One of my back burner writing projects is a book chapter about generative AI in music education and why I think it’s a Bad Thing. In preparation, I reread Ted Chiang’s New Yorker essay about why AI isn’t going to make art, which I completely agree with. To put ourselves in the right frame of mind for this discussion, let’s enjoy some beautiful patriotic AI output.
Anyway, I’m thinking about this paragraph from Chiang’s essay:
Whether you are creating a novel or a painting or a film, you are engaged in an act of communication between you and your audience. What you create doesn’t have to be utterly unlike every prior piece of art in human history to be valuable; the fact that you’re the one who is saying it, the fact that it derives from your unique life experience and arrives at a particular moment in the life of whoever is seeing your work, is what makes it new. We are all products of what has come before us, but it’s by living our lives in interaction with others that we bring meaning into the world. That is something that an auto-complete algorithm can never do, and don’t let anyone tell you otherwise.
Chiang defines the creative process as a huge series of tiny choices. Every word choice in a story, every brush stroke in a painting, every note in a composition, every knob twist in a recording: the work of art is the result of all of those choices, and the artist communicates through their choices. Since the AI isn’t communicating and it’s not making choices, it can’t make art, however much its output superficially resembles art. Chiang compares AI slop to the eyespots on a butterfly’s wings: they look like predator eyes, but fundamentally are not.
Chiang makes a convincing argument for why AI slop sucks, but I also think it’s a useful new conceptual category for bad art made by humans. Sometimes art is bad because it’s inept; that’s not what I’m talking about. I’m talking about highly polished products made by professionals that leave your senses feeling dulled.
Think of the emptiest, most boring top 40 pop song you know. We all have our anti-favorites. For me, it’s “Levitating” by Dua Lipa. On first listen, it’s fine, but the more I hear it, the more it bothers me. All the ideas sound like they were interpolated from other pop songs. Every phrase in the melody sounds vaguely like a phrase from a better song. Every synth and drum sounds like everything else on the radio. The track has been through too many hands, there were too many meetings about it, you can feel the corporate consensus. There’s no point of view, no possibility of specific thoughts or feelings being communicated. I have no problem with mainstream pop, and I also have no problem with that mainstream pop being derivative. However, if the music is so derivative that it sounds like AI slop, that is a problem.
AI slop is also a useful aesthetic category because it enables us to point to actually good pop songs and say, this does not sound like AI slop. We can identify the aspects that you couldn’t possibly arrive at by interpolating from a giant database of already existing pop songs. When you listen to Jimi Hendrix or Michael Jackson or Talking Heads or Björk, sometimes you hear them making contact with the consensus opinions of their time and place, but they frequently surprise you with awkward and counterintuitive choices. This is true for entire stylistic movements, too. I love 1980s hip-hop because there’s no way to predict it from projecting trends in 1970s pop. It questions basic assumptions! It’s missing elements that you might have thought were necessary (like melodies) and it brings in elements you didn’t know you wanted until you heard them (like samples and turntable scratching.) Now that 1980s rap exists, it’s easy to feed it into an AI as training data and get more 1980s rap, but there’s no way that AI could have produced 1980s rap if it was only trained on 1970s pop.
It’s a common assumption that technology drives pop music. However, many of the musical developments that we think of as resulting from technology actually predate the technology itself. For example: it would be easy to say that loop-based pop song structures and perfect metronomic timekeeping are the result of drum machines, MIDI sequencers and DAWs. But it turns out that those aesthetic qualities were present in pop music well before the technologies were widely available. “Stayin’ Alive” by the Bee Gees and “Dreams” by Fleetwood Mac are both built on top of drum loops—they’re tape loops, not digital samples, but the effect is the same. People heard those songs and thought, oh yeah, this is the sound, I want to have this kind of steady and predictable groove! So musicians like Nile Rodgers and Bernard Edwards worked hard at machine-like timekeeping in their live instrumental performance. DAWs and MIDI made that style of groove vastly easier to attain, but by then, the style already existed. And by the way, this is not to equate Chic’s aesthetic with AI slop. I love Chic! Just to say that Chic’s music came before DAWs made it more attainable.
