Most of the AI music debate feels like it’s happening one layer too early.
Artists are (fairly) pissed that companies trained generative models on huge amounts of copyrighted music. The proposed fix that keeps coming up is some version of “AI companies should license the music they train on.”
On the surface, that sounds reasonable. Even pro-artist. But I think we need to be a bit paranoid here about unintended consequences. Because there are two obvious places you can attach compensation in a system like this…
Input and/or output.
Picking the input might quietly hand the entire future of AI music to the exact institutions everyone already hates dealing with.
The Input Model
Say an AI company wants to train a music model. Under a licensing-first world, they need permission to use copyrighted recordings. Who actually has the ability to grant that at scale? Not artists. Rights holders. Labels, publishers, catalog owners, licensing aggregators.
In other words: the people who already sit between artists and the market suddenly become the gatekeepers of permission to build music models.
They negotiate the deals, set the terms, collect the checks, and the model gets trained. Now you’ve got a system where a relatively small number of institutions benefit from the resulting models.
We’ve basically monetized scarcity (music rights) right before creating abundance (AI generation) and that’s… not a stable equilibrium.
Now Flip It
Forget what went into the model for a second.
Look at what comes out.
Imagine I generate:
Toxicity by System of a Down, reimagined as 1970s funk.
It’s good. People actually listen to it. It does numbers. There is an obvious underlying asset here. Toxicity. Instead of pretending it doesn’t exist because the remixer can’t share it…
We release it, point it back to the artist, and route a defined share of revenue back to the rights holders of the underlying composition, the same way licensing already works in other parts of music.
Now something interesting happens. Every person with an AI music tool becomes a potential remix engine for the entire recorded history of music.
Every time a remix gains traction, the underlying rights holders get paid.
AI stops being “the thing that eats the catalog” and becomes the thing that keeps the catalog economically alive.
We Already Have Pieces of This
Music has already been through versions of this problem. You generally can’t just copy someone else’s recording, but U.S. copyright law already has a compulsory mechanical licensing system that allows qualifying covers of released compositions without negotiating every recording individually.
That system doesn’t map cleanly onto AI. Transformative arrangements, sound recordings, voice likeness, and other rights make this messier in the current framework. However, we already let people cover songs without asking permission every single time, as long as the original rights holders get paid.
And the law already recognizes some distinctions that the AI debate tends to flatten.
A composition is not the same thing as a recording. And neither of those is the same thing as a voice, performance, “style”, etc. The current AI discourse keeps smashing all of those into one bucket called “stealing,” which is emotionally satisfying but legally and economically sloppy.
If an AI output substantially reproduces protected elements of a specific song or recording, that raises ordinary copyright questions. If someone clones an artist’s voice, that raises a different set of likeness, identity, and potentially other legal questions.
But if a model generates something new that’s influenced by thousands of inputs, trying to reverse-engineer “who owns what percentage of this neural net” starts to break down pretty fast.
Follow the Money
This is the part I wish more musicians would zoom in on, because the incentives are very different depending on where you attach value.
Training-side licensing:
Artist -> Rights holder -> AI company -> infinite generation
The real leverage sits with whoever controls catalogs and whoever controls models. Artists are mostly downstream of both.
Now compare that to…
Output-side participation:
Listener -> generated track -> identifiable underlying works -> recurring royalties
Now every successful output is a new monetization event tied back to existing music.
One model monetizes access to the past.
The other monetizes reuse in the future.
The Part That Feels Off in the Current Debate
I’m not saying “just let AI companies scrape everything, no rules, lol.” Training rights are a real question. But it feels like we’ve let that question dominate the entire conversation, and I think that might be strategically backwards for artists.
The biggest shift here isn’t “Spotify gets more training data.” It’s that basically everyone is about to have access to something like a studio, a band, a producer, and a remix engine in their pocket.
So the question isn’t only:
how do we stop AI from taking from artists?
It should also be:
how do we ensure musicians can benefit from the inevitable derivation of their work?
One builds walls around data moats while the other builds infrastructure to channel the current. If we’re actually heading into a world of generative abundance, I’d personally worry less about who gets to stand at the gate and more about whether we’re building anything that actually carries value back to the people who made the original material in the first place.
If we get that wrong, we don’t just lose control of music. We also route the benefits the same place they’ve always gone. Inside the walls, inside the moat, and inside the castle.