John Maynard Keynes warned in 1930 about “technological unemployment,” the shock when tools leap ahead of our ability to rearrange the world of work. Keynes also predicted those tools would eventually buy us leisure time. But as Morgan Housel points out, we mostly raised our standards instead of lowering our hours. Every advance made life better in some ways and more demanding in others.
I was asked recently about my hopes and trepidations around AI, and music, and musicology and my first thought was Housel’s framing Keynes, sharing his words from a hundred years ago but as though they were an op-ed in today’s New York Times. AI is in its phase where we’re reckoning with clear permanent upheaval and dealing with the shock. What’s it going to mean for music? I thought about tech shocks I’ve lived through. MIDI let me connect synthesizers and play one from another; then drum machines from my keyboards. Then I had computers “recording” not the sound, but my keystrokes. And then digital audio recording, samples, virtual orchestras, and now Splice.
Joe Zawinul’s old Korg ad summed it up: “MIDI is a powerful weapon. Any idiot can pull the trigger.” Those technologies were all so enabling, for better and worse, and AI is more of that same tension, upgraded. It will let bad musicians make bad music faster. It will let good musicians make better music faster. Whether AI is a net good or a net bad will be about what we choose to reward. If we prize discernment, authorship, accountability, craft, and taste over throughput, music might improve.
I’m a forensic musicologist. My lane is evidence around “does one work significantly reproduce protectable expression from another?” I’m both optimistic and wary. On the positive side, AI is an instrument, a powerful weapon. In skilled, thoughtful, dilligent hands, it expands palette and pace. Composers can summon harmonies, orchestration, and full-color previews in minutes that not long ago took days. Idea density obviously soars and exploration “costs” diminish (I still have a toe in economics). There will be more experimentation, far more choice among the many quickly realized versions of “the musical idea you had,” and an even faster on-ramp for, I’d love to think, many many talented creators who lack access to the more traditional tools that were yesterday’s bedroom studios, much less actual recording spaces, and players. That’s good for music. (Though even typing “players” brings out the crank in me.)
My trepidation is the Zawinul part. Lower the effort required to produce competent-sounding tracks and you flood the zone with much less effortful output. Musical chops get traded for prompting chops, and the worst part is when people lap it up; it’s like everyone’s pretending not to hear the difference. That new norm, like all technological revolutions, is bound to shift our attention. It’s Keynes and Housel again; we were supposed to get shorter work days, but we demanded more and more stuff. I expect a tidal wave of AI enabled junk music, and worse, an audience that happily to binges. There will be more music and less listening than ever before. And probably that’s where my day job heats up. As music generation increases in tonnage, proximity will too. The successive waves of AI lawsuits in the news will do nothing but increase observations of similarity. We’ll see more works that are “pretty close” to something famous, and famous songs that are “pretty close” to that song you wrote last month. More accusations, some justified, many not, but definitely more.
Speaking of the lawsuits, I was looking at UMG v Anthropic. Right now we’re mostly looking at training and whether the AI’s themselves, by virtue of their process which necessitates making a ‘copy,’ violate copyright law that prohibits such copying, and if so whether there’s a fair use defense for it, and if so, does it extend to training on all materials regardless of how they were obtained, and whether it’s okay to maintain those copies as a runtime library. As a forensic musicologist, I’m thinking:
- AI, I’ll judge your systems by their outputs. Learning from large corpora, to me, seems transformative. I transcribed plenty of music in my day. But obviously regurgitating recognizable lyrics or melodies is different. If ordinary prompts reproducibly yield protected expression, that’s the sort of copying that clearly violates copyright, and guardrails need to work, or musicologists’ reports will abound.
- Creatives, please let these tools be instruments and not autopilots. Be noble. Use AI to expand your own palatte, but don’t outsource authorship and yield mundanity. Keep the originality bar high; and maybe document your process a bit, just in case.
- Music lovers, expect better. Reward intent, taste, and arrangement, and do not settle for “generators.”
I’ll do more preventative work, stress-testing songs before release. More forensics, discerning influence from regurgitation. And as these lawsuits progress, less fighting about the aura and morality of AI, and more about measurable claims: prompts, reproducibility, recognizability, and as ever, substantial similarity of outputs. If Keynes is right about shock, shock absorbtion, and adjustment, and Zawinul is right about the good and bad of easy triggers and trigger-fingers, then the path forward is to be super virtuous about treating AI like an instrument, enforcing the lines against recitation, regurgitation, and derivation
Raise standards, not just outputs.
That’ll be the day.
