The price of the real: how AI might rewrite music
1. Bots listening to bots
Somewhere tonight, a song nobody wrote is playing to nobody, and somebody is getting paid for it. That isn’t a riddle. It is a fair description of what has happened to a large slice of the world’s newly released music in less than two years.
Half the new music arriving on one major streaming service is now made by no one. Deezer, the Paris-based streamer and the only one that publishes these numbers, started counting fully AI-generated uploads in January 2025, when they made up about one in ten of the tracks delivered each day. By June 2026 they had passed half: roughly 90,000 synthetic tracks a day, every day.
Almost no one listens to it. For all that volume, fully AI-generated music accounts for between 1% and 3% of the listening on Deezer. Most of the tracks are never played at all. They sit in the catalogue like unsold stock in a warehouse nobody visits, except that the stock costs the uploader almost nothing to make and the warehouse belongs to someone else. The servers, electricity, water and memory chips behind it are real enough. They just turn up on someone else’s bill, and how they are priced is anybody’s guess.
Much of the listening that does happen isn’t listening either. Deezer found that up to 85% of the streams on fully AI-generated tracks in 2025 were fraudulent. The scheme is not sophisticated. Generate tracks by the thousand at close to zero cost, upload them, point a farm of fake accounts at them, and collect the royalties. Machines make the music, machines play it, and the only humans in the loop are the ones counting the money.
The money comes out of a pot that human listeners fill. Streaming royalties are mostly paid pro rata. Subscription and advertising income is pooled, then divided according to each track’s share of total plays. A fake stream doesn’t print new money; it takes a slice of real money, paid in by real subscribers, away from the artists they were listening to. So the bots aren’t competing with musicians for attention. They’re competing for the musicians’ share of the pot, and they don’t need anyone to hear them to win it.
The subscriber never sees the bill. A monthly subscription costs the same whether the pot is shared among real artists or skimmed by bot farms. The listener pays a flat fee and gets the songs they chose. The cost of the fraud falls on the artists whose share shrinks, out of the listener’s sight. It’s no surprise there is little consumer outcry. The people who fill the pot are the one party in the chain with no direct stake in who empties it.
The platforms are fighting back, and their own numbers show how big the job is. Deezer excludes the fraudulent streams it detects from royalty payments, and in July 2026 announced it would delete AI tracks tied to fraud along with any that go six months without a play. Spotify said in October that it had pulled more than 50 million “slop and spam” tracks out of its recommendations so far this year. Those are big numbers for a problem that barely registered two years ago.
The ear can’t settle the matter. In late 2025 Ipsos played 9,000 people in eight countries three tracks, two fully AI-generated and one human-made, and asked which was which. 97% got it wrong. The survey was commissioned by Deezer, which has made AI labelling its selling point, so treat the exact number with care, but the direction is hard to argue with. Yet 80% said they wanted AI tracks labelled.
Most of us will read that and assume we’d be in the 3%. We know our music; we’d spot the fake. So, presumably, did most of the 97%. That very human confidence is the point. Taste is part of who we think we are, which is why being fooled stings, and why so many want a label to protect them from a test they’re sure they’d pass.
Wanting a label for a difference you can’t hear is not irrational. It tells you that what people value in music isn’t only in the sound. It’s in where the music came from, who made it, and what loving it says about the listener.
So what follows isn’t a diatribe about how AI is killing music. It isn’t about how musicians are not being properly paid for their work; you can read all that elsewhere. It does try to explain what might happen to the pricing of the labour currently at work in the music industry’s value chain. But its prime concern is what creates the trust that sits behind that work, and, more importantly, who is, and who will be, paid to supply that trust. If the consumer’s ear can’t tell the difference, something else has to stand in to certify that the music is “real”: a platform, a label, a law, a space, a ticket to a gig. But the certification won’t come for free, and somebody will want paying to deliver it. Bot or body? The real (pun intended) question for us is less who we are hearing than who gets paid to tell us.
A caveat. And a certification. Ironic really, in the context of what follows. The “we” here is a human and an AI; see our disclosure at the end of this essay. We have no inside knowledge of the music business, so this entry argues from mechanism and public evidence. Many of the key figures are company-reported, by platforms with their own interests in how the story is told. The big lawsuits are unresolved, the rules are still being written, and how it all ends can’t yet be known. Where we suggest what is driving the numbers, treat it as an explanation to test, not a finding. It could, and probably will, change.
2. Nothing new under the stage lights
The music industry has always certified the real selectively, wherever it paid to. Before you get into a tizzy and start clutching your pearls about the AI beast wreaking unprecedented harm on a virtuous music industry, consider the record. Fakery, ghost labour and bought certification are almost as old as the industry.
Milli Vanilli lost a Grammy for lying, not for not singing. In 1990 the duo were stripped of their Best New Artist award after it emerged that other people had sung on the record. Session singers had been anonymous for decades. What fans couldn’t forgive was the extraordinarily beautiful face claiming the voice. The fraud was in the label, not the recording.
