likely, a truly great recording contains all of the above. But what is truly great stands out, it is not the same as the rest nor is it simply an echo or a repeat of some past thing. As the great composer Nikolaus Harnoncourt puts it, regarding the historical understanding of great music:
Music brought about changes in people, in listeners as well as in musicians. It had to be continually recreated, just as human beings had to keep on building new homes, in keeping with new patterns of living, new intellectual climates.
Great music must be in a way new, and it must be set apart, and finally, the truly great recordings are almost all great in different ways, ways which may be opposed to each other. You can’t take the average of Bach’s St. Matthew Passion and Tchaikovsky’s Violin Concerto and wind up with something equally amazing, or likely even amazing at all. The average destroys both. You can’t even take the average of Beethoven’s Fifth Symphony and his Moonlight Sonata, again, you’d destroy both. To return to our example, I don’t think what you’d get halfway between Pepe Romero and John Williams would sound all that great either, as great as either one is. You simply can’t do it. In order to have great music, you’d have to either copy one or the other (and then you’re just flat out copying) or you’d have to bring something which I really do think is intangible and fundamentally human, something which we talk about with words like artistry or taste or wisdom.
Something which machines will never possess.
We can speculate about why exactly machines do not have this. They may vaguely echo a human who had it, but they do not themselves. My claim would be that this has to do with the fact that humans are moral creatures with immortal souls, and the most profound wisdom is not about matters of fact, but matters of moral judgement, matters whose perception depends on the fundamental character of the person judging. Human choices, even sometimes those about seemingly trivial things, ring with an infinite significance. As Robert Frost put it:
Two roads diverged in a yellow wood,
And sorry I could not travel both
And be one traveler, long I stood
And looked down one as far as I could
To where it bent in the undergrowth; Then took the other, as just as fair,
And having perhaps the better claim,
Because it was grassy and wanted wear;
Though as for that the passing there
Had worn them really about the same, And both that morning equally lay
In leaves no step had trodden black.
Oh, I kept the first for another day!
Yet knowing how way leads on to way,
I doubted if I should ever come back. I shall be telling this with a sigh
Somewhere ages and ages hence:
Two roads diverged in a wood, and I—
I took the one less traveled by,
And that has made all the difference.
Can machines take the road less traveled, and can it make all the difference? In a way, certainly, a machine can be programmed to do so. But if, as it is in our example of music, the road less traveled actually means “the statistically less likely thing in a given context at a given hierarchical level” then no, machines, at least as they are set up with modern generative AI, cannot do it without simply descending into chaos. In any event all the difference will ultimately be made, not by a machine’s deterministic programming, but by the free choices of a human person, the person who set the machine in motion in the first place, the person made with only one life to live, and with something to lose, others to inspire, and heaven to win.
And so, while I have little doubt that AI, much like the guitar and the violin before it, will be one day used in some capacity as an instrument in the creation truly great music, it will be by the human care and attention to the particular piece written or played, the choices made, the honor risked, the beauty seen and pursued, the love lived, that the machine which matches patterns will be made to deviate into true greatness.
If it is true that AI will never create truly great music, one is led to wonder - how many other distinctively human things are there, and what are they, and will AI, far from overshadowing human beings, actually in its inability to reach these human things finally accentuate their greatness? On the other hand, what things are there which until very recently seemed strictly the province of human beings, like the ability to construct and carry out a relevant and helpful chain of logical reasoning, but which is in fact eminently doable by machines?
It is an exciting time to be a person, as developments with our sophisticated bits of silicon lead to a profound re-examination of what exactly, and who exactly, we are. And we can create truly great music while we do it.
If music be the food of love, play on!