Great Artists Steal

AI music won't be weird for long.

My creative process never begins with an idea. Music doesn’t hit me like a bolt of lightning. I sit down, I play a note or three, and I listen.

But I didn’t start by listening to the notes I play. I started by listening to the music of others: my mother lullabying me to sleep, the dulcet trumpets of Herb Alpert’s Tijuana Brass, VeggieTales. My ears are trained on everything I’ve ever heard. If I like a song, I listen to it more. It becomes part of my musical vocabulary: a network of sounds that make me feel things.

I play notes and more notes until a sequence of them together makes something that I find beautiful and moves me, the way music I like moves me. Perhaps I put words to that music. Eventually I’ve written a song. My mother’s melodies, the singing christian vegetables, and a steady diet of Sufjan Stevens coalesce into what I call my “taste” in music, and it’s that taste that guides each note. 

The process by which a music-generating AI is trained is similar to my own creative process. It listens in its own way, hearing a lifetime of songs in a minute. It learns a taste through what it listens to in those few moments. It notices patterns. It listens some more. It begins to create.

The first time I heard a piece of AI-generated music, I felt queasy. I don’t even remember what it was—Drake singing Dolly Parton, perhaps. I remember thinking it was interesting, funny, and completely threatened the sanctity of that method of composing I’d come to value so much. There was an air of uncanniness about the robotic collage, sure, but nothing that couldn’t be sorted out with a few months of debugging.

Nevertheless the chasm between process and product has existed since sound waves first hit wax. Recorded music means the artist need not be present for the listener to hear the music. Little by little, that chasm has widened. Multitrack recording, for example, made it so the musicians themselves do not even need to be playing at the same time or in the same studio, but we hear them together nonetheless. Sometimes, as is common among “bedroom” artists of today, a single producer is creating the whole song, dubbing layers upon layers. AutoTune, too, has become a staple of recorded music, allowing a singer to sound pitch-perfect, ironing out the natural drift and vibrato of the human voice. 

Each of these advances has at one time been decried by luddites as ruining music by imbuing it with an air of dishonesty. Notwithstanding the merit of those complaints, none of these technological innovations have cleaved the relationship between artist and listener the way that AI-generated music will.

I expect that within the next five years, the majority of music people “discover” on their preferred streaming apps will be entirely artificially-generated. Furthermore, I expect that with inevitable advances in computing efficiency, these streaming companies will be able to create new pieces of music on-the-fly, algorithmically tailored to listeners’ extrapolated preferences. One’s musical taste may become as insular as one’s social media feed.

Of course, I’d be remiss to say I never benefit from AI as an artist. I enjoy a healthy listening audience on Spotify not because of any marketing budget but because the algorithm, for reasons unknown to itself and me alike, just “likes” my music. I’ve known relative stability as an artist because my recordings have been recommended to people by an app, and I get to connect with a wide international audience because of that. 

In my creative process, too, I benefit. Tasks such as mastering have become easier because music software can smooth out dynamics enough to make a song listenable on any device. The result is far from perfect, but given enough time, this art at the end of the recording process may also find itself flattened into a robotic craft.  

The only element of recorded music that can survive technological advances is listening itself. We will continue to listen to songs, and they will continue to move us. They will be divorced from the context of a creator, however. We won’t be aware of poignant details of the artist’s life, for example, that may lend a song more emotional weight. Risk will be artificial, as AI music carefully restrains its creations away from the edge that your taste can tolerate. Auteurship in our listening diets will fade, and perhaps we’ll accept it as readily as we accepted that Duke Ellington doesn’t live in our gramophones, ready to play “In a Sentimental Mood” at the drop of a needle. 

Given the inevitable wave of extinctions that these unfeeling songmakers will precipitate, we must ask ourselves whether the artist’s survival is worth slowing down progress. One way to incentivize streaming companies’ continuing to pay artists would be to impose a royalty system whereby any artist whose work has been ingested by a model is paid whenever that model’s work is streamed. Even though the artist’s song is not heard per se, they can still reap some financial reward for their efforts. 

However, this proposal falls short in at least two ways. For one, new recording artists without a body of work would be left out. If the most widely listened to model is trained on popular music up until 2021, an artist whose first work was released in 2022 will have missed the boat and find themselves without a royalty check. And since companies like Spotify pay artists using a revenue-share model, artists’ royalties from playback of their own songs will approach zero as more computer-generated music is added to streaming catalogs.  

More crucially, models will soon (if not already) be trained on the works of other models, weaving an ouroboros of AI-inspired AI music. Royalty-collecting agencies will need to use some sort of forensic analysis to determine what artists are owed. It will become increasingly difficult if not impossible to do so, as each layer obscures the original authors’ works. 

There are not yet any viable solutions for the existential crises facing recording artists. Technological advances favor profitability, not patronage. Nevertheless, if we want to continue listening to the artists we love, and build relationships with new artists’ music, we as listeners must be disciplined enough to ask ourselves how artists can survive in spite of capitalistic conveniences. In other words, do we care if the artist is present?

Meanwhile, as an artist I wonder how much of my creative process I’m willing to leave up to the judgment of machines. And once I’ve finished a song, does it need a listener?