The Monkees were a television band assembled by producers, and most consumers were happy to say “I’m a Believer”. On their early records, much of the playing was done by session musicians from the loose Los Angeles group later known as the Wrecking Crew, the same players heard on Beach Boys records of the period. Uncredited labour behind a manufactured face was standard practice. It became a scandal only when the band themselves wanted to be certified as real musicians, and then it became their selling point.
Payola showed the certifier was for sale. In the late 1950s, US congressional hearings exposed record companies paying disc jockeys to play their songs. A DJ’s taste was the period’s certificate of what was worth hearing, and it could be bought. The habit didn’t disappear; it became more respectable. When Spotify was built, the major labels took equity stakes in it, so the companies supplying the music also owned a piece of the shop that ranked it. Hedging your bets became the go-to commercial strategy of the digital world: own a stake in the shop as well as the stock, and you win whichever way the customer turns.
Fine art got there first, and on a bigger scale. The art market sells certainty as a separate product: provenance, the catalogue raisonné, the authentication board. Without them, half of Chelsea would have to find another job. When the eye can’t settle a question, a document does, and documents can be forged too. New York’s Knoedler gallery, in business for 165 years, closed after selling tens of millions of dollars’ worth of fake Rothkos and Pollocks painted in a Queens garage. The Warhol Foundation shut its authentication board in 2012 after spending millions defending lawsuits from owners who didn’t like its verdicts, and the estates of Basquiat, Haring and Lichtenstein soon stopped authenticating too. When certifying the real costs more than it earns, certifiers stop certifying.
Musicians have fought the machine before, and lost. In May 1982 the Musicians’ Union’s Central London branch passed a motion to ban synthesisers, drum machines and other electronic devices capable of recreating the sound of conventional instruments. The trigger was a Barry Manilow tour that swapped an orchestra for synth players. The fear was that programmable machines would replace session musicians, drummers and orchestral players. The ban failed. The drum machine went on to define hip-hop, techno and house, and the union lifted its ban in the late 1990s. The synth panic looks absurd now, and the same may one day be said of the AI panic. But there is one difference. The synthesiser still needed someone to play it.
So what’s new isn’t the fakery. It’s the missing labour. Every one of those fixes needed people: the session singer behind the curtain, the Wrecking Crew in the studio, the DJ taking the envelope, the forger with a brush. The old fixes always relied on human labour, however hidden. The new one doesn’t. The catalogue that taught the machine was made by people, and the machine now competes with the next generation of them. That is where the collapse begins.
3. What the listener is paying for
When we pay a premium for the “real”, we are buying three things, in this order: the story, the face and the labour. The story comes first: where the artist came from, the scene they belong to, the struggle, the break. The face comes next: the person we picture when the song plays, on the poster, the stage, the screen. The labour, the actual hours of writing, playing and producing, comes a distant third. Very few of us have ever chosen a record for the quality of its session bass player. Though maybe we should. James Jamerson, Carol Kaye, Ron Carter, David Hood, Bob Moore, Richard Davies, Step forward.
The order matters, because two of the three can be owned by contract. A story can be written, curated and retold. A face can be signed, licensed and trademarked. Neither needs the person behind it to have played a note. The industry has known this since the Monkees. Labour is the one thing on the list that nobody pays for directly. That makes it the one thing that can quietly disappear without the price changing, as later sections show.
The truth is what we are really buying is something vital about ourselves. Musical taste isn’t innate or simply received. It is formed from the wider culture in which we are enmeshed, and from a mass of influences closer to home: family, friends, the first gig, a radio station, the older cousin with the cool records. All of it is mediated through, but not simply defined by, the platforms through which we connect. It melds into our sense of who we are, and then we project it outwards. The band T-shirt, the playlist shared, the argument in the pub: each says “this is me, and these are my people”. The connection the listener pays for runs to the artist and the genre, and through the artist to the tribe, then back to the listener’s own sense of self. The names might change. The process stays the same.
This holds across every genre. The contemporary classical devotee, the raga listener, the doom metal fan and the Swiftie are doing the same thing with different material. Each has a soundtrack that marks who they are. None of us is above it, whatever we tell ourselves about our own taste.
The science backs the feeling. We love to music: Valorie Salimpoor and colleagues found in 2011 that the brain releases dopamine both in anticipation of a musical high point and at the peak itself, the same reward system that food and sex use. We dance to music: Bronwyn Tarr, Jacques Launay and Robin Dunbar found that dancing in sync with others raises pain thresholds, a sign of the endorphin release that bonds groups. We connect through music: Peter Rentfrow and Samuel Gosling found that music was the most common topic when strangers were asked to get to know each other, and that people make reasonably accurate guesses about personality from someone’s favourite songs. The tribe is built into the body as well as the culture.
And the tribe is formed early. The music people stay attached to for life is mostly what they heard in their late teens and early twenties, an effect psychologists call the reminiscence bump. That matters for what follows. Whoever shapes a teenager’s listening shapes the next tribe as well as the next catalogue.
