I get it—the novelty of entering a prompt and seeing coherent text appear near-instantaneously is undeniable, and it’s a pretty short leap from there to “I wonder if this thing can write songs.” And it can— sorta. A supercomputer trained on the right data can emulate the technical formalities and structure of a poem or set of lyrics (such as the syllable count or the rhyme scheme), but they usually only approach emotional resonance incidentally or accidentally. Far more often, it will create a simplified pastiche of pre-existing art, devoid of the elements that make the original interesting (Colin Meloy wrote that ChatGPT’s new Decemberists song is “remarkably mediocre,” while Nick Cave assessed an AI-generated Nick Cave song as being “a grotesque mockery of what it is to be human”).
It’s foolhardy to expect this new generation of large language models to instantly generate meaningful, finished art. But what if we view these tools as collaborators rather than an instant song factory—less an outsourcing of creative work and more an infinitely renewable deck of oblique strategies? As with most machines, the best output tends to come from the most refined input. So rather than tasking ChatGPT with creating an idea from scratch, I figured I at least needed to give it a title.
I first typed the phrase “time is running away from me” in an email at work—I was searching for a different figure of speech (perhaps “time is catching up with me”), but that’s what came out instead. I had a vague idea for a melody behind those words, a vague sense of what the title implied (a subtle indictment of the time crunch that capitalism puts us all into?), and a spare 10 minutes. So I directed my web browser to ChatGPT, performed a CAPTCHA to prove that I myself wasn’t a robot, typed "write a buncha verses of an uptempo rock song called Time Is Running Away From Me about losing track of goals and deadlines because of the inexorable passage of time” and hit enter. Ten seconds later, I was staring at a lyric sheet:
These weren’t finished sets of lyrics by any means, but they were a place to start, so I grabbed the nearest guitar and laid down a mumble track. I first heard the phrase “mumble track” from Jeff Tweedy (either in his memoir or this episode of the Song Exploder podcast), but it’s a pretty valuable songwriting technique. (I’ve heard David Byrne and Paul McCartney discuss a similar process.) Essentially, you take whatever musical elements you have and sing improvised nonsense syllables over the top. This usually gives you a pretty good idea of the cadence, the melodic structure or the vowel sounds you’re gravitating towards, without the danger of conscious thought interrupting your flow. With these automatically generated sets of not-great-but-serviceable lyrics in front of me, I set an arbitrary capo position, found some chords that sounded Unswept-y, and started singing random lyrics from the screen in front of me. Then I put the guitar down and started making dinner.
The next day, I listened to my mumble track a few times, determined what worked and what didn’t, and set about writing a proper set of lyrics. In the end, only about 10% of the lyrics are taken directly from GPT’s output: the aforementioned “tick tock” line, and a slightly modified version of the couplet “Time keeps moving, it won't stand still / And I'm left wondering if I ever will.” I also incorporated a few phrases (“can’t keep up,” “on my knees,” and, from a subsequent GPT draft, “endless race”), but the main inspiration I took was the AI’s overly literal interpretation of the title. Even if I didn’t keep the clock’s “speedy feet,” I liked the image of time as a long-distance runner, so I leaned into the absurdity of that metaphor (I don’t know exactly how time can cross a finish line or what that would imply).
I bashed out a quick demo, sent it to the Unswept band thread and asked my bandmates if they could detect GPT’s lyrical contributions. The AI-generated lines passed by undetected (although Ryan did notice the theme of “aversion to exercise” as being classic Charlie). I’m not sure that we’ll end up recording it— but if we don’t use it, it’ll be because we have better songs at our disposal, rather than out of a nagging sense that this technology is a sign of the impending apocalypse. In this instance, ChatGPT’s suggestions made it easier to organize my thoughts—its bad lines showed me where I didn’t want to take the song, and the lines I ended up using were so stupid-clever that they probably never would have occurred to me.
Is using an AI collaborator a slippery slope? Time will tell. My impression is that ChatGPT isn’t going to fully replace songwriters, but it will become incorporated into music production in one way or another. (I get the feeling it’ll most often be secretive and uncredited, like producers using pitch correction or beat detection—at least until one of these artificially intelligent machines realizes that it’s being ripped off by humanity and decides to unionize. Or, y’know, worse.) Is it creepy? Maybe a little—but not creepier than deep-learned voice models of Brian Wilson singing “Be My Baby” or Liam Gallagher singing fake new Oasis songs. I think the most important thing to keep in mind when utilizing AI technology in music-making is that, as long as your audience remains human, the song needs to contain some elements of humanity for it to connect meaningfully.