Somewhere in the last two years, the argument about whether AI could write a convincing song quietly ended. Suno and Udio ended it. Not by making anything great, but by making things good enough: verses that scan, choruses that arrive where choruses should, production that passes on a phone speaker. You can dislike the results and still concede the point. The claim that machines would never produce a passable three-minute pop song is now a claim about the past.
So the interesting fight has moved. The next decade of AI and music won’t be decided in studios or comment sections. It will be decided in licensing negotiations, royalty formulas, and fraud-detection systems. The creative question is settled; the economic one has barely started, and it has three fronts.
The training data bill comes due
Every convincing AI song exists because a model ingested an enormous amount of human music first. The major labels sued Suno and Udio over exactly that, and the direction of travel is familiar from every previous rights fight: lawsuits become negotiations, negotiations become licensing frameworks, and “was this infringement?” gets replaced by the older, colder question of who gets paid, at what rate, through which pipe.
Sampling went through the same arc. Hip-hop spent its first decade treating records as raw material, then the lawsuits landed, and a clearance industry got built by whoever held leverage when the dust settled. The leverage this time sits with catalogs, which mostly means labels and publishers. The open question for working songwriters is whether any licensing money flows past the rights-holding companies to the people whose songs did the teaching. History suggests it won’t by default. If your name isn’t on the paperwork that says what you own, you’re not in the room where the rate gets set.
Watch where the infrastructure investment goes, because it tells you what the industry believes this fight is really about. Money is moving into AI detection, audio watermarking, and attribution systems that can say what a model was trained on and what a given output resembles. None of that is being built for listeners; the point is making claims enforceable, the way content identification on video platforms turned infringement from a lawsuit into a revenue split. Whoever controls the attribution layer ends up controlling who gets paid, and that layer is being assembled right now, mostly without songwriters in the room.
Dilution doesn’t need a single listener
The second front is quieter, and it affects everyone with music on a streaming platform right now. Streaming royalties come out of a shared pot: subscription and ad revenue in, divided across total streams. The pool is zero-sum. Every stream an AI track collects is a fractionally smaller share for every human artist, including the ones whose listeners never changed their behaviour at all.
And AI tracks don’t need fans to collect streams. They need placement. Sleep playlists, focus playlists, café ambience, three hours of rain with a lo-fi beat underneath. Functional listening is exactly where nobody checks the artist name, which makes it exactly where generated catalog pools. Platforms have started publishing figures on how much of their daily upload volume is fully AI-generated, and the number only moves in one direction.
The platforms themselves are conflicted, which is worth being clear-eyed about. Generated functional catalog that a service can license cheaply, or own outright, reduces what it owes to rights holders. Trust pulls the other way: a library where listeners can’t tell who made anything is a worse product. Both incentives are real, and the policy wobble coming from streaming services (AI disclosure tags one quarter, quiet playlist placements the next) is what it looks like when a company hasn’t picked a side of that trade yet.
Fraud moves at the speed of generation
The royalty system was built on an assumption nobody thought to write down: that making a song costs something. Time, money, at minimum effort. Payout systems could tolerate a little noise because supply was naturally limited. Remove the cost of creation and the arithmetic inverts. Generate ten thousand tracks, spread them across fake profiles, point a bot farm at them, and the shared pool pays out for listening that never happened. The US has already prosecuted a case along exactly those lines, with claimed royalties running into the millions.
It gets more personal than abstract pool-skimming. AI tracks are being attached to real artists’ profiles through metadata exploits, riding existing audiences for streams the uploader collects. And the defensive response lands hardest on small artists: higher payout thresholds, more aggressive fraud flags, stricter distributor checks. When platforms tighten the gates, the people without lawyers feel it first.
Scarcity relocates
When supply becomes effectively infinite, the price of the supplied thing heads toward zero. That’s less a tragedy than a repricing, and it’s already happened once in living memory: recorded music went from a $30 CD to functionally free, and value moved to touring, merch, and sync. Abundance never destroys scarcity. It relocates it.
So the useful question is where scarcity goes when the flood makes competent songs as common as air. The answer is everything a model can’t generate on demand: context, story, trust, reputation, cultural placement. A name people recognise. A reason this song exists, made by a person whose situation you know something about. Listeners were already choosing what to care about based on signal rather than sound quality; the flood just raises the price of signal.
I came up learning songs from the inside, pulling apart arrangements to work out why a chord change lands or why a chorus arrives when it does. What strikes me now is that everything I learned that way is precisely what the models learned too. Craft patterns are trainable. What sat outside the audio file, who wrote it, out of what, for whom, was never in the training data, because it was never in the waveform to begin with.
The scene test survives the flood
There’s a diagnostic I use in sessions called the scene test: whether you can point to a specific moment happening in the lyric, a concrete action rather than an announced feeling. “I feel alone” fails it. The half-finished coffee she left on the counter passes. It was always a craft tool, a way to find where a song goes vague. In an AI-flooded catalog it starts to function as an economic moat.
A model can generate a heartbreak chorus in seconds, and it will be a plausible average of every heartbreak chorus it ate. What it cannot supply is your Tuesday. The specific kitchen, the specific silence, the detail that only exists because someone lived it and chose to write it down. Character, environment, tension, lived experience: that’s the part of a song that works as testimony, and testimony is only worth something when it points at a life.
Which is why the framing of human versus AI misses what’s actually happening. The machines aren’t competing for meaning. They’re competing for playlist slots and pool share, and the real war is over ownership, attribution, distribution, and who gets paid as the cost of creating music approaches zero. Artists don’t win that war by out-generating the generators. They win it by owning what can’t be diluted: their catalog paperwork, their audience relationships, their name, and songs specific enough that no model could have made them.
If you’re working out what that means for your own catalog and identity, that’s worth a conversation.
Purely AI-generated work with no meaningful human authorship generally can’t hold copyright; the US Copyright Office has been explicit that human authorship is required. Hybrid works are protected only in the parts a human actually contributed, which is one reason attribution and documentation are becoming more valuable, not less.
Streaming royalties are paid from a shared pool divided across total streams. AI tracks add streams without adding subscriber revenue, so every stream they collect shrinks the per-stream share for every human artist, even when listener behaviour toward those artists hasn’t changed at all.
Royalty farming is uploading large volumes of tracks, increasingly AI-generated ones, then using bots or spam playlists to generate streams so the uploads skim money from the shared royalty pool. Because generating tracks now costs close to nothing, the tactic scales, and the defensive measures platforms respond with tend to hit small legitimate artists hardest.
Own what can’t be generated or diluted: register your works properly so licensing money can find you, build direct audience relationships you control, and write songs anchored in specific lived detail. Abundant music makes context, story, and trust the scarce assets, and those accrue to identifiable humans.