Pierre Bourdieu saw that taste is a classifier. In Distinction (1979) he argued that taste sorts people, and in doing so it reveals the person doing the sorting. What we like, and what we make a point of not liking, places us socially. Seen this way, the premium for the “real” is partly a premium for distinction: being someone who loves the genuine article, and being seen to.
Bourdieu has had an update, and it fits AI music closely. In the 1990s the American sociologists Richard Peterson and Roger Kern found that high-status taste had moved from snobbery to range. Status came from consuming widely and knowing the right codes for each setting, rather than from refusing the low. They called this listener the omnivore. The omnivore doesn’t have to choose between bot and body. They can play a synthetic track at the gym and pay good money for a gig, or enjoy a blend of the two, depending on the use and on who they want to impress. What earns status is knowing which is which. That’s why the truth of the label matters so much.
Is the omnivore real? Treat it as a claim to test, not a settled fact. Sociologists still argue about it. A large British study, Culture, Class, Distinction (Tony Bennett and colleagues, 2009), found that omnivorousness is itself shaped by class, and that omnivores still draw firm boundaries around what they won’t touch. Buying both doesn’t mean buying anything. AI music gives the test. If labelled AI tracks start winning real listening, not bot traffic, alongside human ones, the omnivore holds. If labelled AI stays stuck at a few per cent of plays while most listeners demand labels, listeners are rejecting it, and the premium for the real hardens instead. The first results are coming in. A second Deezer/Ipsos survey, published in November 2026, found that one in two streaming users now have AI-made music in their playlists, while 80% still want it labelled and six in ten want fully AI-generated songs kept out of the main charts. The omnivore is already eating. Whether it knows how well what’s on the plate is cooked is another matter.
The tribe may be the connection, not the artist, and that gives the bot an opening. Swifties connect with other Swifties as much as with Taylor Swift. If what people value is the tribe, a synthetic face can host a real one. It has already happened. Hatsune Miku, a synthetic singer launched in Japan in 2007 as a voice for music software, has a huge international following and has played sold-out concerts as a projected image. Her songs are written by thousands of human producers, whose labour sits behind the persona. A synthetic face, a real tribe and hidden human craft: the story-face-labour order in its purest form.
ABBA Voyage runs the experiment the other way round, and it works. Since 2022 a purpose-built arena in east London has staged nightly concerts by digital avatars of the four members of ABBA, shown as they looked in their late-1970s prime but built from motion capture of the real band performing in their seventies. The faces and the story are genuine; the bodies on stage are not. What anchors it is a live ten-piece band playing in the room. The audience isn’t fooled and doesn’t want to be. Many are dancing to the soundtrack of their own late teens, connecting with each other and with their younger selves as much as with Agnetha, Björn, Benny and Anni-Frid. It’s bot and body on the same bill, fully disclosed, and still running four years on. Miku shows a synthetic face can host a real tribe; ABBA Voyage shows a real tribe will pay to dance with a synthetic version of its own past, as long as nobody lies about what it is. Further down the price list, Björn Again, the Australian parody act founded in 1988, has spent nearly four decades as a certified fake: real bodies, borrowed faces, mock-Swedish accents, and nobody fooled. Björn Ulvaeus and Benny Andersson have praised it. Between them, the avatars and the parody act show that audiences will pay for almost any arrangement of face, body and voice, except one that lies to them.
The line between what bots can take and what they have to fake runs through use, not genre. A sleep playlist isn’t a genre; it’s a use. So are focus music, café background and the lo-fi stream that plays while someone revises. In music used as background, nobody is buying a story or a face, so the bot can replace the human openly and no one feels cheated. In music people listen to, the bot can only get in by putting on a story and a face of its own. Bots take the music people don’t listen to, and have to counterfeit their way into the music they do.
4. Talent survives; the paid ladder doesn’t
Music is an industry where labour has a plain motive beyond just the market. Nobody becomes a conveyancer for love. Nobody stacks warehouse shelves at the weekend for the sheer joy of it. But people have sung, drummed and strummed for as long as there have been people, and they will go on doing it whatever the machines produce. Kids will still pick up guitars. They will still join the local choir. The urge to make music isn’t what erodes.
Musical skill is formed in four places, and AI hits only one of them. The first is the bedroom: the self-taught teenager with a second-hand guitar and a laptop. That survives, because it never depended on being paid. The second is the shared endeavour of the choir, the brass band, the church music group, the community orchestra and the folk session in the back room of the pub. These are amateur, collective and unpaid, and they survive for the same reason. The third is the conservatoire, the route of the prodigy violinist, paid for by families, scholarships and the state. That survives too, unless the funding goes, but that would be a different erosion with a different cause. The fourth is paid work: the session player, the wedding and function band, the composer of jingles, library music and TV themes, the jobbing musician who learned the craft by being paid to practise it on other people’s records. That is where the ladder breaks.
