The Absentees

When I started making AI music it was 2020, and I had a lot of time on my hands. OpenAI had just released Jukebox, and I’d found my way to a Collab Notebook to run it. I generated a series of samples — usually about 14 seconds — and brought them into my laptop’s Digital Audio Workstation (DAW) for sound, and I’d chop the samples up and write music around them. That was the birth of The Organizing Committee, and one of the ways I had integrated AI into my music workflow.

I can’t lie: it was incredibly inspiring, in part because AI became a vehicle for my imagination. It asked me: what kind of genres do I want to mash together? What kind of band do you want to hear? I’d only get a 16-second glimpse at max into that possible world, and it would give me something to move toward as I made it into something bigger. I thought deeply about the role of influences on my sound; what pieces of the things I loved I wanted to smash together into something new.

I say this because I don’t think a lot of folks know what it’s like to be intoxicated, as an artist, by generative AI. It is a way to tease out a glimpse of something possible, a way of giving you a taste of something you might want to build out.

Since then, it’s become much more complicated to make these claims. I first started feeling a pang of concern back then around training data. This was 2020, before any of this was a mainstream conversation. I started making beats based on the Wu Tang Clan. When I got those beats back and started sampling them — (I never used the resulting tracks, for reasons I’m about to explain) — it felt too much like the Wu Tang Clan.

It felt like digital blackface. Here was an archive of these songs, taken without permission, written by black men, and here I was, an AI Elvis, about to reap the result of their legacy in ways they wouldn’t ever benefit from. I know I’m not, you know, a threat to RZA. But the light bulb over my head reminded me of my work with race and GANs in 2018. Wasn’t this just the same as white performers making music on the backs of black musicians who never got credited?

So I stuck to weirder samples — I fused noise and Francois Hardy songs, because look, she’s rich enough already. I had invented a more crass version of the ethics I’ve since heard elsewhere from AI artists: “don’t prompt with anyone who still has to pay rent.”

Eventually I stopped using Jukebox, because it wasn’t really how I wanted to make music. Most of The Organizing Committee records barely use it, and are written and composed on my own, with the occasional dip into some interesting model or process.

AI Music

This week I’m in Dublin, Ireland to talk with the Music Current Festival about AI and music and “AI music,” or what I am calling Gaussian Pop. As part of my prep for that, I’ve been going deeper into a new AI music generator called Suno, which in essence has taken OpenAI’s Jukebox (a GAN-based system) into the Diffusion era.

My feelings are complicated, because I know I’ll cringe when I see Suno’s training data. But I have to say: Suno is the first AI product I’ve been impressed by in about 4 years.

It’s much like Midjourney, in that you can describe a song you’d like to hear. It will generate between 30 seconds to two minutes of fully “recorded” audio in response. You can extend that, or tell it where you’d like to splice in and regenerate from. Then you have a bunch of disconnected fragments, and Suno will piece them back together seamlessly and up-sample the track to a single mp3 file.

The sound quality is like an early 2000’s CD rip. It’s slightly worse than Spotify streaming gives you. But what’s compelling about it is that it’s a complete track, with customizable lyrics, that follows your prompt. (Jukebox, by contrast, only let you identify up to two genres and two artists).

The cringey examples people are posting online are reflections of bad taste and/or AI-drunkenness. I admit I am at risk of the second one myself. The moon song I posted last week is not “good,” but it’s technologically impressive.

But look, I’m not here to sell Suno (if you’re wondering, nobody is paying me to hype anything here). My thoughts aren’t really about the company or even the software.

Instead, I’m remembering what it’s like to be AI-drunk.

It’s disorienting: suddenly, I don’t know if I like what it’s making, or if I am just impressed that it can make it. I have to recalibrate my ear for music. At the same time, it’s like gambling: you hit a button and you see what you get. By then the song is basically complete, and you say yes or no. You can extend it, or render the song into a file and be done.

That’s quite different from when I make music in a DAW — which lots of musicians don’t think is “real,” either.

But often, with Suno, I just hit “next.”

Then what? Do I settle for the new fragment? Do I regenerate an alternative minute? Crucially, I have to ask: Did I like the thing, or did I like the feeling of shaping the thing? Did I like what I discovered, or did I like discovering it?

I’m easily immersed in imaginary problems — I teach storytelling for video games. Most of that storytelling is about tapping into the imagination so the user tells their own story about their interaction with an interface. I can play Rimworld in four hour stretches, I can make AI music for even longer.

