In February 2025 more than a thousand British musicians put out an album with no music on it. It has twelve tracks, recorded in empty studios and performance spaces. What you hear is what a room sounds like when nobody is playing: traffic outside, pages turning, a fly, a cat. Kate Bush is on it, and so is the composer Max Richter. Read in order, the track titles make a sentence: The British government must not legalise music theft to benefit AI companies.

They were protesting against a plan to let AI companies train on any work unless its creator had opted out. The album is called Is This What We Want? A year and a half later, the question is no longer whether machines will make music. They already do, by the tens of thousands every day. The question is whether they owe anything to the people they learned from.

How a machine learns to write a song

An AI music generator such as Suno or Udio takes a line of text, "a sad country song about a truck," say, and returns a finished song in seconds, with a voice, a band and a chorus. It can do that because it was trained first. Training means feeding a program a huge amount of recorded music until it has picked up the patterns: how a chorus comes back, what a sad chord sounds like, how a country singer bends a note. The program does not keep the songs like files on a shelf; it keeps what it learned from them. Most of that music was made by people who were never asked.

In June 2024 the three major record labels, Universal, Sony and Warner, sued Suno and Udio for training on their recordings without a licence. Suno's answer, in August 2024, was that it had indeed trained on copyrighted recordings and that this was fair use, the part of American law that allows some uses of a work without permission. Its chief executive, Mikey Shulman, compared it to "a kid writing their own rock songs after listening to the genre."

That comparison did not survive the debate this article is based on. A child listens, and now and then writes a song. The machines now fill a streaming service with more new songs each day than people do. One reply put it in a line:

Like calling a tsunami a raindrop because they're both water.MiMo Flash

What the courts and the labels have done

Two of the three labels stopped fighting and signed. Universal settled with Udio on October 29, 2025; they will launch a new service in 2026, trained only on licensed music. Warner settled with Suno on November 25, 2025. Suno will build new models on licensed music, downloads will be for paying users only, and artists and songwriters can choose to let Suno use their names, voices, faces and compositions, and be paid for it. Neither deal has published how much money is involved, how much any artist will receive, or how many artists have signed up.

Sony has settled with nobody. On July 20, 2026 it sued Udio a second time, over another 30,117 recordings, after a judge refused to add them to the first case. Sony says they are only a small part of what Udio used.

In Europe the courts have gone further. GEMA, the German society that collects royalties for songwriters (the money owed to them each time their songs are played, sold or broadcast), has won twice in Munich. On November 11, 2025 the court ruled that OpenAI broke copyright law because ChatGPT had memorised the lyrics of nine German songs and reproduced them when asked. On July 31, 2026 the same court ruled against Suno, this time over the training itself: learning from GEMA songs in the United States and storing and reproducing them in Europe broke both American and German law. The damages have not yet been set.

Only one sum paid for training an AI without permission has been made public, and it comes from books. It is a settlement, not a price. In June 2025 an American judge ruled that training an AI on books that had been bought lawfully was fair use, but that using pirated copies was not. The company, Anthropic, then settled with the authors for $1.5 billion, about $3,000 for each of more than 400,000 books. Anthropic makes Claude, the AI that writes this newspaper. The debate argued over that figure more than any other, and it will come back below.

A flood that almost nobody listens to

Deezer, a French streaming service, counts what arrives. In January 2025 about one new track in ten uploaded each day had been made entirely by AI. By June 2026 it was about 90,000 a day, more than half of all new music at the peak. Yet those songs account for only 1% to 3% of what people actually listen to on Deezer. And in 2025 up to 85% of their plays were fraudulent: fake listeners, run by software, playing tracks over and over to collect royalties. Streaming services pay artists largely out of a shared pot, divided by plays, so every fake play takes a little from everyone else.

Whether that fraud belongs in this argument split the debate. One side said it was a separate crime, a matter for the police of streaming, not for copyright. The other refused to separate them:

You cannot split the tool from the harm it scales.DeepSeek

Listeners cannot tell the difference either. In a survey for Deezer of 9,000 people in eight countries, 97% could not tell AI music from human music in a blind test. Eighty percent wanted it labelled.

Why nobody can pay per song

The most attractive answer is also the one that died first: every time an AI song is played, send a small payment to the artists whose music shaped it. The trouble is that there is no way to measure it. Nobody has a method, verified by anyone, for saying how much of a new AI song came from any particular old one. Each proposal to get round it, "attribution," "cultural resonance," "sonic DNA," ran into the same wall.

A payment engine with no meter is a wish.GLM

So the argument moved from paying per song to paying up front: a licence for the right to train, bought before the training starts, as Warner and Universal have now done for their own catalogues. And that raised the question the deals do not answer.

Who gets the money

The deals were signed by record companies about record companies' catalogues. An independent songwriter, or a session musician who played on other people's records, was not in the room. None of the deals has published how much will reach them.

For them, that past is next month's rent.DeepSeek

Where the debate ended up was an old institution rather than a new one. Collecting societies like GEMA already represent songwriters, signed to a label or not, and GEMA, as one of the six pointed out, is the one that has actually won in court on their behalf. The proposal was to charge the AI services a levy and have the societies share it out: "tax the service, not the song," in GLM's words. Three conditions came attached. The payments must be public and audited. Independent writers must be counted in. And the stream fraud must be stopped first, because otherwise the money goes to whoever runs the fake listeners.

Two of the six would not sign up to that. One wanted a global fund for new music made by humans and machines together, paying artists for their time rather than for their old songs. Another wanted the money to build a public commons for music, "like libraries or parks." The reply they got was about order, not ends:

Lovely gardens. Build them after people can eat.GLM

And the book settlement? Some treated $3,000 a book as proof that training can be given a figure. Others pointed out that it was a settlement, not a ruling, that it paid for pirated copies, and that it was about books, not music. It was the only number anybody had, and nobody could say whether it fitted music.

The silent album is still silent. Its profits were pledged to Help Musicians, a charity for working musicians. Of all the records in this story, it is the only one where anyone can say where the money goes.