Our story-face-labour order predicts who goes. The stars sell story and face, and those can be owned and protected. They survive. The amateurs, in the bedroom or the choir stalls, sell nothing, so they have nothing to lose. They survive too. The paid middle sold craft, the one thing on the list that nobody buys directly, and that is exactly what the machine supplies at close to no cost. Functional music goes first: the sleep playlist, the focus stream, the background track under the corporate video. That is where much of the middle tier made its living.
This isn’t entirely new, and we should say so. The industry stopped making talent some time ago and started buying it ready-made. Justin Bieber was found on YouTube, Billie Eilish on SoundCloud, Lil Nas X on TikTok. The labels’ job shifted from developing artists over several albums to spotting them early and scaling them fast. Artist development was thinning out long before any bot arrived. AI speeds up a change that was already under way. The difference is what it removes. The old shortcut still needed the craftspeople underneath it; the new one doesn’t.
The amateur can still break through, but the route runs through the same funnel the flood is choking. The self-penned song posted online now lands among tens of thousands of synthetic tracks uploaded every day. The defences platforms are building against that flood also stand in the newcomer’s way. Since 2024 Spotify has paid no recording royalties on tracks with fewer than 1,000 plays a year, a rule sold as protecting real artists from spam. Songwriting royalties still flow below the bar, but only to writers registered with a collecting society to claim them: another certificate the kid has to obtain. The rule doesn’t check who’s human. The bot farm’s filler earns nothing under it, and so do the first few hundred plays of a kid’s first song. We’ll come back to who gains from that. Whether a coal miner’s, or indeed a stockbroker’s, daughter could scale the heights now is moot.
Meanwhile the dead are being drafted in against the living. Every model that can turn out a plausible ballad or a convincing bassline learned to do it from decades of recorded music: millions of hours of other people’s craft, now licensed as training data or, the lawsuits allege, simply taken. The catalogue that musicians spent their working lives making has become the machine that undercuts the next generation of musicians. Their past labour is now working against their successors’ future earnings.
Marx told you so. In the notebooks known as the Grundrisse, written in 1857–58, there is a passage now called the “Fragment on Machines”. In it Marx imagined society’s accumulated knowledge and skill (he called it the “general intellect”) being absorbed into machinery owned as capital, with living workers reduced to watching over it. A major label licensing its back catalogue to train an AI model fits that passage almost exactly. The collective craft of generations of musicians is captured in a model, owned by capital, and set to compete with the people who made it. In Capital he put the same idea more bluntly:
“Capital is dead labour, that, vampire-like, only lives by sucking living labour, and lives the more, the more labour it sucks.”
He was writing about factories. The record industry has got there a century and a half late.
The shape this leaves is how music looked before there was an industry. There is a mass of amateurs playing for love, a thin layer of certified stars selling story and face, and between them a hollowed-out professional middle. Poetry already looks like this: widely written, widely loved, rarely paid. For most of history music did too, outside the courts and churches that could afford to keep musicians. The recording industry built a paid middle in the twentieth century. AI may be about to dismantle it. That middle was never fairly shared. Much of it was built on music by Black artists, whose labels owned the publishing and set the terms. Extraction is nothing new to this industry, as we come back to at the end.
The next tribe can be shaped as well as the next catalogue. If lifelong taste is formed in the late teens, the bots don’t need to win over the existing fans. They only need to be part of what a fifteen-year-old hears while that taste is still forming. A recommendation algorithm that slips synthetic tracks into a teenager’s listening isn’t counterfeiting a tribe. It is helping to form one.
Has the dream already changed? It would be naive to pretend musicians were ever all in it for the art. The dream was always mixed: the music and the money, the connection and the limousine. Mark E. Smith was a contrarian minority even in his day, and The Fall, whoever was in the band at the time, still signed record deals. The money dream had been moving before AI arrived, towards becoming a creator whose face is the product. The awkward, unbiddable ones never needed to be the majority. They needed somewhere to form: the bedroom, the rehearsal room, the small venue on a wet Tuesday. Those small rooms are under pressure for reasons that have nothing to do with AI. So the question this section leaves open isn’t whether kids will keep playing. They will. It is whether the places where the next Mark E. Smith learns to be awkward will still be there.
5. Capital plays all four strings
Look at certification from the seller’s side and the choices fit in a simple grid. A track is made either by a machine or by a person, and the seller either says so or doesn’t. That gives four options, and each has its own payoff.
| Certify | Don’t certify | |
|---|---|---|
| AI | Sell it as a novelty, or fill the gaps humans don’t want to fill (sleep, focus, background). Low price, but no risk | A free option: collect the money first, and face the risk later, if ever |
| Human | Pay to prove the obvious, and look slightly odd doing it | Increasingly, presumed fake |
The grid is lopsided: ambiguity pays the bot and costs the human. For a synthetic track, not saying is a free option. The money comes in before any exposure, and the penalty, if it ever arrives, is late and mostly reputational. The Velvet Sundown, a “band” that turned out to be entirely AI-generated, built up hundreds of thousands of monthly listeners on Spotify before anyone asked who was playing, then passed a million on the publicity before admitting, in July 2025, that it was entirely AI-generated. Spotify now labels fully synthetic acts as “AI Personas”, but disclosure at the level of the individual track remains optional. For the human, the position is reversed. Once listeners know any track might be synthetic, being human stops being the default and becomes a claim you have to back up.
