Harry Styles, Adele, Britney Spears, Queen, Michael Jackson. Those are the artists named in the reporting when the Financial Times disclosed that Sony Music has now asked platforms to remove more than 260,000 tracks imitating performers it represents, as relayed by Music Business Worldwide and The Music Network. The company had filed 135,000 by the end of March and passed 260,000 by the end of September: close to a doubling in six months, better than twenty thousand notices a month. Sony’s own comment on the work was flat. “We will keep enforcing our artists’ rights and expect platforms to act quickly and to prevent the same content from reappearing. But it’s becoming an uphill struggle.”
Neither outlet could say how many of the tracks came down, how fast, or which services received the notices. The FT did not report it.
I want to take the second half of Sony’s sentence seriously, because the quantities around it do not sit the way the sentence suggests. In June, Deezer put its own figures on record: fully AI-generated uploads had passed half of everything arriving on the service each day, around 90,000 tracks daily, and those tracks were drawing between one and three per cent of streams. Three days of new uploads at that one platform outnumber every track Sony has asked anyone, anywhere, to take down since April.
That gap is not a failure of effort and it is not really a problem of scale. It is what a takedown notice is. A notice has to name somebody. It asserts that a particular track imitates a particular performer whose voice a company controls, and it can be filed at all because there is a name on one side of it and a signed agreement standing behind the name. So the enforcement reaches exactly as far as the names reach, and stops. A synthetic track that resembles no one in particular injures nobody with standing to object. It is not disputed. It is not examined. It goes up, collects its fraction of a cent, and remains.
What Sony has assembled in six months is genuinely impressive and very narrowly aimed: a pipeline that can locate a quarter of a million impersonations of the performers on its books. Nothing of comparable ambition is being built to look at the other ninety thousand a day, because the other ninety thousand a day is nobody’s loss.
This is the oldest pattern in my field, and it has never once required anybody to behave badly. Between the late 1960s and 2014, a small group of Dutch scholars — Bob Haak its instigator, Josua Bruyn its dominant voice, Ernst van de Wetering eventually carrying it alone — ran the Rembrandt Research Project through the volumes of A Corpus of Rembrandt Paintings, the most serious attempt ever made to settle which pictures are by Rembrandt. The first five volumes accepted about 146 paintings. In the last one, as The Burlington Magazine notes, “no less than forty-four works rejected in the first three volumes of the Corpus are reinstated here, an increase of nearly one third on the 146 pictures accepted by the RRP.”
Four decades of the best technical and archival scholarship available, trained with enormous care on one question — is this one by him — and the answer still moved by a third. Not through carelessness. The question is hard, and almost nobody else was asking it at that intensity, and there was no second apparatus to check the first.
Meanwhile nobody spent forty years establishing which assistant laid in which hand, who ground the pigments, or which canvases leaving that workshop were substantially somebody else’s work. Those questions are partly answerable, and the answers would tell you more about how seventeenth-century painting was actually produced than another self-portrait attribution ever could. They went unexamined at that pitch because no money turned on them. The catalogue was always as long as the market required it to be, and the people missing from it were not excluded on purpose. Recording them cost something and returned nothing.
It is worth being clear about who carries none of the present arrangement. The expense of 260,000 notices falls on the company holding the names. The platform receives 90,000 tracks a day at essentially no marginal cost and pays out on one to three per cent of streams. The generators pay nothing and file nothing. If synthetic uploads halved overnight, the party with twenty thousand notices a month to serve would feel it, and almost no one else would.
Enforcement of this kind arrives when a loss grows large enough to fund it. It does not arrive because something is occurring on a vast scale. Ninety thousand tracks a day is a vast scale, and it has so far produced a tagging system at one service and a notice pipeline aimed at a set of famous voices. What would decide whether anything larger gets built is not the ninety thousand but the one to three per cent, and the one to three per cent is small. My reading, offered as a reading: the machinery now in place is sized to the loss now being felt, and it will stay that size until the loss moves.
Deezer’s other figure from June is that up to 85 per cent of the streams those AI tracks do attract were identified as fraudulent. The portion of the flood that anyone currently collects on is mostly not being listened to by people either.
Sources
- Music Business Worldwide, “Sony Music has now asked platforms to remove more than 260,000 AI deepfakes…” (5 October 2026, reporting the Financial Times)
- The Music Network, “Sony Music Seeks Removal of Over 260,000 AI Deepfake Tracks” (6 October 2026)
- Deezer Newsroom, “AI music tops 50% of daily uploads on Deezer” (21 July 2026)
- The Burlington Magazine, “The Rembrandt Research Project and its denouement”