Why AI music is flooding streaming

Two numbers arrived in the same year and they do not agree with each other. Deezer said on 21 July 2026 that fully AI-generated tracks reached about 90,000 a day in June, more than half of everything uploaded to it at peak, while the same tracks take between 1% and 3% of its plays; a Deezer/Ipsos survey of 9,000 listeners in eight countries found 97% could not tell a synthetic song from a human one. In Chicago, about twenty musicians stood in the rain on 1 October to launch a campaign against AI music and to announce a platform blackout on 26 and 27 October. Nobody in that argument is claiming a machine cannot make a listenable song, because the listeners have already settled it. What is being argued about is who gets paid, and by what rule.

Deezer has been counting this longer than anyone else, because it built the detector first: about 10,000 fully AI-generated tracks a day in January 2025, 90,000 by June 2026 — a ninefold rise in eighteen months — and 13.4 million tracks tagged across 2025. The share passed half of all daily deliveries for the first time in June. The same announcement carried the number that matters more to anyone who makes music: those tracks account for only 1% to 3% of listening on the platform, and Deezer said it would begin removing AI tracks used to inflate streams fraudulently, as well as ones that have not been played in six months.

We would read that second figure carefully rather than cheerfully. Part of the reason AI tracks are barely listened to is that these platforms keep them out of recommendations and editorial playlists, so the 1% to 3% measures a policy as much as an appetite. What it does show is that flooding a catalogue is not the same as finding an audience, and that the uploads are being driven by people who do not need one.

The cleanest fact in the whole argument is the blind test. Deezer commissioned Ipsos to have 9,000 adults in eight countries listen to three tracks — two made entirely by AI, one by people — and pick out the human one. Ninety-seven per cent failed, and 52% said being unable to tell made them uncomfortable. Eighty per cent wanted AI-generated music labelled, which is the one thing most of the industry now broadly agrees on and implements differently.

That result should end a certain kind of argument. Any claim that AI music is obviously worse is one the public cannot verify. And any claim that it is indistinguishable is not a compliment to the machines so much as a statement about how much popular music is already a formula. The label does the work the ear cannot, which is why the listening public asks for it.

The Chicago protest is small and specific, which is why we take it seriously. Around twenty artists gathered outside the Federal Center on 1 October, in black shirts reading Human Music Only, demanding legal safeguards against AI copies of their work; organiser and singer Melody Angel put the ask as policy that says which music is real and which is not, so listeners have the choice. The group's next step is a blackout on 26 and 27 October asking people to avoid Spotify and Apple Music for two days. It follows an open letter published on 22 September criticising the platforms' payouts and their handling of unlabelled AI music, which reporting has put at between 70 and 80 signatories.

The platforms are already moving, unevenly. Spotify said in August 2026 that it would label AI-generated artist profiles — an AI Persona badge — and keep their music out of editorial and algorithmic recommendations, though the badge speaks to an artist's public identity rather than to how a particular track was made. Deezer became the first service to tag AI tracks for listeners in June 2026. The gap between those two positions is the whole difficulty: one labels the machine, the other labels the person, and neither yet tells you what is in the four minutes you are hearing.

The most useful local contribution is not a protest but a request for a framework. In a Straits Times forum letter, the chief executive of COMPASS, the society that administers music copyright in Singapore, argued that making music available on streaming has never meant consenting to commercial AI training, and asked Singapore to convene creators, developers, policymakers and rights organisations to build a practical licensing framework — transparency, recognition and payment, rather than a ban. The letter points to a ruling by the Munich Regional Court on 31 July 2026 that the AI music company Suno infringed copyright by training on songs represented by the German rights body GEMA without authorisation, and ordered it to disclose revenues and pay damages, subject to appeal.

We prefer this shape of answer to the rallying kind, for a plain reason: it is the only version that can end in a payment. A framework that tells a creator whether their catalogue was used, by whom and for what, and pays them when it was, addresses the same grievance as a blackout without pretending the tools will go away. It also puts the cost where the money is, on the training rather than on the listener's conscience.

Two cautions. The 90,000-a-day figure is Deezer's own detector counting Deezer's own inbox, and it is not audited by anyone else; it is the best number we have, not an independent one. And a three-track blind test measures recognition under artificial conditions, not taste over years — we would not conclude from it that audiences will accept synthetic music in the places where they actually care what they are hearing.

We would not spend a minute arguing about whether AI music is real music, because the audience has already told us it cannot tell the difference and does not particularly want to. The fight that pays is about the catalogue and the credit, and it is being fought in the right two places: a court in Munich, and a rights society in Singapore asking for licensing before the case arrives rather than after.

The practical advice we would give a creator is unromantic. Label your own work before someone labels it for you; keep the record of how it was made; and build the audience somewhere you can reach them without asking a platform's permission, because a playlist is rented and an email list is not. Nothing about that requires being against the tools.

Our worry is that labelling becomes a tax on honesty. The people who disclose will be the ones already playing straight, while the fraud farms that upload 90,000 tracks a day carry on. A label only helps if the platforms enforce the same rule on everyone, and so far they have each decided separately.