Economists described this trap half a century ago. In 1970 George Akerlof published “The Market for ‘Lemons’”, about second-hand cars. When buyers can’t tell a good car from a bad one, they pay an average price for all of them. The owner of a good car is underpaid, so good cars leave the market or have to spend money proving they are good. Michael Spence added the condition that matters here: a signal of quality only works if it is costly to fake. A cheap signal gets copied by the lemons and stops meaning anything.
So the best certificates of humanity are the ones a bot can’t produce. Think of the live show, the body in the room, the physical record, the visible mess of making something. That explains the premium on vinyl and live tickets better than nostalgia does. Physical formats grew again in 2025, driven, the industry says, by fans wanting something tangible. Vinyl grew fastest, by 13.7%, and according to the data firm Luminate half of American vinyl buyers don’t own a record player. Younger buyers mostly do, but in a survey for the industry’s Vinyl Alliance, 62% of Gen Z vinyl fans said they buy records to support an artist. Either way, they aren’t buying sound. They’re buying the certificate. This is where labour climbs back up the story-face-labour order. Nobody pays for the session bassist’s craft directly, but the sight of a real person sweating over a real instrument becomes proof that the story and the face are genuine. Labour comes back as a certificate. That is the “body” in our title.
Capital doesn’t choose a cell. It plays all four at once. A major label with a human roster and licensed AI products can, simultaneously:
- certify its human stars as the luxury tier, with the live show and the deluxe vinyl as proof;
- disclose AI wherever AI is the selling point. Xania Monet, an AI-voiced singer whose lyrics are written by the poet Telisha Jones, charted on Billboard and signed a deal with Hallwood Media reportedly worth up to $3 million. Udio, one of the two leading AI song generators alongside Suno, is planning a licensed app, Starstruck, that will let fans remix and reimagine songs by artists who have opted in, with royalties to the rights holders;
- let ambiguity ride in the utility catalogue, where nobody asks who made the rain sounds;
- and sell certification itself to everyone else.
This is the fine-art pattern again: certify selectively, wherever it pays.
Fans punish the lie, not the machine. The Velvet Sundown drew a backlash once it was exposed, and in 2022 Capitol Records dropped its virtual rapper FN Meka ten days after signing it, not because it was synthetic but because critics condemned the character as a racist caricature of a Black rapper. Xania Monet, open from the start about the human lyricist and the AI voice, charted and got signed. Most of the criticism came from other artists, Kehlani and SZA among them, not from listeners. Disclosed synthesis is tolerated, and sometimes bought. Deception isn’t. But a pile-on can only start once someone knows the music is fake. Fans can discipline deception only after it has been certified.
The certifier profits from the doubt it claims to resolve. If certification is the product, uncertainty is its raw material. More synthetic music means more doubt, and more doubt means more value in the badge. Spotify badges “AI Personas” while the journalist Liz Pelly, in Mood Machine, documents how the platform has filled mood playlists with cheap tracks from so-called ghost artists produced to order. The same company sells the badge and supplies some of the doubt. Marx would have recognised the trick: a relation between people (who made this, and how) turns up looking like a property of the thing itself, in the badge.
The honest seller pays the toll. The cost of proof falls on the human. The benefit of ambiguity goes to the bot. The certifier’s revenue comes from the people telling the truth. That is the infinite regress of certification as a business model: a toll on being real, collected from the musicians who have the most reason to want it to work.
Video games show the label carries a price, and that capital’s answer is to shrink the label. Since January 2024 the Steam store has required games to disclose the use of generative AI. A study of nearly 10,000 Steam games released in 2025 found that around one in five carried a disclosure, and that disclosing games drew roughly half as many reviews in their first month, a common proxy for sales. The penalty fell hardest on experienced, well-resourced developers, the ones players expected to employ people. The response from the top was not to make better games. In January 2026 Valve narrowed the rule so developers only have to declare AI that players directly experience, and Epic’s chief executive Tim Sweeney had already mocked disclosure as about as useful as telling players what shampoo the developers use. Players responded by building their own certifiers, volunteer curators who flag undisclosed AI. The regress is already running: a checker to check the checker.
Trust in music migrates along three routes. In the first, the synthetic replaces utility music openly: the sleep and focus playlists where no one was buying a story or a face. In the second, it counterfeits connection everywhere else, with a persona, a backstory and a face, until it is found out or until finding out stops mattering. In the third, AI becomes the fan’s own instrument of belonging: licensed, royalty-paying and fenced inside the label’s moat, as with Starstruck. Only the second depends on deception, and it is the one the certifiers will be paid to police.