I like shaping it. I think of the genre and what words I used to get a result. Suno doesn’t let you use artist’s names, so I describe genres and vibes and instruments. I’m arriving at a theory of what Gaussian Pop is. It’s noisy, because shaping and constraining noise is how diffusion makes pictures. It’s also how diffusion makes songs. So I use white, pink, and brown noise to create these weird soundscapes: not songs, but sonic spaces.

Sometimes the process creates melodies. I want those, but as with images, I want to push away from the training data. So I prompt Atonal: the absence of melody. I get two minutes of noisy, pulsating treble. So I add some things to get me closer to the center of the training data, but not so close. Then a thing happens, a song. And I start piecing it together.

It’s a flow state, and it involves a kind of creative rumination. But that rumination has less control over expression than I usually have, when I write or compose music, or when I play Rimworld. In a sense, what I share when I share a piece of AI generated music is an artefact of rumination, severed from individual control over expression. It’s a weird form of creativity, and a disorienting form of art “making.”

Some of these tracks I listen back and say, why did I like that? Some of them, I get chills. It’s uniquely self-alienating. Art without craft. Putting the album up on Bandcamp — even when I know I am engaging with this technology critically, and making music that ideally pushes against itself — is still a bit bleak. It feels like being the avant-garde to the automated, inevitable enshittification of Bandcamp. But at the moment, there’s nowhere else to make this available. And, look: as experimental music goes, I think it’s worth sharing.

The Absentees

I wrote a song about making a song. It’s called “The Absentees.” I’m sharing it here, as evidence of a thought process. As a way of sharing the words fused with a vibe. Which is what music-as-communication is, and it succeeds or doesn’t, based not on my level of engagement with its production but by whether it expresses that fusion properly.

A voice without a singer,
Music from an empty room.
A house without an occupant
or maybe, it’s all of you?
A voice without an occupant -
Or maybe it’s all of you.

What ghosts haunt
the absentees?
What should we do
What ghosts haunt
the absentees?
What should we do
with all of your memories?

This song is whatever you want it to be, but it’s also about training data, and the phenomenon of expressing myself through an archive constrained by the performances and emotions of other people. Perhaps even you.

I’m trying to create lyrics that call attention to the emptiness of expression, before we all forget. Some of the songs are singing punctuation marks or code mark-up.

I made an album with Suno, because I like to keep examples of how people are making AI generated music in the archive at Latent Space. It’s not The Organizing Committee, but something else, more suitable to Diffusion and training data and noise as a medium: The Absentees.

You can listen to the whole thing if you want.

I’m getting interesting textures — maybe not “songs” — by requesting combinations of noise: white noise, pink noise, and brown noise. Because diffusion models strip noise out of sound, my theory is that the result of this noise prompt would be something interesting. But it turns out that sonically, it’s still just noise. Sometimes there are patterns of clusters in there that become a sound, or the noise moves into something resonant and harmonic. Likely, that’s the model’s movement toward something recognizable to the training data.

It’s complicated, because noise is also a sound, and a genre. It’s also not something people usually like to listen to, so many of these experiments, at the moment, are really challenging, akin to glitch music and drone. Meanwhile, the model is capable of generating anything from dance pop to acoustic folk field recordings. Even with noise, it sometimes generates “songs,” especially when I have lyrics.

The record is a collection of experiments made from attempts at either creative misuse, or pointing out the challenges of creative AI in the first place. What, or who, is absent from this music: “The Absentees” is the name of the band to call attention to it. Though a lot of the “doo be doo” type tracks are a result of asking the system for unpronounceable lyrics: punctuation marks, such as “;;;” or “!!!”, or singing the ASCII art characters of a paper clip.

I would never change my workflow to make AI generated music exclusively. But I am interested in what creative misuse of the system sounds like. The song linked above is not a direct result of that process. A few — the tracks titled “Gaussian Splat” 1-3, are more in line with that approach.

What Debris?

If you know what you want, Suno is not for you. When I had lyrics and a kind of tone in mind, I found it was more frustrating to generate something with Suno that matched what I wanted to hear. Abstract experiments, where I am asking the machine to process white noise fields, were interesting for me. Precise song ideas were, perhaps obviously, really difficult.

“What Debris?” was meant to be a specific sound: I wanted noisy, 1920s style recordings with lots of ambient room tone.

Ambient room tone, and field recordings, are an interesting aspect of sound generation. Like radio pieces, or images, or language, there is a scanning for signal that our brains naturally gravitate toward. If we hear a recording of a room, aren’t we looking for evidence of something in that recording? But nothing is delivered to us, just as there’s no intent in the production of an AI generated image or text. With sound recordings, we know it’s not a recording of anything. Still, we want to listen. These tools are built that way, to direct our attention to empty rooms that don’t exist. (“What are you listening for? There’s nothing here.”)