AI slop existed long before AI did. It’s all the boring, committee-driven copycat filler that has been clogging up the pop charts forever. I think that our job as educators is not just to push back on AI slop made by AI, it’s also to push back on AI slop made by humans. Bob Dylan and the Weavers are better than Peter, Paul and Mary because Peter, Paul and Mary sound like an AI trained on Bob Dylan and the Weavers. The Beach Boys are better than Jan and Dean because Jan and Dean sound like an AI trained on the Beach Boys. Eric Clapton sounds like an AI trained on Albert King. The Eagles sound like an AI trained on Neil Young. Everybody is inspired by somebody, but there’s inspiration and there’s finding correlations in a giant metaphorical database, and they are not the same thing.
In class a few weeks ago, I wanted to talk about The Velvet Sundown, but I misremembered their name as Velvet Sunset. I was playing one of their songs and saying, see, this is the kind of bland, featureless pap that you get from AI. But then someone in class pointed out that we were listening to the wrong band. As it turns out, Velvet Sunset is actual humans. I’m sorry, Velvet Sunset! But you know what? If your music is so easy to mistake for AI, that is on you!
It has been suggested to me that commercial pop music has to be AI-like in order to come across, but that isn’t true. Maybe you can only find moderate success by interpolating what’s come before you, but the truly huge generation-defining smashes are always weird and are frequently annoying. Good commercial art is much more similar to good personal art than either of them are to AI slop. George Michael is closer to Daniel Johnston than he is to the Velvet Sundown.
By the way, if you haven’t read Ted Chiang’s short story collections, do not deprive yourself another minute. They are haunting and excellent.
This discussion also reminds me of a thing Keith Johnstone said that’ll always stick with me: “You should be as obvious as possible. At its heart, your obviousness is unique because it is only obvious to you.” Context: https://creativecreativity.com/2017/12/03/keith-johnstone-and-obvious-creativity/
BTW mention of Fela Kuti reminds me of one of my FAVE examples of building on earlier work, which is Ryuichi Sakamoto’s “Riot in Lagos”, which is an homage to Afrobeat with drum machines, and then ended up massively influencing hip-hop breakdance etc. It still sounds catchy ‘n’ fresh today!
> But there’s a difference between transforming your source material and merely interpolating between pre-existing data points.
^ ➕➕➕➕ YES this gets to the heart of it! 💖
I believe “AI slop” is passing the buck of what was originally “human slop”, and like how you articulated it. I’m a big Ted Chiang fan (gosh how long until we get another story?!), and what we’re seeing is how AI EXCELS at well-worn tropes — like Top 40 pap-pop that Suno/Udio can do better than human now, arguably, and sugarcoat clichés better than a Frankensteined Max Martin. Our reality is where the vast majority simply can’t tell “Is it AI?” when it comes to common forms, whether they be visual or audio.
But where it gets harder are those individual maximalized edges… I’m still not getting satisfying generations of rarer idioms like IDM/braindance. Maybe someday, but there are still plenty of good reasons to make the music YOU want to hear that you aren’t getting anywhere else. The aggregate of those many small, purposeful choices, paying creative dividends.
Yeah, recent Drake and Taylor Swift tracks might as well be AI output trained on earlier Drake and Taylor, but meanwhile all the edgy weirdos are perfectly safe. Even if you could get decent IDM out of Suno, it’s not like Suno is going to be generating the next IDM-like idiom.
You aware, I suppose, that the second half of this article contradicts the first part. You initial premise was that AI can never be art because art fundamentally is about communication and relating to people (even if the art is mediocre and derivative) and a machine can never relate to people.
Then, in the second half if the article you trash genuine artists as “AI slop before Al” whose work is a little more derivative than you would like.
It’s hardly worth mentioning but JS Bach being very late Baroque was very derivative of earlier Baroque composers (and was nearly lost to history for 100 years under Mendolsohn rediscovered him) and yet today is considered the definition of the best of Baroque music (perhaps because he perfected and improved what he took from his musical ancestors). Likewise, Handel is considered almost divinely inspired despite near plagiarized some of his contemporaries.
While I agree with you on AI, I prefer not to trash real human beings.
There is nothing wrong with being derivative. It’s unavoidable! I don’t think there’s any creativity without it. But there’s a difference between transforming your source material and merely interpolating between pre-existing data points. The members of Talking Heads are transparent about who they are imitating, from the Velvet Underground to Kurtis Blow to Fela Kuti, and you can clearly hear the ideas they took from those sources, but Talking Heads still manages to sound utterly unlike any of those people, they just sound like themselves. I trash Eric Clapton because he sounds like a smoothed-out average of a bunch of better blues players. I trash Peter, Paul and Mary because the entire project was conceived by Albert Grossman to specifically be a less political version of the Weavers and a smoother version of Dylan.