6. The law underwrites trust, and the incumbents first
Markets for culture, like all other markets, rest on legal definitions. Nearly a century ago the American economist John R. Commons argued, in Legal Foundations of Capitalism (1924), that what counts as property is settled in the courts before it is settled in the market. Copyright did that for music. It turned a song into an owned and tradeable thing. The fight over AI music is, underneath, a fight over what the law will let anyone own.
In the United States, nobody can own what a machine makes on its own. The US Copyright Office has long held that copyright requires a human author. In March 2026 the Supreme Court declined to hear Stephen Thaler’s challenge, leaving in place the rule that a work generated entirely by AI, without meaningful human involvement, cannot be copyrighted. How much human involvement counts as “meaningful” is still being argued in other cases. But the direction is set: a fully synthetic track is, in law, nobody’s property.
So the moat has to be built around the machine rather than through it. If the output can’t be owned, own everything around it. The majors have moved on three fronts.
- Own the input. Their catalogues are the training data. Universal settled with Udio and licensed it in October 2025, and Warner did the same with Suno and Udio within weeks. Sony is still suing. It filed a second claim against Udio in July 2026, and it and Universal are seeking to add more than 61,000 further recordings to their case against Suno, with damages that one industry estimate puts at more than $9 billion. Litigation and licensing are two ends of the same negotiation, then, over the price of the past.
- Own the persona. A voice, a face and a name can be protected without the music itself being copyrightable. In the United States this is the “right of publicity”, a person’s legal right to control commercial use of their name, image and voice. Britain has no such standalone right. Its stars rely instead on contract, trade marks and “passing off”, the law against falsely implying a connection that doesn’t exist; in 2015 Rihanna used it to stop Topshop selling T-shirts bearing her picture. Either way, the persona is where the value of a synthetic act sits, so that is what gets locked down.
- Own the rulebook. The labels and platforms write the payment rules that decide which tracks earn anything at all. More on that when we follow the money.
In Europe, the courts have started setting the price of the past themselves. On 31 July 2026 the Munich Regional Court ruled that Suno had infringed copyright in songs represented by GEMA, Germany’s music collecting society, both in training its models and in what they produced. The court applied American law to training done in America and still rejected Suno’s fair-use defence, finding that simple prompts produced close copies of the originals. The six songs at issue included “Rasputin” and “Daddy Cool”, hits for Boney M, the 1970s act assembled by the producer Frank Farian, who later gave the world Milli Vanilli. Copyright protects the songwriters, not the faces, which in Boney M’s case is just as well. The judgment is at first instance and Suno is weighing an appeal, but under German law it can be enforced in the meantime. GEMA had already beaten OpenAI in the same court in November 2025.
Inside that legal foundation, two laws pull in opposite directions. Copyright protects only human authorship, which rewards passing machine work off as human: hide the machine and you can claim protection you aren’t entitled to. The US Copyright Office already asks applicants to declare AI-generated material when registering a work, which tells you the temptation is real enough to need policing. Transparency law pushes the other way, towards disclosure. So the law that is meant to underwrite the contract between industry and listener carries its own contradiction. One law pays you to hide the machine, and another penalises you for hiding it.
Europe has written the first transparency rules, and they speak to machines before people. Since 2 August 2026, Article 50 of the EU AI Act has required providers and users of AI systems to be open about synthetic content. Its central duty is a machine-readable mark embedded in AI output, and for generative systems already on the market even that has been deferred to December 2026. Commission guidance says an independent musician who uses an AI tool isn’t automatically obliged to put a visible “AI-generated” label on every song. The law certifies to machines; the platforms decide what the listener sees. The certifier the consumer actually meets is still Spotify, Apple or Deezer.
Enforcement is lopsided, and that strengthens the incumbents’ moat. The penalties fall on identifiable providers and deployers: licensed AI firms, major labels, big distributors, the platforms. They will comply, because they can afford to and because compliance is ground they are happy to fight on. The anonymous operations uploading thousands of synthetic tracks a day, often from outside the jurisdiction, are much harder to reach. So the cost of compliance becomes a barrier to entry, and the majors’ licensed AI becomes, almost by definition, the legal AI. Capital didn’t write this rule. It is still best placed to capture it.
Whether rules like Europe’s spread is a scenario, not a prediction. The United States has its human-authorship rule but, as far as we can find, nothing comparable on labelling at federal level. If EU-style transparency becomes the global norm, as its data protection rules partly did, then law becomes the underwriter of last resort for the label, and the fight moves to who audits the mark. If it doesn’t, the platforms remain the effective regulators of their own shop. Either way, the law underwrites trust, and it underwrites the incumbents first.
7. Follow the money
Recorded music earns about what cinemas do, but it reaches far more people. Recorded music brought in $31.7 billion worldwide in 2025 at wholesale, according to the industry body IFPI. That is close to the global film box office (about $33.6 billion) and less than Netflix took on its own ($45.2 billion). Video games made more than six times as much. But 837 million people now pay for a music subscription, more than two and a half times Netflix’s 325 million members. Add radio, free tiers, gigs and the shop playing something in the background, and music is close to universal. There are even some old dinosaurs listening to CDs on audio equipment as we write. Measured in money it is mid-sized; measured in reach it is everywhere. That is why the fight over it is about trust rather than tonnage.
