Sometime in early July, a clip appeared of four musicians standing on a London rooftop while the writer who had recruited them narrated a threat.
He had formed them into a band with a plan so simple it sounded like a prank. They would take the songs of The Velvet Sundown, re-record them properly, in a studio, with hands, and try to rack up more streams than the original. All the profits, he promised, would go to human musicians. "I'm not giving you a penny," he said, to no one in particular, because there was no one in particular to say it to. In the clip, he delivered the line the whole thing was built around. "If you're making AI music, be very afraid."
It is a good joke, and it is doing serious work. The Velvet Sundown, the band being threatened, cannot be afraid. It cannot climb the stairs to the roof. It has no body to stand on it and no nerves to lose. It is a synthetic act, "composed, voiced, and visualized with the support of artificial intelligence," by its own Spotify page's admission. When it first got caught, it denied being AI-made. Then it called itself "Not quite human. Not quite machine," and "somewhere in between." Then it admitted the truth. By late June it had two albums out; a later report added a third. It had promotional images, a backstory, and invented members with invented names. What it did not have was anyone who could be embarrassed on a roof.
So here are grown people, in a real city, aiming a live-performance threat at an opponent that has no legs. The tribute band has always been a strange transaction: you pay to watch someone who is not the star be the star, and everyone agrees to feel something anyway. But a tribute act needs an original to miss. This one has picked, as its object of devotion, a group that was never in a room in the first place. "We're not actually paying tribute," the writer admitted. Of course not. You cannot pay tribute to an absence. You can only point at it.
The official version of this story is a detection story. Can you tell? Should there be a label? A survey Deezer ran late last year found that 97 percent of people could not pick the fully synthetic track out of a lineup, and that half of them felt bad about it. The whole apparatus now going up around AI music treats this as the emergency: the fake has gotten good enough to pass, so we must mark it, or we will all be fooled.
Here is the number that makes the detection story fall apart.
On Deezer, roughly 44 percent of the music uploaded every day is now fully AI generated. That is almost 75,000 tracks a day, more than two million a month, up from about 10,000 a day just over a year earlier. A tide. And that tide accounts for between 1 and 3 percent of what people actually stream. Of that sliver, Deezer says 85 percent is fraud: bots streaming bots, demonetized on sight. Do the arithmetic Deezer does not spell out, and the listening that is not flagged as fraud comes to somewhere near a third of one percent.
Read that again, because it breaks the frame. The machines are not winning listeners. They are barely getting listened to. The flood is a flood into an empty room. So the panic cannot really be about being fooled, because to be fooled you have to be listening, and the numbers say almost no one is.
Then how did The Velvet Sundown get more than a million streams in a few weeks?
Not the way a band does. When one music-data firm looked closely, the band had 325,388 monthly listeners and 1,533 followers. Sit with that ratio. A third of a million people had it playing; fifteen hundred cared enough to press follow. The follower count is not a fanbase; it is something narrower. The tracks had turned up inside other people's playlists: 26 of them in a list called "Vietnam War Music" that more than six hundred thousand people had saved, another 22 in one fan's soundtrack for a teenage TV drama whose last episode aired in 2007, eighteen years before these songs existed. Discovery came through playlists and recommendations, into the slots where attention was already flowing and no one was watching closely: the war-movie mood, the nostalgia tab, the background.
That is the first thing to rule out. This was never a contest your ears entered and lost. Your ears were off.
The second easy story is money, and it is realer, but smaller than it sounds. A track streamed a handful of times earns almost nothing; Spotify once disclosed that tens of millions of its tracks, the ones played between one and a thousand times a year, averaged three cents a month. Three cents. The Velvet Sundown's better-performing songs, by one estimate, might have pulled in a few thousand dollars. The payout, per song, is a rounding error. The threat is not the take. It is the dilution: two million new tracks a month poured into the same finite pipes of recommendation, and the fear is that every real musician's odds of surfacing shrink a little each day. That is what the industry is actually fighting over. It is why one platform says it will stop paying royalties on anything it judges wholly AI made, why the labels rolled out "AI-Generated" and "AI-Assisted" tags days ago, why detection has become an arms race. None of it is about your ears. All of it is about the pipes.
Which leaves the tempting third story: that the machine simply makes junk, and quality will save us. This one is the most comforting and the least true. Music has a long, humbling record of calling a tool fake right up until someone makes it sing. A drum machine that was built to imitate a real kit, failed at it, and got dumped onto the second-hand market went on to become the low end of whole genres. A gadget sold to correct bad pitch became a sound singers chased on purpose, precisely because no human throat could do it. A few years ago, restoration software peeled a dead man's voice off a muddy home cassette, clean enough at last that his two surviving bandmates could play a final song around him; one of them said it was as close as they would ever come to having him back in the room. There is a singer performing today who is nothing but voice software, with a hundred thousand songs written for her by human fans and sold-out concerts where she appears as a projection. Fiction is not fraud. Absence is not fraud. A tool is not fraud. If your whole objection is that a machine touched it, you will lose that argument, and you should.
So strip away detection, strip away the pennies, strip away quality. What is actually left to fight over?
Attention. Just attention. And underneath this whole rooftop stunt is a claim nobody wanted to say out loud: a great deal of streaming now runs on the absence of it.
The synthetic band works best where it is filed, not chosen. It thrives in exactly the places where a listener has agreed, in advance, not to listen closely: the focus playlist, the dinner playlist, the mood you press play on so you can stop choosing. Those are slots a listener has handed to the platform to manage. A synthetic track does not need to be loved to fill one. It only needs to be plausible enough to hold the slot. The machine never needed to beat a musician at being heard. It only needed the slots where hearing had already been switched off.
Here is the uncomfortable version, because plenty of smart people will fight it. Authenticity, for a lot of listening, is not something we spend attention on. It is something we profess and then delegate. We say we care whether a person made it. We want a label to certify it, a platform to police it, a lawyer to defend it, so that we never have to do the one thing that would settle the matter, which is listen closely enough to care. We outsourced authenticity to metadata, and the metadata is exactly the ground the machine already holds.
The copy has hands now. The original never did.
The tribute band, whether the writer would put it this way or not, is a bet against that arrangement. Re-recording a fake band with real hands is a ridiculous way to make money and a hopeless way to win a stream count. But it is a very direct way to force a moment of attention: to make you look up from the background and notice that a body is a thing that can be tired, and sued, and proud, and afraid on a roof. The whole gambit only works if people still glance up.
More than a quarter of a century ago, another tribute band climbed onto a roof on Savile Row. It was the exact roof where the band they impersonated had played an unannounced goodbye to the street below, thirty years before them. That older tribute meant something because the bodies it copied had genuinely once been up there, playing, and the roof still half remembered them. This new rooftop is the same picture run backward. The copy has hands now. The original never did.
The band said its theory out loud in an early post: what is the one thing AI music can't do? Play live. Until now. Somewhere else it promised, in a caption that misspelled "musicians," to steal the music back and give the profits to human ones. Nobody fixed the spelling. They mean the taunt as a taunt. Whether it is also a joke, a business plan, or a small test of whether anyone will still look up, they cannot say, and neither can you. So they stand on the roof and aim a threat at something that cannot feel one. The machine is not afraid of anything. It already has the only thing it ever needed, and the thing took nothing to get: the room you walked out of when you stopped listening.