On 21 July 2026 Deezer said more than half of its daily new-music deliveries, at a June peak, were fully AI-generated. Those tracks account for 1-3% of streams — and by Deezer’s own count, up to 85% of the plays they drew in 2025 were fraudulent. The aesthetic debate is masking a far more mundane and far more expensive question: who collects the royalties.
On 21 July 2026, from Paris, the French streaming service Deezer released a number engineered to become a headline: roughly 90,000 fully AI-generated tracks uploaded every day, more than 50% of daily new-music deliveries at the peak recorded in June 2026. Within hours the figure was circulating in the form everyone expected — half of all new music is now made by a machine, and something is over.
The same press release contained a second number that travelled much less far. Those tracks account for between 1% and 3% of total streams on Deezer. And a third, more awkward still: according to the company, up to 85% of the streams generated on those tracks in 2025 were fraudulent and were demonetised. Read together, the three figures tell a different story from the aesthetic one. It is not a story about what music we will listen to. It is a story about where the money goes.
Where the AI actually sits
It is worth being precise, because artificial intelligence appears twice here, in opposite roles.
The first is obvious: music generators — Deezer names models such as Suno and Udio purely as identifiable sources — produce finished audio from a text prompt. Marginal cost near zero, production time in minutes.
The second is the interesting part. Since January 2025 Deezer has run every delivery through a machine learning system trained to recognise the acoustic signatures left by the major generative models. It does not judge whether a track is good, or original: it looks for statistical regularities in the audio signal — synthesis artefacts, spectral patterns typical of a specific architecture — that a human ear does not catch. The company claims 99.8% accuracy and fewer than one false positive in ten thousand, applies the tag at album level, and filed two patents in December 2024, published by the European and US patent offices in June 2026. In 2025 alone the system detected and tagged 13.4 million tracks. From June 2025 Deezer became the first platform to make that tag visible to users; since January 2026 it has licensed the technology to third parties, with trials alongside Sacem — the French collecting society for authors and composers — and use by Billboard to flag AI tracks on its charts.
So: a classifier trained on synthetic audio, producing a label, which becomes a business decision. That is the real subject of the news.
The problem is not that the music is bad
Here is the contrarian part, and it is not an aesthetic opinion. It is arithmetic.
A catalogue that makes up half of deliveries and 1-3% of listening is not winning audiences. But be careful how you read that gap. The temptation is to say listeners recognised it and rejected it. The data does not say that. Deezer itself attributes the tiny share to one of its own choices: tracks flagged as fully AI-generated are excluded from algorithmic recommendations and editorial playlists. On a streaming platform, staying out of the recommender means staying out of the traffic. The 1-3% measures a distribution policy, not a preference.
If anything, taste points the other way. An Ipsos survey commissioned by Deezer in November 2025, covering 9,000 people across eight countries, found that 97% could not tell an AI track from a human one in a blind test. The same respondents demanded transparency: 80% want labelling, 52% would keep AI tracks out of the main charts. That is a demand for provenance, not for quality. People cannot tell the difference and do not claim they can — they want to be told.
The product isn’t the song. It’s the stream
If audiences don’t listen and platforms don’t push them, why upload ninety thousand a day?
Because for a meaningful share of uploaders the song is not the product. The stream is. Under a pro-rata royalty pool — where the pot is divided in proportion to plays — every simulated listen is a withdrawal from everyone else’s share. For that purpose synthetic music is ideal raw material: no advances, no session musicians to pay, producible at industrial volume, and needed only as a vehicle for automated playback. It is spam with a waveform.
Which is why the policy announced on 21 July 2026 is more interesting than the headline it generated. Deezer did not ban AI-generated music. It announced the removal of two categories: AI tracks used to generate fraudulent streams, and tracks unplayed for at least six months. The criterion is not who made it, but what it is for. A synthetic track that finds listeners stays. A synthetic track that exists only to be played by bots goes.
Three cautions the press release does not foreground
All of that said, these numbers need handling with gloves, for structural reasons.
- They are self-reported. The 85%, the 1-3% and the 8% fraud figure across Deezer’s whole catalogue in 2025 are internal metrics, unaudited by third parties, produced by the same proprietary tool Deezer has been selling to others since January 2026. The party measuring the phenomenon also sells the thermometer. That does not make the figures false; it means nobody has checked them.
- The yardstick moved. In September 2025 Deezer spoke of “up to 70%” of AI streams being fraudulent, and about 0.5% of total streams. By July 2026 the figures for the same year had become 85% and 1-3%. There is no explanation of whether the phenomenon changed or the method did. And the company’s research director, Manuel Moussallam, acknowledged to Billboard that part of the observed growth reflects better detection, not only a growing problem.
- The denominator doesn’t add up. Luminate estimates roughly 150,000 tracks delivered daily across all platforms combined; Deezer’s percentages would imply around 180,000 daily deliveries to Deezer alone. The bases are not comparable, and in the release the “90,000 a day” appears once as a June monthly average and once as a peak daily flow.
There is also a useful benchmark. France’s Centre national de la musique, a body under the Ministry of Culture, estimated in 2021 that 1-3% of streams in France were fraudulent; the analytics firm Beatdapp puts global fraud at around 10%. Deezer’s declared 8% for 2025 sits inside that range. Streaming fraud was not invented by music generators: it already existed, using dirt-cheap human catalogues. Generative AI did not create the problem — it collapsed the cost of entry.
What is actually being decided
The rest of the industry is moving along different lines. Spotify announced in September 2025 that it had removed more than 75 million tracks classified as spam over twelve months, alongside an anti-impersonation policy and AI disclosure in credits. Bandcamp banned AI music in January 2026; Tidal removed its monetisation in June 2026. In July 2026 the RIAA, IFPI and others launched a shared tagging system. No major competing platform, however, is reported to have adopted Deezer’s detector.
That is where the real contest lies. If the “AI” tag becomes shared infrastructure, the classifier stops being an editorial tool and becomes an economic referee: it decides who gets into the royalty pool. With one consequence nobody has yet addressed publicly. The reliability of AI-content detectors is contested in the research literature, and Deezer’s claimed 99.8% has no published external validation. One false positive in ten thousand sounds negligible — until you remember the tag is applied at album level and that the catalogue tagged in 2025 runs to 13.4 million tracks. The risk is not that a machine writes a bad song. It is that a machine decides, with no appeal and no audit, that you didn’t write yours.