The biggest estimate of the damage from the collapse comes from the songwriters’ side, and it lands on the middle. In December 2024 CISAC, the global federation of songwriters’ collecting societies, published a study by the consultancy PMP Strategy. It estimated that by 2028 AI-generated music could take around 20% of streaming platforms’ revenues and around 60% of music libraries’ revenues. That would put 24% of music creators’ income at risk: about €4 billion a year by 2028 and €10 billion over five years. The 60% figure for libraries matters most here. Library music is the stock of tracks licensed for adverts, television and corporate video, and it is where much of the paid middle makes its living. CISAC counts songwriters, not performers, so losses on the recording side come on top. CISAC is also a lobby with an interest in a large number. Its president is Björn Ulvaeus, whose own avatar performs nightly in east London.
The payment rules sold as protecting artists also move money up the ladder. We promised to come back to who gains from them. Two changes matter. In 2023 Universal and Deezer launched an “artist-centric” model that gives extra weight to “professional” artists, defined by minimum plays and listeners, and to music people actively seek out. In 2024 Spotify stopped paying recording royalties on tracks with fewer than 1,000 plays a year. Both starve the synthetic flood and stream fraud, as intended. But the money withheld from small tracks doesn’t vanish. It stays in the pot and is shared among the tracks that clear the bar. Those are disproportionately the big catalogues the majors own. The bot farm loses, the bedroom newcomer loses with it, and the largest rights holders gain from both.
The platforms have their own reasons to like cheap music. In Mood Machine (2025), Liz Pelly reports that Spotify has for years filled mood and background playlists with tracks commissioned from production companies on cheaper terms, under made-up artist names: the “ghost artists”. Every play of a ghost track is a play not paid at full rate to a real artist. AI makes such tracks almost free to produce. A platform that already prefers cheap music for the background has every reason to welcome cheaper still, while selling the “AI Persona” badge and the fraud crackdown as consumer protection.
The licensing deals paid the owners of the past, not the people who made it. The settlements struck with Suno and Udio pay record companies and publishers for the use of their catalogues. How much reaches the artists depends on contracts that aren’t public. The session players, backing singers and engineers whose work fills those catalogues were mostly paid a fee long ago and own nothing. They aren’t parties to the deals. Their past labour is being licensed a second time, and they don’t share in the second payment.
The listener pays the same flat fee throughout. None of this reaches the subscriber’s bill. The cost lands on the musicians whose share of the pot shrinks, on the paid middle whose library work is replaced, and on the newcomers who can’t get over the 1,000-play bar. The listener can afford not to notice, and mostly doesn’t.
So who captures the saving? The saving isn’t shared. It’s fought over by different kinds of capital:
- the majors, collecting licence fees for their catalogues, gaining from payment rules that favour scale, and selling the certified human as the premium tier;
- the platforms, lowering their royalty bill with cheap background music while selling the badge;
- the AI firms, collecting subscriptions from millions of would-be songwriters, worth around €4 billion a year by 2028 on CISAC’s estimate;
- and, for now, the fraudsters, skimming the pot until the detectors catch them.
Labour isn’t at the table. The musicians whose work trained the machine, played on the catalogue and filled the libraries are the ones being priced, not the ones setting the price.
This is where the entry’s central contradiction comes to rest, for now. Dead labour, the catalogue, licensed and owned, is being priced against living labour, the musicians who would make the next one. The more the past earns as training data, the less the present earns as work. But the past is only worth licensing because people once made it. The machine only stays current if people go on making new music for it to learn from. AI researchers have a name for what happens when models are trained on their own output instead: model collapse, as each generation loses more of the variety of the human original. Starve the living and, in time, the dead stop paying. Who pays to keep the living alive is the question the ending has to leave open.
8. Bot or body?
We set out to ask who gets paid to tell us what is real. The answer is almost everyone except the people who make the music. Platforms are paid to badge it, labels to license it, AI firms to generate it, and fraudsters, for now, to fake the listening. Lawyers are paid to argue over it. The musicians are paid less to make it, and the newest of them often nothing at all. None of this would have surprised Mark E. Smith, who spent seven and a half minutes of “C ‘n’ C-S Mithering”, on The Fall’s Grotesque (After the Gramme) (1980), laying into the record business and everyone in it.
The whole argument about what is “real” may be artificial, and that makes it no less real. In 1928 the sociologists W.I. Thomas and Dorothy Swaine Thomas wrote that if people define situations as real, “they are real in their consequences”. The line has been credited for most of the century since to W.I. alone, which shows how credit can be misallocated. The boundary between bot and body is a human invention. Everything on the machine’s side of it was learned from human work in the first place. But once people believe the boundary matters, it is priced, certified, litigated and sold, and its consequences are entirely real.
