AI music is built from the bones of human performance. So why are we calling it artificial?
Start here: before a single AI-generated track was ever streamed, before Suno served up its first suspiciously competent lo-fi groove or Udio conjured a passable soul ballad from a text prompt, somebody had to sit down and play something real. Somebody had to sweat under studio lights, bend a string, hold a note too long, crack on a high phrase, and pour something irreplaceable into a microphone. That sound that human act is the raw material of everything we are now calling “AI music.” And if you are comfortable with that framing, you should not be.
This is not a screed against technology. It is a question about honesty. About who gets the credit when a machine produces a melody that moves you. About whether the label “AI music” is, at its core, one of the most successful misdirection campaigns the music industry has ever run.
The Dataset Is the Dirty Secret
The Magenta team at Google built something called the MAESTRO dataset – Musical Instrument Digital Interface and Audio Edited for Synchronous TRacks and Organization, if you need the acronym. More than 200 hours of high-quality classical piano recordings, every note timed, every velocity captured, every pedal depression logged, drawn from International Piano-e-Competition performances recorded between 2004 and 2018 on Yamaha Disklavier pianos. The whole thing was designed to give AI systems a granular model of human piano performance not just the notes, but the dynamics, the rubato, the phrasing, the articulation that separates a technically correct performance from one that actually breathes.
A 2025 study published in Nature Scientific Reports used MAESTRO as the foundation for training and evaluating deep learning models – LSTMs, Transformers, GANs – on music generation and performance modeling. The paper was careful and honest about what it found: the AI-generated outputs consistently fell below human compositions in perceived expressiveness and emotional depth. The machines learned from human performances. They could not replicate them. And here is the part worth sitting with: they were never meant to. They were meant to approximate them well enough that the difference might not register on a Spotify shuffle.
The gap between approximation and authorship is everything. One of them is a tool. The other is a lie.
What the Machine Cannot Steal
There is a line worth borrowing from a recent essay by Charlotte Dubois at Skaala: “The ghost in the score is the performance itself.” The essay was written about AI music transcription tools — systems that listen to a performance and produce sheet music but the insight cuts to the bone of the whole enterprise. An algorithm can hear the pitch. It cannot hear the ache.
Think about what that means in practice. James Brown did not invent the one-beat drop by running a rhythmic calculation. He arrived at it through years of performing in front of Black audiences in the American South, through call-and-response church music, through sweat and showmanship and the particular kinetic intelligence of a body that knew how to make a room lose its mind. Chopin’s rubato was not a tempo deviation. It was grief. The blue note in the blues is not a microtonal error. It is a cultural wound that found a pitch.
AI systems trained on recordings of these performances captured the what – the intervals, the timing deviations, the dynamic contours. What they ingested was the statistical shadow of the why. The ache, the groove, the communal electricity of a gospel choir rehearsing in a church basement on a Tuesday night – that is not recoverable from a FLAC file. It was borrowed. It was laundered. It was repackaged and sold back to us with a new brand name and a text-prompt interface.
The Lawsuit Tells You Everything
In June 2024, Sony Music, Universal Music Group, and Warner Records filed suit against Suno and Udio in New York and Massachusetts, accusing both companies of massive copyright infringement- specifically, using the labels’ catalogs without permission or compensation to train AI systems capable of generating music at scale. The complaints named artists including the Temptations, Mariah Carey, James Brown, Michael Jackson, and Bruce Springsteen. The labels sought up to $150,000 per infringed song. The companies were accused of being deliberately secretive about what training data they used.
Read that last part again. Deliberately secretive. Because if the public knew exactly which human performances these systems were trained on, the “AI music” framing would collapse under its own weight. The lawsuits described the practice bluntly: the companies copied music without permission to teach their systems to create music, at the direct expense of human artists’ work. If that theft is real – and multiple federal courts seem inclined to think it is – then what the AI output represents is not original machine creation. It is a distillation of stolen human artistry, remixed at scale and served through a clean consumer interface. The output is human music by proxy. It always was.
The Copyright Paradox
The law is having an identity crisis over exactly this problem, and it is worth watching carefully. As of 2026, the U.S. Copyright Office maintains4 that works created entirely by AI without meaningful human involvement cannot be copyrighted. A machine is not an author. On March 2, 2026, the U.S. Supreme Court declined to hear Stephen Thaler’s appeal in Thaler v. Perlmutter, leaving intact lower court rulings5 that purely autonomous AI output is ineligible for protection. The constitutional requirement of human authorship is holding.
Here is the paradox that the law has not yet figured out how to address: the music the AI generates was built from the authorship of millions of human artists – artists whose recordings, whose performances, whose creative choices were ingested, processed, and probabilistically recombined without their consent. The output is unprotectable because a machine produced it. The input was protectable because humans created it. So we have a product built entirely from human intellectual property that belongs, legally speaking, to no one and credits no one.
Who holds the ghost? Not the artists whose performances were fed into the model. Not the consumers who believe they are listening to something new. Certainly not the algorithm, which holds nothing and knows nothing. The ghost floats free, profiting a platform, crediting a machine, and erasing the humans who actually made the music.
The Folk Tradition Problem
In oral and folk traditions, a song is not a fixed object. It is a living thing, shaped differently by every performer, every community, every generation that passes it along. When Alan Lomax drove through the American South with a portable recorder, he was not capturing the definitive version of anything. He was catching one moment in an ongoing human conversation.
AI systems trained on folk recordings do not understand this. They canonize. They pick up a version of a song, model its patterns, and generate outputs that reflect those patterns – which means they flatten variation, smooth out regional accent, and quietly erase the evolutionary drift that keeps tradition alive. That is not creation. That is extraction dressed up as production. The algorithm does not know it is performing an act of cultural preservation gone wrong. It does not know anything. The humans who built and deployed it, however – they know.
The Branding Lie
Here is the argument at its sharpest: if AI music is a statistical remix of human performances – patterns distilled from billions of human creative choices, made by human bodies, in human rooms, for human audiences – then calling it “AI music” may be the most successful branding lie in the history of recorded sound.
It is more accurate to call it crowd-sourced human music, laundered through a machine. The crowd did not agree to contribute. The laundering was done in private. The machine got the credit.
The Only Honest Conclusion
This is not nostalgia. It is not technophobia. It is a demand for accuracy.
The engineers who built these systems are talented. The tools are remarkable. The music they generate can be genuinely useful and, sometimes, genuinely interesting. None of that is the point. The point is the name on the door.
Recognize the human at the root – not because machines are threatening, but because the truth demands it. Because James Brown’s sweat is in that algorithm. Because some kid who played piano at a competition in Vienna in 2012 and never became famous is in there too, their timing data and velocity curves now powering a product they never consented to feed. Because the entire edifice of what we are calling artificial music was built, brick by brick, from the most human thing we have ever made.
The machine did not write this music. It inherited it. And an inheritance you did not earn, from an estate you were never invited into, is not a creation.
It is a heist with a very good publicist.
About The Author