Point is, if you’re a musician and you want something precise to come into the world, it is a different process from being a musician who wants to discover a version of a thing that is inside of your head. The negotiation is different: from “what will I do to the system if I do this?” to “How can I get the system to make this thing that I want?” We can argue over whether one is “creative” and the other isn’t. I think they’re different forms of creativity, one of which is expressive: the idea in my head makes it into the world, or the idea in my head accommodates the variety of possible responses within the boundaries of possibility that I define.

Generative art, in the style of folks like John Cage or George Brecht, is not really about expressivity, but they couldn’t help but to express an idea: an idea of generative music, its value, and its limits. Making music, or art, with AI is confusing for similar reasons. It’s only expressive because I set the confines of my expressivity to “see what the machine does and make it visible.” The rest was filled in. This is different from “I want a song that sounds like a ghost singing from a 78rpm record with lots of ambient room tone” — then getting a disco track and sharing it anyway.

Comparisons

I don’t have the confidence to know if the results are interesting to anyone else. Music is so subjective. That’s why there’s a shred of bravery in sharing the songs you make, when you write your own music: people you love can just not dig your sound. It’s perfectly reasonable.

That’s true of all my music making. But when I’m writing a song, I listen to smaller fragments over and over and over again. The music I write and perform evolves through that attention and scrutiny and the changes I can make are literally beat by beat. With AI, some of that investment is stripped away.

I’m reminded that the history of generative AI has been marked, chiefly, by longer capacities for what’s called “attention.” We used to predict one word based on the most recent word, then one word based on the last 5, then 7, etc. Now they can look at whole essays and write whole essays.

But each word, each drum beat, is an invitation to human attention and attendance. I listen to smaller chunks of the work as I write music — literally starting at individual beats and rhythms, then to loops, then to song structures. I shift between all of those scales of attention regularly.

Suno gives me two-minute bursts, and a decision to say yes, no, or revise.

I like some of these songs. I want to hear them on repeat. Suno is perhaps closer to Spotify than any musical instrument I’ve ever worked with. Of course, Suno can be both. If Spotify was paying musicians, I might be worried for artists, but at the moment this just seems like one more batch of giant-label licensing deals waiting to happen.

The politics of training data are gross. But I wanted to make a record of what it is like to engage with generative AI when it still feels impressive, and when you still hate it, and are equally torn between being excited and wanting to break it. To show what my thought process is like, and how I navigate the tensions between creative and expressive mind states. I think that’s clear here. I am disoriented between thinking of it as unlistenable, hypothetical music and something that says something about unlistenable, hypothetical music.

I hope you hate it, and it ends all AI music forever. But also that you like it, somehow. I want to believe I found a way of making this music my own. But I don’t want the music to carry the responsibility of proving that.

Things I Am Doing In April

Dublin, Ireland April 2 & 3: Music Current Festival

Flowers Blooming Backward Into Noise, my 20-minute “Documanifesto” about bias and artist agency in AI image making, will be part of an upcoming event at the Music Current Festival in Dublin. The Dublin Sound Lab will present a collaborative reinterpretation of the original score for the film alongside the film being projected on screen. It’s part of a series of works with reinterpreted scores to be performed that evening: see the listing for details.

I’ll also be leading a workshop for musicians interested in the creative misuse of AI for music making, taking some of the same creative misuse approaches applied to imagery and adapting them to sound. (I’m on a panel too - busy days in Dublin!)

London, UK: All Month!

As part of my Flickr Foundation Residency I’ll be in the city of London for the majority of April, with a few events in the planning stages in a few UK cities. I will keep you posted when dates are final! But if you’re in London — or Bristol or Cambridge — keep a lookout, or let me know if you want to do something.

Sindelfingen, Germany April 19, 21: Gallerie Stadt Sindelfingen

For the “Decoding the Black Box” exhibition at Gallerie Stadt Sindelfingen, which is showing Flowers Blooming Backward Into Noise and Sarah Palin Forever, I’ll be giving an artist talk, “The Life of AI Images,” at 7pm on the 19. On the 21st, tentatively, I’ll be part of a daylong workshop showing new work linked to “The Taming of Chance,” a project by Olsen presented alongside new works by Femke Herregraven, the !Mediengruppe Bitnik, Evan Roth & myself, who will all be there in person.