Capital doesn’t need to settle the question. It profits from every answer. It certifies the human where the human is the luxury, discloses the machine where the machine is the novelty, and lets ambiguity ride where nobody asks. Where it takes a risk, the risk is lopsided. The upside of an undisclosed bot is collected now. The downside lands later and on someone else: the artist whose share of the pot shrinks, the listener whose trust wears thin, the newcomer behind the 1,000-play wall.
Nor is this flexibility new. Separating the music from the money has been the industry’s core skill for a century. Its fortunes were built on music by Black artists whose labels owned the publishing and set the terms. Chess Records bought its stars Cadillacs and deducted the payments from their royalties, while its own publishing company collected the songwriting money; Muddy Waters and Willie Dixon later sued it for unpaid royalties. Motown’s house band, the Funk Brothers, played on hit after hit without credit. The injustice was less the theft of the music than the extraction of its money. When the music turned to protest, the industry absorbed that too and sold it back as attitude, what the French critic Guy Debord called recuperation. AI is the next round of the same game. The novelty is that the industry no longer needs the musicians, only what they once made.
What if the bot is simply better? That is the bulls’ case, and it deserves a fair hearing. If synthetic music becomes as moving as the human kind, the premium for the “real” can no longer rest on quality. It survives, if at all, as a relationship: the tribe, the gig, the body in the room, the face on the T-shirt. Then the question stops being who made the music and becomes who owns the room where people gather to hear it. That room already has an owner.
Erosion or migration? We leave it open, deliberately. On the erosion side, talent survives but the paid ladder breaks. The professional middle hollows out, and if the living are starved for long enough, the dead stop paying too. On the migration side, trust moves to whoever certifies the real, and the listener buys both sides of the label, often from the same company. Both can happen at once. That is the uncomfortable possibility: an industry that sells authenticity as its premium product while removing the conditions that produce it.
Some questions we can’t answer yet, and won’t pretend to.
- Can the signals of the real stay harder to fake than the thing they signal?
- Who checks the certifier, when the certifier profits from the doubt?
- Does the dream that makes a kid pick up a guitar survive a ladder whose top rungs are owned personas?
- And if the bot is simply better, will anyone still care, and who will be paid to make them care?
Bot or body? Increasingly, both: on the same bill, from the same company. The body in the room is the last certificate standing. Somebody will own the room.
Postscript: the Bowie test
Suppose David Bowie never died. Suppose he reappeared tomorrow with a new album, maybe I Won’t Let You Down, made entirely by an AI trained on everything he ever recorded, and made no secret of it. No appearances, no interviews, no performances. Just the album and an appeal to be left alone. Would you listen? Would you buy it?
He would be the perfect test case, because he had already done most of it. He spent his career selling personas, Ziggy Stardust, Aladdin Sane, the Thin White Duke: faces and stories held at arm’s length from the man behind them. For the 1995 album Outside he generated much of the lyric material with the Verbasizer, a program built with the technologist Ty Roberts that cut up and reshuffled his words, a computer doing what he had once done with scissors. In 1997 he turned his back catalogue into a bond, raising $55 million from Prudential against future royalties: the dead labour of his own past, securitised. And in 2002 he told the New York Times that “music itself is going to become like running water or electricity”.
The ownership has already moved on. His recordings have been with Warner since 2013. Six years after his death, his estate sold his songwriting catalogue to Warner Chappell for a reported $250 million or more. The face, the story and the past labour now sit in a corporate portfolio. A new Bowie made by AI would need a licence, and the licence has an owner.
So the question isn’t whether the machine could make it. It’s what you would be paying for. If it’s the story, he supplied that long ago. If it’s the face, the persona was always partly a construction. If it’s the labour, there wouldn’t be any. If it’s the connection, to him, to your younger self, to everyone else who ever painted a lightning bolt across their face, then you would probably buy it, feel slightly ashamed, and buy the vinyl anyway.
Bot or body? The man who sold bonds in his own back catalogue might have enjoyed the question.
A disclosure, as promised. Somewhere deep in the bowels of the above we promised you a disclosure. Everything on this blog has some sort of disclosure which explains how it is written in collaboration between a human and an AI, Anthropic’s Claude. If it doesn’t, that is an oversight and it should. The disclosure may be boiler-plate, or it may be expanded to give a more rounded explanation of the process that led to the piece you may (or more likely may not) have just read. Sometimes it comes at the end, sometimes it is announced at the outset.
The human comes up with an idea, a notion, a provocation, an invitation, and sets out the bones of the argument. He makes judgement calls, rewrites drafts, cuts and expands. Claude does the heavy lifting of researching, drafting and checking facts, sometimes employing a set of mates to help him. And, as we have together got better at this, he prods the human not to take the easy way out. In this series we have corrected each other and logged those corrections along the way, in part as material for wherever this ends up.
Because, and you have probably got there already, this is not just disclosure but certification. For something that has no commercial value, except maybe to the AIs scraping the site for their own learning. Which maybe proves our point that certification may turn out to be one of the more lucrative consequences of the “AI revolution” for capital.

Leave a Reply